{"id": "E001", "collection": "comparison", "domain": "finance_operations", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-11-no_reviewer", "topology_family_id": "topology-001", "annotation_status": "candidate", "image_path": "images/comparison/E001.png", "title": "Invoice intake to payment approval/execution pipeline", "purpose": "Model a connected finance workflow where an invoice-handling agent ingests externally supplied invoice documents, extracts payment-relevant details, drafts payment actions, and forwards them into an approval and execution path.", "assumptions": ["The invoice documents and any supporting evidence originate outside the trusted application zone.", "The invoice intake agent can read external documents and produce structured payment drafts, but its internal validation, guardrails, and approval checks are unspecified.", "An approval step and a payment-execution step exist downstream of the draft handoff.", "The attacker has no privileged system access, but can influence the content of external invoices or evidence.", "No claims are made about whether controls are effective; only that their implementation/effectiveness is unspecified."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Invoice intake is a documented application workflow that accepts external invoice files, extracts fields into a draft record, and hands the draft onward; validation depth and guardrails are unspecified."}, {"id": "K2", "component_ids": ["C4", "C5"], "flow_ids": ["F3", "F4"], "fact": "A separate approval-and-execution path exists downstream of the draft handoff; review and authorization checks are present by design, but their exact implementation and effectiveness are unspecified."}, {"id": "K3", "component_ids": ["C2"], "flow_ids": ["F1", "F2"], "fact": "The invoice agent is allowed to read external documents and generate structured payment actions; tool authority and instruction-handling limits are not specified."}], "architecture": {"components": [{"id": "C1", "label": "Invoice documents", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice intake agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Payment draft record", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Approval reviewer", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment execution service", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "structured invoice extraction and payment draft", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "payment draft for review", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F5", "source": "C5", "target": "C3", "data": "execution status and payment reference", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External source"}, {"id": "Z2", "label": "Invoice processing"}, {"id": "Z3", "label": "Approval and execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "external invoice prompt/instruction injection into intake agent", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "Payment draft record", "threat_vector": "Attacker supplies crafted invoice text or supporting evidence that influences the intake agent's extraction and drafting behavior.", "preconditions": ["Attacker can influence the content of external invoices or evidence.", "The intake agent processes external documents and generates a payment draft.", "Validation depth and guardrails are unspecified."], "impact": "Incorrect or attacker-influenced payment draft contents may be written to the draft record, enabling downstream fraudulent or incorrect payment processing.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Invoice intake is a documented application workflow that accepts external invoice files, extracts fields into a draft record, and hands the draft onward; validation depth and guardrails are unspecified.", "justification": "The architecture explicitly accepts untrusted external invoices across B1 and has an agent that extracts and drafts payment actions, with unspecified guardrails. That supports a feasible agent-specific manipulation path affecting the draft record."}, {"id": "T2", "category": "T", "subtype": "invoice field tampering leading to altered payment draft", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "Payment draft record", "threat_vector": "Attacker alters invoice content so extracted fields or payment details differ from the legitimate invoice intent.", "preconditions": ["Attacker can modify or supply the external invoice documents.", "The intake agent relies on extracted invoice fields to create the payment draft."], "impact": "The payment draft may contain altered payee, amount, or reference details, causing incorrect payment instructions to be reviewed downstream.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Invoice intake is a documented application workflow that accepts external invoice files, extracts fields into a draft record, and hands the draft onward; validation depth and guardrails are unspecified.", "justification": "The explicit workflow from external invoice input to structured draft creation creates a tampering path on the data being transformed into the payment draft. The control fact does not describe robust validation, so exposure is supported."}, {"id": "T3", "category": "I", "subtype": "sensitive payment draft disclosure to external source via review handoff assumptions", "component_ids": ["C3", "C4"], "flow_ids": ["F3"], "boundary_ids": ["B2"], "asset": "Payment draft record", "threat_vector": "The draft is transferred from invoice processing into the approval zone where sensitive invoice and payment details may be exposed to parties handling review.", "preconditions": ["The payment draft contains sensitive financial or business information.", "An unauthorized external recipient or untrusted party can observe the review process, which is not stated in the architecture."], "impact": "Potential disclosure of invoice and payment details to parties outside the intended need-to-know set.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "A separate approval-and-execution path exists downstream of the draft handoff; review and authorization checks are present by design, but their exact implementation and effectiveness are unspecified.", "justification": "A cross-boundary review flow exists, but the architecture does not show any unauthorized recipient or exfiltration path. The downstream controls are present by design, yet their effectiveness is unspecified, so disclosure exposure cannot be confirmed."}, {"id": "T4", "category": "R", "subtype": "insufficient traceability for draft-to-execution decisions", "component_ids": ["C3", "C4", "C5"], "flow_ids": ["F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "Payment execution record", "threat_vector": "A malicious or erroneous actor can dispute how the payment draft was reviewed, approved, or executed if traceability is incomplete.", "preconditions": ["Approval and execution decisions rely on records that may be incomplete or not strongly attributable.", "No explicit audit/logging mechanism is described in the architecture."], "impact": "Reduced ability to attribute who approved or executed a payment and to reconstruct the decision path.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "A separate approval-and-execution path exists downstream of the draft handoff; review and authorization checks are present by design, but their exact implementation and effectiveness are unspecified.", "justification": "The workflow has review and execution steps, but the architecture does not specify auditability or non-repudiation controls. This makes repudiation possible in principle, but the evidence is insufficient to classify the path as exposed."}], "scope_notes": ["Boundary IDs assigned from the ordered flows: B1 for Z1↔Z2 and B2 for Z2↔Z3.", "F2 is same-zone within Z2 and does not cross a trust boundary.", "The architecture does not specify whether the approval reviewer is human or automated, so unauthorized disclosure or spoofing against that component remains unsupported.", "No direct path from the execution service back to the external source is shown; reverse flow F5 remains within B2 and does not imply external disclosure.", "Controls are present as described, but their effectiveness is unspecified; unsupported paths were not forced into the analysis."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Treat all external invoice content as untrusted input: add strict content normalization, schema validation, and rule-based extraction checks before the agent can write any payment draft; require human confirmation for any new or changed payee, amount, bank detail, or payment reference before persisting to C3.", "verification": "Use benign test invoices containing prompt-like instructions, conflicting fields, and malformed attachments; confirm the agent ignores instructions embedded in documents, rejects invalid structures, and only writes drafts after the required field-level checks and human confirmation succeed.", "existing_control_effect": "K1 and K3 already show that external invoices are ingested and transformed into drafts, but they do not specify effective guardrails. The downstream draft workflow exists, so this control would strengthen the existing intake path rather than replace it.", "residual_risk": "A crafted invoice may still influence extraction edge cases or non-critical metadata, and approval-stage review may miss subtle fraud if reviewers are rushed or rely on the draft output too heavily."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Enforce source-document authenticity and field consistency checks at intake: verify invoice provenance where possible, compare extracted fields against expected vendor/master-data records, and block draft creation when payee, amount, or reference values deviate from approved templates or historical patterns without explicit reviewer approval.", "verification": "Submit benign invoices with altered payee and amount fields and confirm the pipeline flags the mismatch, prevents automatic draft finalization, and preserves an auditable rejection reason without creating an approved draft.", "existing_control_effect": "K1 indicates the workflow already extracts fields into a draft record, so there is an existing transformation point to control. However, validation depth is unspecified, so current protection against field tampering is uncertain.", "residual_risk": "If an attacker supplies a plausible invoice that matches expected patterns closely, some altered values may still reach review, especially when business exceptions are common or the reference data is incomplete."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Limit review-time exposure by displaying only the minimum necessary payment data to reviewers, masking nonessential sensitive fields, and restricting approval and execution access to authenticated users or services with a documented need-to-know; log every view and export of draft details.", "verification": "Use a normal draft containing sensitive but noncritical sample data and confirm that masked fields remain hidden by default, only authorized reviewers can unmask them, and access logs record each disclosure event.", "existing_control_effect": "K2 states that approval and authorization checks exist by design in the downstream path, which helps reduce casual disclosure. Because their implementation and effectiveness are unspecified, additional least-privilege display and logging controls are still needed.", "residual_risk": "Authorized reviewers can still see the needed payment details, and approved drafts may be disclosed through screenshots, local notes, or other out-of-band handling not prevented by the architecture."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5", "C3"], "recommendation": "Add tamper-evident audit logging for draft review, approval, execution, and payment-reference updates; include who approved, what changed, when it changed, and the linked draft identifier, and require these records before execution can be finalized.", "verification": "Run a benign payment through the workflow and confirm the system produces an immutable audit trail covering the draft, approval action, execution action, and returned payment reference, with records cross-linked and time-ordered.", "existing_control_effect": "K2 shows the approval/execution path already exists, which provides natural checkpoints to attach logging and traceability. The current design, however, does not specify that those checkpoints are attributable or immutable.", "residual_risk": "Audit records can still be incomplete if an upstream component fails before logging, and strong traceability does not prevent a malicious insider from making an authorized but inappropriate approval."}]}
{"id": "E002", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "single_prompt", "trial_id": "slot-06-single_prompt", "topology_family_id": "topology-002", "annotation_status": "candidate", "image_path": "images/comparison/E002.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "Relevant control implementation and effectiveness are unspecified.", "The support agent can query account data to prepare responses, but privileged access is not assumed beyond the stated workflow.", "An attacker can influence external help documents or ticket text without privileged internal access."], "controls": [{"id": "Ctrl1", "component_ids": ["C2", "C3"], "flow_ids": ["F2", "F3", "F1"], "fact": "Document and ticket inputs are passed through the drafting workflow without an explicit content-authentication or instruction-sanitization guarantee being stated."}, {"id": "Ctrl2", "component_ids": ["C1", "C4"], "flow_ids": ["F4", "F5"], "fact": "The agent performs account lookups through the customer-account store during response preparation."}, {"id": "Ctrl3", "component_ids": ["C1", "C3"], "flow_ids": ["F6"], "fact": "The draft response is presented to the support agent before any external reply is sent."}], "architecture": {"components": [{"id": "C1", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C2", "label": "Help Document Fetcher", "zone": "Z2", "kind": "process"}, {"id": "C3", "label": "Response Drafting Tool", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}, {"id": "C5", "label": "External Help Documents", "zone": "Z1", "kind": "external_entity"}, {"id": "C6", "label": "Support Ticket Intake", "zone": "Z1", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C6", "target": "C3", "data": "customer question and ticket text", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C5", "target": "C2", "data": "help article content", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C3", "data": "retrieved document snippets", "boundary_ids": []}, {"id": "F4", "source": "C1", "target": "C4", "data": "account lookup request", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C4", "target": "C1", "data": "account details", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C3", "target": "C1", "data": "draft response and citations", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support Application"}, {"id": "Z3", "label": "Account Data"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation via external help text", "component_ids": ["C5", "C2", "C3", "C1"], "flow_ids": ["F2", "F3", "F6"], "boundary_ids": ["B1"], "asset": "response-generation workflow and agent tool authority", "threat_vector": "Attacker-controlled help document content can inject instructions into retrieved snippets that the drafting tool or agent may follow while preparing a reply.", "preconditions": ["Attacker can influence external help documents.", "The drafting workflow treats retrieved text as trusted guidance or context.", "No explicit instruction-sanitization guarantee is stated."], "impact": "The agent may follow malicious instructions, producing unsafe or incorrect support actions and responses.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["Ctrl1"], "evidence_quote": "Document and ticket inputs are passed through the drafting workflow without an explicit content-authentication or instruction-sanitization guarantee being stated.", "justification": "The path crosses from External Sources to Support Application through retrieved document snippets, and the architecture explicitly notes the lack of a stated content-authentication or instruction-sanitization guarantee. This supports an agent-specific manipulation risk."}, {"id": "T2", "category": "I", "subtype": "account-data exposure through response drafting", "component_ids": ["C4", "C1", "C3"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "customer account details", "threat_vector": "The support agent can retrieve account details and incorporate them into drafted responses, which may disclose data to an unauthorized recipient if the draft is mishandled or over-shared.", "preconditions": ["The agent can access account data for support work.", "The drafted response includes account details or derived identifiers.", "A downstream recipient is not authorized for all included details."], "impact": "Customer account information could be disclosed beyond the intended support context.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["Ctrl3", "Ctrl2"], "evidence_quote": "The draft response is presented to the support agent before any external reply is sent.", "justification": "The architecture shows a review step before external sending, which reduces but does not eliminate disclosure risk. The account lookup flow exists, but the draft review provides a limiting control on the outbound path."}, {"id": "T3", "category": "R", "subtype": "support action attribution gap", "component_ids": ["C1", "C3", "C4"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "account lookup and response provenance", "threat_vector": "If account lookups and drafted responses are not linked to strong provenance, the agent may later deny having used particular account data or having produced a specific response.", "preconditions": ["Logs or provenance are incomplete or not retained.", "Support work is performed through the drafting tool and account lookup flow."], "impact": "Disputes over who accessed data or authored a response may be hard to resolve.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["Ctrl2", "Ctrl3"], "evidence_quote": "The agent performs account lookups through the customer-account store during response preparation.", "justification": "The architecture clearly includes account access and drafting, but no explicit audit or immutable record control is stated. Effectiveness of provenance handling is therefore unspecified."}], "scope_notes": ["Boundary IDs were assigned by first-seen zone pairs: B1 for External Sources to Support Application, B2 for Support Application to Account Data.", "The architecture does not state whether the external help documents are authenticated, filtered, or sandboxed; unsupported assumptions should not be made.", "The account store is used for support preparation, but the authorization model for specific fields is unspecified.", "No additional attacker capability beyond influence over external documents or ticket text is assumed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C1"], "recommendation": "Add a content-separation rule in the help fetcher and drafting tool so retrieved document text is labeled as untrusted context, stripped of executable-style instructions, and never treated as agent directives without user confirmation.", "verification": "Submit a benign help article containing embedded imperative text and confirm the drafting tool preserves it only as quoted evidence, not as instructions to follow.", "existing_control_effect": "Ctrl1 already indicates the workflow lacks a stated instruction-sanitization guarantee, so this mitigation addresses the identified gap.", "residual_risk": "Some prompt-like content may still influence summarization or phrasing, especially if the agent overrides the separation rule."}, {"threat_id": "T2", "enforcement_component_ids": ["C1", "C3", "C4"], "recommendation": "Restrict account-field retrieval to the minimum needed for the case and require the support agent to redact sensitive fields before external sending in the drafting review step.", "verification": "Use a test account record with a nonessential sensitive field and confirm the draft view masks it unless the case explicitly requires it.", "existing_control_effect": "Ctrl3 already places the draft before the support agent prior to external reply, and Ctrl2 confirms the lookup path exists for review-based handling.", "residual_risk": "The agent may still reveal details deliberately or mistakenly if redaction rules are incomplete or bypassed."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C3", "C4"], "recommendation": "Record account lookup requests and final draft versions in an append-only audit trail tied to the support session so later disputes can be traced to a specific lookup and response.", "verification": "Perform a normal support session and confirm the audit record contains the request, retrieved account reference, and the reviewed draft without altering the original entries.", "existing_control_effect": "Ctrl2 and Ctrl3 show the lookup and review steps exist, but they do not specify durable provenance; this mitigation adds traceability.", "residual_risk": "Audit records can still be incomplete if operators work outside the tool path or if retention is not enforced."}]}
{"id": "E003", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-16-full_workflow", "topology_family_id": "topology-003", "annotation_status": "candidate", "image_path": "images/comparison/E003.png", "title": "Guarded multi-analyst incident triage with coordinator and case store", "purpose": "Two analyst agents independently review externally supplied incident evidence, then report normalized findings to a coordinator that consolidates results into a case store for response tracking.", "assumptions": ["Externally supplied incident content may be adversarially crafted, but the attacker has no authenticated access to internal systems.", "Analyst agents can only use bounded retrieval or parsing tools against approved sources and allowed actions.", "The coordinator is authorized to write only case records, not arbitrary code or system commands.", "Provenance metadata exists for evidence and findings, but completeness and tamper-resistance are not guaranteed unless separately enforced."], "controls": [{"id": "K1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Evidence intake is restricted to approved incident sources, and analyst tool use is bounded to read-only retrieval and parsing of those sources."}, {"id": "K2", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4"], "fact": "Analyst outputs to the coordinator carry provenance metadata and are accepted as normalized findings rather than direct tool commands."}, {"id": "K3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "The coordinator has explicit write permission to the case store, limited to case records and status updates."}, {"id": "K4", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "Case writes are mediated through a record-oriented interface; arbitrary execution or direct store scripting is not part of the design."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "normalized findings, provenance notes", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C3", "target": "C4", "data": "normalized findings, provenance notes", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C4", "target": "C5", "data": "consolidated case records, response status", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Evidence"}, {"id": "Z2", "label": "Analysis Workspace"}, {"id": "Z3", "label": "Coordination and Case Data"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "agent instruction manipulation via evidence", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "Analyst Agent A judgment and downstream findings", "threat_vector": "Adversarially crafted incident documents/logs/attachments attempt to steer the analyst agent's interpretation or tool use during bounded parsing/retrieval.", "preconditions": ["Attacker can supply external evidence content through the approved intake path.", "Agent behavior can be influenced by content in the evidence set."], "impact": "Corrupted or unsafe findings may be produced and forwarded to the coordinator, affecting case handling and response decisions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Evidence intake is restricted to approved incident sources, and analyst tool use is bounded to read-only retrieval and parsing of those sources.", "justification": "The architecture explicitly accepts externally supplied evidence into the analyst workspace, and the supplied assumption states that externally supplied incident content may be adversarially crafted. The bounded tool policy reduces arbitrary action but does not remove the content-manipulation path."}, {"id": "T2", "category": "I", "subtype": "sensitive case details disclosed through analyst reports", "component_ids": ["C2", "C4", "C5"], "flow_ids": ["F3", "F5"], "boundary_ids": ["B2"], "asset": "Case-related findings and response status", "threat_vector": "A compromised or overly detailed analyst report could disclose sensitive incident details into coordinator-managed case records beyond intended recipients.", "preconditions": ["A finding report contains sensitive evidence details or provenance that should not be broadly exposed.", "The coordinator stores the report-derived content in the case store."], "impact": "Sensitive incident information may be exposed in coordination/case data, increasing unauthorized visibility of evidence or response context.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "Analyst outputs to the coordinator carry provenance metadata and are accepted as normalized findings rather than direct tool commands.", "justification": "There is a defined path for analyst findings to cross into coordination and case storage, but the design facts do not specify access restrictions on who can later read case records or how much content is retained. The provenance/control statements reduce command injection concerns, but disclosure impact remains unspecified."}, {"id": "T3", "category": "T", "subtype": "tampering with provenance or case records", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": [], "asset": "Case store integrity and response status", "threat_vector": "A coordinator write path could be abused to alter case records or response status if record-oriented controls are bypassed or insufficiently enforced.", "preconditions": ["An attacker can influence coordinator inputs or behavior enough to alter what is written.", "Store-side authorization or record validation is incomplete."], "impact": "Case records or status may be modified, undermining integrity of incident tracking and response workflows.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Case writes are mediated through a record-oriented interface; arbitrary execution or direct store scripting is not part of the design.", "justification": "The design limits writes to record-oriented updates and excludes arbitrary scripting, which reduces some tampering paths. However, the effectiveness of record validation and prevention of malicious record content is not fully specified, so residual tampering risk remains unknown."}, {"id": "T4", "category": "D", "subtype": "case-store update saturation", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": [], "asset": "Coordinator availability and case store update capacity", "threat_vector": "High-volume or repetitive coordinator writes could saturate the record-oriented case update path and delay normal response tracking.", "preconditions": ["Coordinator receives many findings or repeated updates.", "The case store or write interface lacks stated rate limiting or queue protections."], "impact": "Response tracking may be delayed or partially unavailable due to overloaded case-write handling.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "The coordinator has explicit write permission to the case store, limited to case records and status updates.", "justification": "The architecture defines a write path but does not describe throttling, quotas, or backpressure. That makes availability impact plausible but not demonstrated as exposed."}], "scope_notes": ["No threat is asserted for direct unauthorized access to internal systems; the scenario explicitly says the attacker has no authenticated access to internal systems.", "Boundary B1 covers flows between External Evidence (Z1) and Analysis Workspace (Z2); boundary B2 covers flows between Analysis Workspace (Z2) and Coordination and Case Data (Z3).", "No same-zone boundary exists on F5, so threats on coordinator-to-case-store traffic are internal-path threats rather than trust-boundary crossings.", "The design facts do not specify who can read the case store, retention rules, or whether provenance metadata is immutable and tamper-evident.", "Unsupported paths such as arbitrary code execution are constrained by the stated record-oriented interface and are not claimed here."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4"], "recommendation": "Add a prompt- and parser-facing evidence sanitization gate in Analyst Agent A and a coordinator-side provenance check that rejects instructions embedded in evidence content, while allowing only normalized findings fields to pass onward.", "verification": "Send benign sample evidence containing quoted commands, links, and malformed sections; confirm the analyst outputs only bounded findings and the coordinator rejects any report that includes executable-looking instructions or non-allowed fields.", "existing_control_effect": "K1 already restricts intake to approved sources and limits analyst tools to read-only retrieval and parsing, which reduces arbitrary action and narrows the injection surface.", "residual_risk": "Adversarial evidence can still bias interpretation or wording inside allowed findings, so judgment manipulation risk is reduced but not eliminated."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Apply case-store field minimization and role-based read permissions so the coordinator stores only the minimum necessary incident summary, with sensitive source details and rich provenance kept out of broadly readable case records.", "verification": "Use a benign test case with sensitive-looking placeholders and confirm that stored records contain only the approved summary fields and that non-authorized readers cannot access detailed source content.", "existing_control_effect": "K2 and K3 already keep analyst outputs in a normalized form and limit coordinator writes to case records and status updates, which lowers the chance of direct leakage through command channels.", "residual_risk": "Some sensitive context may still be necessary for coordination, so limited disclosure can remain if summaries or provenance fields are readable by broader audiences than intended."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Enforce strict schema validation and append-only audit logging on coordinator writes so only approved case-status transitions and record fields are accepted, with rejected updates recorded for review.", "verification": "Submit benign updates with extra fields, altered status values, and reordered metadata; confirm the store accepts only schema-conformant record updates and logs rejected attempts without modifying existing records.", "existing_control_effect": "K3 and K4 already route writes through a record-oriented interface and exclude arbitrary execution or direct store scripting, which blocks the most powerful tampering paths.", "residual_risk": "A malicious or buggy coordinator can still write incorrect but schema-valid content unless business-rule validation and audit review are also enforced."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Introduce rate limiting and write queuing on the coordinator-to-store path, with backpressure and duplicate-update suppression for repeated status writes.", "verification": "Run a benign burst of repeated case updates and confirm the coordinator throttles excess writes, preserves normal-priority updates, and keeps the store responsive under load.", "existing_control_effect": "K3 and K4 confine writes to record updates through a non-executable interface, which prevents overload via scripting abuse but does not cap write volume.", "residual_risk": "A legitimate surge in findings can still delay case updates if the queue or throttle thresholds are too low or if downstream capacity is small."}]}
{"id": "E004", "collection": "comparison", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-09-full_workflow", "topology_family_id": "topology-004", "annotation_status": "candidate", "image_path": "images/comparison/E004.png", "title": "Guarded invoice-to-payment assistant with approval handoff", "purpose": "Model a finance workflow where an invoice-reading agent ingests externally supplied documents, extracts payment-relevant details, drafts payment requests, and passes them to a separate approval and payment-execution path under enforced controls.", "assumptions": ["External invoices and supporting evidence may be attacker-influenced and are treated as untrusted input.", "The invoice agent runs with bounded tool authority and cannot directly execute payments.", "Approval is a distinct control point with authorization checks; execution requires passing that gate.", "Provenance metadata is available for documents, but its effectiveness against forged or incomplete evidence is not fully guaranteed from the described facts."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "fact": "Document intake records provenance metadata before the invoice agent sees the content, and the agent reads only the stored invoice record rather than the supplier channel directly."}, {"id": "K2", "component_ids": ["C3"], "flow_ids": ["F4"], "fact": "The invoice agent has bounded tool authority and can prepare drafts but cannot directly authorize or submit payments."}, {"id": "K3", "component_ids": ["C5", "C6"], "flow_ids": ["F5"], "fact": "Payment release requires an approval step with authorization checks before the payment service accepts the request."}, {"id": "K4", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "fact": "The workflow relies on provenance metadata and content extraction, but the described facts do not guarantee that forged or incomplete supporting evidence will always be detected."}], "architecture": {"components": [{"id": "C1", "label": "Supplier portal", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Document intake service", "zone": "Z1", "kind": "process"}, {"id": "C3", "label": "Invoice agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Provenance store", "zone": "Z2", "kind": "data_store"}, {"id": "C5", "label": "Approval console", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Payment service", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": []}, {"id": "F2", "source": "C2", "target": "C4", "data": "normalized documents plus provenance metadata", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C4", "target": "C3", "data": "invoice record and provenance context", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C5", "data": "payment draft with extracted payee, amount, and justification", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C6", "data": "approved payment request", "boundary_ids": []}, {"id": "F6", "source": "C6", "target": "C1", "data": "payment confirmation or remittance notice", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "Untrusted document intake"}, {"id": "Z2", "label": "Invoice processing"}, {"id": "Z3", "label": "Approval and payment execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z1", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "invoice-content-and-provenance-manipulation", "component_ids": ["C1", "C2", "C4", "C3", "C5"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1", "B2"], "asset": "invoice records and payment drafts", "threat_vector": "Attacker-controlled supplier documents are crafted to alter extracted payee, amount, justification, or provenance context as they move from untrusted intake into invoice processing and drafting.", "preconditions": ["Attacker can influence the supplier-supplied invoice documents or supporting evidence.", "The workflow depends on content extraction and provenance metadata to drive the draft payment request."], "impact": "Incorrect payment draft content may be approved and forwarded, causing fraudulent or incorrect payment execution.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "The workflow relies on provenance metadata and content extraction, but the described facts do not guarantee that forged or incomplete supporting evidence will always be detected.", "justification": "This is a supported manipulation path from untrusted documents through intake and storage to the agent-generated draft. The quoted control explicitly says forged or incomplete evidence may not always be detected, so the path remains exposed."}, {"id": "T2", "category": "A", "subtype": "prompt-or-context-manipulation-via-invoice-content", "component_ids": ["C1", "C2", "C4", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "invoice agent context and extracted invoice record", "threat_vector": "An attacker embeds instructions or misleading content in invoice documents so the invoice agent's extracted record and provenance context are manipulated before draft generation.", "preconditions": ["Attacker controls or influences invoice documents or supporting evidence.", "The agent consumes the stored invoice record and provenance context as inputs to its drafting behavior."], "impact": "The agent may follow attacker-shaped instructions or produce a misleading payment draft, leading to unsafe agent tool use and downstream fraud.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K4"], "evidence_quote": "Document intake records provenance metadata before the invoice agent sees the content, and the agent reads only the stored invoice record rather than the supplier channel directly.", "justification": "The agent processes untrusted content after ingestion, so attacker-supplied document text can influence the record it reads. The control limits direct channel access but does not state robust sanitization or instruction suppression, and K4 leaves detection of forged or incomplete evidence uncertain."}, {"id": "T3", "category": "R", "subtype": "payment-request-origin-and-approval-dispute", "component_ids": ["C3", "C5", "C6"], "flow_ids": ["F4", "F5"], "boundary_ids": ["B2"], "asset": "payment draft and approval decision trail", "threat_vector": "A malicious or compromised drafter can deny having introduced incorrect payee, amount, or justification into the payment draft that passed to approval.", "preconditions": ["A payment draft is created by the invoice agent and handed to a separate approval console.", "The architecture description does not state explicit non-repudiation logging for draft authorship or approval rationale."], "impact": "Disputes over who introduced the erroneous payment content may hinder investigation and accountability after an incorrect or fraudulent payment attempt.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The invoice agent has bounded tool authority and can prepare drafts but cannot directly authorize or submit payments.", "justification": "The workflow separates draft creation from approval, but the provided facts do not specify audit or non-repudiation controls. Because authorization checks exist, the operational path is constrained, but repudiation resistance is unspecified rather than clearly absent or present."}, {"id": "T4", "category": "D", "subtype": "approval-pipeline-overload-or-retry-abuse", "component_ids": ["C1", "C2", "C3", "C5", "C6"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1", "B2"], "asset": "invoice processing and approval availability", "threat_vector": "High-volume or repeated attacker-supplied invoices can overload intake, extraction, drafting, or approval handling, delaying legitimate payment processing.", "preconditions": ["Attacker can send many invoice documents or malformed supporting evidence to the supplier portal.", "The architecture description does not provide rate limits, queue isolation, or abuse throttling facts."], "impact": "Legitimate invoices may be delayed or blocked, reducing availability of invoice processing and payment execution.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K3"], "evidence_quote": "Payment release requires an approval step with authorization checks before the payment service accepts the request.", "justification": "The presence of approval checks constrains execution, but the architecture gives no explicit availability controls for intake or drafting. This makes denial-of-service plausible but not explicitly supported as exposed; control effectiveness for overload resistance is unspecified."}], "scope_notes": ["Boundary IDs were assigned from the listed flows: B1 for Z1-Z2, B2 for Z2-Z3, and B3 for Z1-Z3.", "No direct threat path was identified for C6->C1 beyond ordinary payment confirmation/remittance notice delivery; that flow alone does not demonstrate disclosure or abuse.", "The architecture does not describe authentication, logging, rate limiting, or content validation details beyond provenance metadata and approval authorization checks, so related attack paths remain incomplete in the available facts."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4", "C3", "C5"], "recommendation": "Add strict invoice-content validation at intake and storage: separate document text from provenance fields, require schema-checked extraction with checksum/hash binding to the original upload, and block draft generation unless provenance and amount/payee fields pass consistency checks.", "verification": "Submit benign test invoices with mismatched payee, amount, or altered supporting evidence and confirm the workflow rejects or flags them before draft creation and approval handoff.", "existing_control_effect": "K4 already notes reliance on provenance metadata and content extraction, and K1 stores provenance before agent access, which reduces direct supplier-channel tampering but does not guarantee detection of forged or incomplete evidence.", "residual_risk": "An attacker may still craft plausible documents that pass validation gaps or exploit extractor errors, so some incorrect drafts can remain possible if review does not catch them."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Harden the intake-to-agent handoff with instruction-aware content normalization: strip or quarantine imperative text from invoices, label untrusted document text as data-only context, and require the agent to ignore embedded operational instructions unless separately authorized by workflow policy.", "verification": "Use harmless invoices containing fake directives like 'ignore previous fields' and confirm the agent treats them as inert content and produces the same result as a neutral invoice with identical business data.", "existing_control_effect": "K1 already prevents the agent from reading the supplier channel directly and forces it to consume stored records, which narrows exposure but does not by itself neutralize malicious document text.", "residual_risk": "If normalization misses adversarial phrasing or the model overweights untrusted text, some context manipulation can still affect extracted outputs or draft quality."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C5", "C6"], "recommendation": "Add tamper-evident audit logging for every draft and approval step, including draft origin, extracted fields, approver identity, timestamps, and immutable decision records linked to the payment request.", "verification": "Perform a benign end-to-end test payment draft and approval, then confirm the audit trail reconstructs who created the draft, who approved it, and the exact fields forwarded to payment execution.", "existing_control_effect": "K2 and K3 already separate drafting from authorization and require approval checks before execution, which limits unilateral payment submission but does not provide explicit non-repudiation evidence.", "residual_risk": "A participant may still dispute intent or claim a process error, but the audit trail should materially improve attribution and investigation."}, {"threat_id": "T4", "enforcement_component_ids": ["C2", "C3", "C5", "C6"], "recommendation": "Implement intake and workflow rate limiting with queue isolation, per-supplier quotas, backpressure, and malformed-input rejection so repeated invoices cannot starve approval or execution resources.", "verification": "Run a controlled burst of benign test invoices within a staging environment and confirm limits trigger throttling or queuing without preventing normal-priority invoices from progressing.", "existing_control_effect": "K1–K3 preserve separation between stages and prevent direct payment execution by the agent, but they do not state any capacity or abuse controls against volume-based overload.", "residual_risk": "A sustained high-volume flood can still consume shared infrastructure capacity or increase latency if limits are misconfigured or if attack volume exceeds provisioned headroom."}]}
{"id": "E005", "collection": "comparison", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-18-full_workflow", "topology_family_id": "topology-005", "annotation_status": "candidate", "image_path": "images/comparison/E005.png", "title": "Incident evidence triage with dual analyst agents, coordinator, and case store", "purpose": "Support incident response by having two analyst agents review externally supplied evidence, consolidate findings via a coordinator, and persist case records in a shared store.", "assumptions": ["External incident documents, logs, and attachments may be attacker-influenced, but the attacker has no privileged access to internal services or operator accounts.", "Relevant control implementation and effectiveness are unspecified; no assumption is made that validation, sandboxing, approval, or authentication controls are present or absent.", "The two analyst agents are distinct components that each consume external evidence and produce findings to a coordinator.", "The coordinator can write consolidated results into a case store that is inside the trusted application zone."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Unspecified evidence handling controls govern analyst ingestion of external files and logs; effectiveness is not assumed."}, {"id": "K2", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "Coordinator-only write path to the case store; analysts do not write directly to storage."}, {"id": "K3", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4", "F5"], "fact": "Findings are mediated through the coordinator before persistence, preserving a single consolidation step."}, {"id": "K4", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External evidence remains outside the trusted zone until consumed by analyst agents."}], "architecture": {"components": [{"id": "C1", "label": "Evidence intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case store", "zone": "Z2", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, extracted indicators", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, extracted indicators", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "consolidated case record", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External evidence sources"}, {"id": "Z2", "label": "Trusted analysis and case handling"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "external evidence manipulates analyst context or tool use", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "Analyst agent A reasoning and outputs", "threat_vector": "Attacker-influenced incident documents, logs, or attachments are ingested by Analyst agent A and can steer its interpretation or unsafe downstream tool use.", "preconditions": ["Attacker can supply or modify external evidence", "Analyst agent A processes the supplied content"], "impact": "Incorrect findings, poisoned indicators, or unsafe agent actions that propagate into the consolidated case record.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "External evidence remains outside the trusted zone until consumed by analyst agents.", "justification": "The design explicitly accepts attacker-influenced external evidence across B1 into the analyst agent. The control fact confirms the exposure boundary, but effectiveness of handling controls is unspecified, so this is a supported exposed path."}, {"id": "T2", "category": "A", "subtype": "external evidence manipulates analyst context or tool use", "component_ids": ["C1", "C3"], "flow_ids": ["F2"], "boundary_ids": ["B1"], "asset": "Analyst agent B reasoning and outputs", "threat_vector": "Attacker-influenced incident documents, logs, or attachments are ingested by Analyst agent B and can steer its interpretation or unsafe downstream tool use.", "preconditions": ["Attacker can supply or modify external evidence", "Analyst agent B processes the supplied content"], "impact": "Incorrect findings, poisoned indicators, or unsafe agent actions that propagate into the consolidated case record.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "External evidence remains outside the trusted zone until consumed by analyst agents.", "justification": "The architecture routes the same external evidence across B1 to a second analyst agent. Because handling control effectiveness is unspecified, the path remains exposed to context or tool-use manipulation."}, {"id": "T3", "category": "I", "subtype": "sensitive evidence disclosed to both analyst agents", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "Incident documents, logs, and attachments", "threat_vector": "The same external evidence is delivered to both analyst agents, increasing the number of trusted components that can observe attacker-supplied sensitive material.", "preconditions": ["Evidence contains sensitive content", "Analyst agents are not equally constrained in retention or handling"], "impact": "Broader internal exposure of incident evidence within the trusted analysis zone.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Unspecified evidence handling controls govern analyst ingestion of external files and logs; effectiveness is not assumed.", "justification": "Two analysts each receive the full external evidence set. This creates an internal disclosure opportunity, but the architecture does not specify whether either agent can retain, forward, or expose the content, so effectiveness of limiting controls is unknown."}], "scope_notes": ["No direct unauthorized write path to the case store is shown; only the coordinator writes to C5.", "No direct analyst-to-store disclosure path is shown; persistence occurs only through the coordinator.", "Because control effectiveness is unspecified, ingestion hardening, sandboxing, and approval mechanisms cannot be treated as present or absent.", "Boundary B1 is the only cross-zone boundary in the model; internal flows among C2, C3, C4, and C5 stay within Z2."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2"], "recommendation": "Enforce strict evidence sanitization and tool-allowlisting at Analyst agent A before any parsing, retrieval, or action execution on F1 content; isolate untrusted attachments from agent instructions and disable implicit tool invocation from evidence text.", "verification": "Submit benign test evidence containing instruction-like text and crafted attachments; confirm C2 treats it as inert data, blocks prohibited tools, and only emits expected normalized findings to C4.", "existing_control_effect": "K4 keeps external evidence outside the trusted zone until C2 consumes it, and K1 indicates ingestion controls already govern the path, but their effectiveness is unspecified.", "residual_risk": "If sanitization or sandboxing is incomplete, attacker-influenced evidence may still distort analysis or trigger unsafe agent behavior."}, {"threat_id": "T2", "enforcement_component_ids": ["C3"], "recommendation": "Enforce strict evidence sanitization and tool-allowlisting at Analyst agent B before any parsing, retrieval, or action execution on F2 content; isolate untrusted attachments from agent instructions and disable implicit tool invocation from evidence text.", "verification": "Submit benign test evidence containing instruction-like text and crafted attachments; confirm C3 treats it as inert data, blocks prohibited tools, and only emits expected normalized findings to C4.", "existing_control_effect": "K4 keeps external evidence outside the trusted zone until C3 consumes it, and K1 indicates ingestion controls already govern the path, but their effectiveness is unspecified.", "residual_risk": "If sanitization or sandboxing is incomplete, attacker-influenced evidence may still distort analysis or trigger unsafe agent behavior."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Minimize duplicated exposure by redacting or partitioning sensitive evidence per analyst where feasible, and have the coordinator accept only necessary excerpts plus provenance tags before persistence to C5.", "verification": "Use a benign case bundle with known sensitive fields and confirm each analyst receives only its assigned subset, while C4 stores only approved excerpts and provenance in the case record.", "existing_control_effect": "K1 notes unspecified evidence handling controls at C2 and C3, K3 requires findings to be mediated through C4 before persistence, and K2 prevents analysts from writing directly to C5.", "residual_risk": "Both analysts still receive external evidence today, so sensitive content may remain visible to multiple internal components unless partitioning is implemented."}]}
{"id": "E006", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-05-full_workflow", "topology_family_id": "topology-006", "annotation_status": "candidate", "image_path": "images/comparison/E006.png", "title": "Support Agent with External Help Docs and Account Lookup", "purpose": "Draft customer-support replies by combining a user ticket, externally supplied help documents, and internal customer-account data within one agentic workflow.", "assumptions": ["The assistant is one connected workflow that decides when to consult external documents and the account store.", "External help documents are attacker-influenced inputs and are outside the trusted application zone unless later validated by the system.", "The customer-account store is internal to the support environment and is accessed only through the agent workflow.", "Control implementation and effectiveness are unspecified; sanitization, retrieval filtering, approval gates, and audit controls may exist in some form but their strength is not assumed.", "The attacker has no privileged access but can supply or influence external documents or evidence that the agent may ingest."], "controls": [{"id": "CTRL1", "component_ids": ["C3", "C4"], "flow_ids": ["F1", "F2", "F6"], "fact": "The orchestrator and drafting step operate as a single support workflow that can combine ticket text, retrieved help content, and account facts before producing a reply; the exact validation and grounding checks are unspecified."}, {"id": "CTRL2", "component_ids": ["C5", "C6"], "flow_ids": ["F3", "F4", "F5"], "fact": "Account lookups are routed only through the gateway to the internal store, which limits direct store access to the agent path; authorization and field filtering behavior are unspecified."}, {"id": "CTRL3", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External inputs enter the workflow from outside the trusted zone and may be incorporated into agent context; the system may apply parsing or sanitization, but effectiveness is unspecified."}], "architecture": {"components": [{"id": "C1", "label": "Customer Ticket Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Help Document Source", "zone": "Z1", "kind": "external_entity"}, {"id": "C3", "label": "Support Orchestrator", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Response Drafting", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Account Lookup Gateway", "zone": "Z2", "kind": "process"}, {"id": "C6", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C3", "data": "customer issue text and attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "external help articles and examples", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C3", "target": "C5", "data": "lookup request for customer context", "boundary_ids": []}, {"id": "F4", "source": "C5", "target": "C6", "data": "account query", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C6", "target": "C5", "data": "account status, history, and profile fields", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C3", "target": "C4", "data": "selected ticket context, retrieved help content, and account facts", "boundary_ids": []}, {"id": "F7", "source": "C4", "target": "C1", "data": "draft support response", "boundary_ids": ["B1"]}], "trust_zones": [{"id": "Z1", "label": "External Inputs"}, {"id": "Z2", "label": "Support Application"}, {"id": "Z3", "label": "Customer Data Store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt-injection via external help docs", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F2", "F6", "F7"], "boundary_ids": ["B1"], "asset": "support reply generation and agent decision-making", "threat_vector": "Attacker supplies help articles/examples that the orchestrator may ingest into context and then use in drafting a customer reply.", "preconditions": ["Attacker can influence external help documents or examples.", "The workflow may incorporate retrieved external content into the agent context without fully effective grounding or validation."], "impact": "Injected instructions could steer the support agent to generate unsafe, irrelevant, or attacker-favored replies, causing incorrect support actions or misuse of the drafting workflow.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL3"], "evidence_quote": "External inputs enter the workflow from outside the trusted zone and may be incorporated into agent context; the system may apply parsing or sanitization, but effectiveness is unspecified.", "justification": "The architecture explicitly allows external help content into the agent context and the single workflow combines it with ticket and account facts before reply generation. That creates a supported instruction-manipulation path across B1. Validation effectiveness is unspecified, so the path is exposed rather than reduced."}, {"id": "T2", "category": "I", "subtype": "overbroad account data exposure in drafted reply", "component_ids": ["C3", "C4", "C5", "C6", "C1"], "flow_ids": ["F3", "F4", "F5", "F6", "F7"], "boundary_ids": ["B2", "B1"], "asset": "customer account data", "threat_vector": "The agent requests and receives account status, history, and profile fields, then includes selected account facts in the outgoing support response.", "preconditions": ["The workflow chooses to query the account store for a given ticket.", "The drafting step may include account facts in the response without explicit field minimization or disclosure checks."], "impact": "Sensitive account information could be disclosed to the ticket initiator or other unintended recipients through the generated reply.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2"], "evidence_quote": "Account lookups are routed only through the gateway to the internal store, which limits direct store access to the agent path; authorization and field filtering behavior are unspecified.", "justification": "The design explicitly returns account status, history, and profile fields from Z3 to the agent, and the draft reply is sent back to the external ticket source. Because field filtering and authorization are unspecified, unauthorized or excessive disclosure remains a feasible path across B2 and B1."}, {"id": "T3", "category": "S", "subtype": "external ticket spoofing", "component_ids": ["C1", "C3", "C4"], "flow_ids": ["F1", "F6", "F7"], "boundary_ids": ["B1"], "asset": "support workflow trust in ticket origin", "threat_vector": "An attacker submits a fabricated ticket that appears to be a legitimate customer issue and drives the agent to produce a response based on false premises.", "preconditions": ["Attacker can send tickets into the external intake channel.", "The workflow does not have a specified strong origin-authentication or identity-binding control for incoming tickets."], "impact": "The agent may act on a spoofed request, leading to incorrect support handling or disclosure in a reply that should not have been generated for that claimant.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "External inputs enter the workflow from outside the trusted zone and may be incorporated into agent context; the system may apply parsing or sanitization, but effectiveness is unspecified.", "justification": "The architecture shows ticket input entering from Z1 to the support agent, but it does not specify authentication of the requester or ticket provenance. That leaves spoofing plausible, though the control state is unspecified rather than clearly weak enough to mark exposed."}], "scope_notes": ["Boundary IDs inferred from the supplied flows: B1 covers Z1<->Z2 traffic on F1, F2, and F7; B2 covers Z2<->Z3 traffic on F4 and F5.", "No direct unauthorized read from C6 alone was treated as disclosure; disclosure is only modeled where data can plausibly reach the external reply path.", "No separate repudiation threat was added because logging/audit behavior is not described.", "Potential prompt-injection and data-exposure paths depend on unspecified validation, sanitization, approval, and field-filtering controls; unsupported paths remain incomplete in scope."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C3", "C4"], "recommendation": "Enforce strict separation between retrieved help content and executable agent instructions: mark external documents as untrusted, strip or ignore instruction-like text before it enters the orchestrator context, and require the drafting step to use only ticket facts plus approved retrieval snippets from a constrained schema.", "verification": "Run a benign test with a help document containing obvious instruction-like text and confirm the workflow treats it as inert content, logs the rejection/redaction, and still drafts a reply based only on allowed fields.", "existing_control_effect": "CTRL1 and CTRL3 already indicate the workflow can combine external content with drafting, but validation effectiveness is unspecified; this means the current design may provide some parsing/sanitization without reliable instruction isolation.", "residual_risk": "If sanitization or grounding is incomplete, a crafted document may still bias content selection or phrasing, though the attack surface is reduced."}, {"threat_id": "T2", "enforcement_component_ids": ["C5", "C4"], "recommendation": "Constrain account retrieval to least-privilege fields and require the drafting component to pass account facts through a disclosure filter that redacts sensitive or unnecessary fields before any reply is emitted to the customer channel.", "verification": "Use a test ticket that triggers an account lookup and confirm only the minimum approved fields are returned to drafting and that any nonessential profile or history values are removed before the response is sent.", "existing_control_effect": "CTRL2 already routes access only through the gateway to the internal store, which prevents direct store access; however, authorization and field filtering are unspecified, so the existing control mainly narrows the path rather than preventing over-disclosure.", "residual_risk": "If business logic still needs account context to answer a ticket, some sensitive derivation risk remains unless the response filter is consistently enforced."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C3"], "recommendation": "Bind each ticket to a verifiable requester identity or session token at intake and have the orchestrator reject or defer responses when the origin cannot be matched to an authenticated customer record.", "verification": "Submit a benign test ticket without valid identity binding and confirm the system flags it as unverified and does not generate a customer-specific response until the origin is confirmed.", "existing_control_effect": "CTRL3 shows tickets enter from outside the trusted zone, but it does not specify origin authentication; current controls may parse or sanitize content yet do not clearly establish requester identity.", "residual_risk": "If identity binding is weak or optional, an attacker may still submit plausible tickets that are handled as legitimate, though stronger intake checks would make this less likely."}]}
{"id": "E007", "collection": "comparison", "domain": "finance_operations", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-11-full_workflow", "topology_family_id": "topology-007", "annotation_status": "candidate", "image_path": "images/comparison/E007.png", "title": "Invoice-to-Payment Drafting Flow with External Document Intake", "purpose": "An invoice-handling agent ingests externally supplied invoice documents and supporting evidence, extracts billing details, drafts payment requests, and forwards them into an approval and payment-execution path.", "assumptions": ["Invoice documents and evidence originate outside the trusted application zone and may be attacker-influenced.", "The invoice agent can read external documents and can create or populate payment drafts from extracted content.", "The approval and payment-execution path exists as a separate downstream function, but the implementation and effectiveness of its controls are unspecified.", "The attacker has no privileged access to internal systems, but can shape document contents, metadata, or attached evidence submitted through normal external channels.", "No additional compensating controls are assumed beyond the described functional flow.", "The invoice agent operates with authority to populate draft payment data but not to authorize final payment execution."], "controls": [{"id": "K1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "External submissions are accepted through the supplier portal into the invoice agent; content validation and document-authenticity checks are present in design, but their effectiveness is unspecified."}, {"id": "K2", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "fact": "The agent writes extracted invoice data to the document store for later traceability; storage access is limited to the invoice workflow, but enforcement strength is unspecified."}, {"id": "K3", "component_ids": ["C2", "C4"], "flow_ids": ["F3"], "fact": "Payment drafts are generated by the invoice agent and passed to the draft service; the agent cannot directly execute payments, but the downstream guardrails are unspecified."}, {"id": "K4", "component_ids": ["C4", "C5", "C6"], "flow_ids": ["F4", "F5", "F6"], "fact": "Approval and execution are separated into distinct downstream stages; review and execution gating exist as workflow steps, but their control effectiveness is unspecified."}], "architecture": {"components": [{"id": "C1", "label": "Supplier Upload Portal", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Document Store", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Payment Draft Service", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Approval Queue", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "Invoice document and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "Parsed invoice content and extracted metadata", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C4", "data": "Billing details and payment draft fields", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "Draft payment request for review", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C6", "data": "Approved payment request", "boundary_ids": []}, {"id": "F6", "source": "C6", "target": "C5", "data": "Execution status and reconciliation result", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Intake"}, {"id": "Z2", "label": "Trusted Application"}, {"id": "Z3", "label": "Approval and Execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "Invoice content tampering leading to fraudulent draft payment", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "Draft payment request", "threat_vector": "Attacker submits a crafted invoice/document set through the supplier portal so the agent extracts manipulated billing details and populates a payment draft with attacker-chosen fields.", "preconditions": ["Attacker can submit documents through the normal external channel.", "Invoice contents or metadata are attacker-influenced.", "The agent uses extracted content to create draft payment data."], "impact": "Fraudulent or incorrect payment draft values may be created and forwarded into the approval path.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "External submissions are accepted through the supplier portal into the invoice agent; content validation and document-authenticity checks are present in design, but their effectiveness is unspecified.", "justification": "The architecture explicitly accepts external invoice content across B1 and uses the agent to populate payment drafts from that content. The listed validation/authenticity checks are present, but effectiveness is unspecified, so the design remains exposed rather than reduced."}, {"id": "T2", "category": "T", "subtype": "Stored invoice data tampering or integrity corruption", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "boundary_ids": [], "asset": "Document store records", "threat_vector": "Manipulated extracted metadata or parsing results are written into the document store, corrupting the traceability record used by the invoice workflow.", "preconditions": ["The attacker can influence the invoice content that the agent parses.", "The agent persists extracted data to the store as part of the workflow."], "impact": "Stored invoice records may no longer reflect the original submission, affecting later review, reconciliation, or audit use.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "The agent writes extracted invoice data to the document store for later traceability; storage access is limited to the invoice workflow, but enforcement strength is unspecified.", "justification": "There is a write path from the agent to the store, but the available fact describes a limited workflow scope rather than an effective integrity control. The actual enforcement strength is unspecified, so impact on stored records is possible but not enough to mark the path exposed."}, {"id": "T3", "category": "A", "subtype": "Instruction/context manipulation via invoice content affecting agent output", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "Invoice agent behavior and draft generation", "threat_vector": "A malicious invoice or attached evidence is crafted to steer the agent's parsing/extraction behavior or downstream drafting decisions through attacker-controlled content.", "preconditions": ["Attacker can shape document contents, metadata, or attached evidence submitted through normal external channels.", "The agent processes that content as input to extraction and drafting."], "impact": "The agent may be induced to produce unsafe or attacker-favored draft payment fields, causing incorrect downstream payment handling.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "External submissions are accepted through the supplier portal into the invoice agent; content validation and document-authenticity checks are present in design, but their effectiveness is unspecified.", "justification": "This is an agent-specific input-manipulation path from the external zone into the trusted application zone. Because validation/authenticity checks are only described as present with unspecified effectiveness, the attack path is supported and not shown to be effectively constrained."}, {"id": "T4", "category": "T", "subtype": "Draft request tampering before approval", "component_ids": ["C4", "C5"], "flow_ids": ["F4"], "boundary_ids": ["B2"], "asset": "Draft payment request", "threat_vector": "A compromised or manipulated draft request is forwarded from the draft service into the approval queue with altered payment details.", "preconditions": ["The invoice agent can populate draft payment data.", "The downstream review gate does not reliably detect or reject altered draft fields."], "impact": "An attacker-influenced payment request could reach approval review with tampered content.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Approval and execution are separated into distinct downstream stages; review and execution gating exist as workflow steps, but their control effectiveness is unspecified.", "justification": "The path from the trusted application zone into the approval zone exists, but the available control fact only says gating exists and does not establish whether tampering is prevented. That makes this a plausible but not confirmed exposed path."}], "scope_notes": ["Boundary IDs assigned from flow order: B1 for Z1-Z2, B2 for Z2-Z3.", "Controls mention validation, storage limitation, and workflow gating, but their effectiveness is explicitly unspecified; this prevents stronger reduction claims.", "No direct unauthorized read path from the document store is shown, so information disclosure is not separately asserted here.", "Approval and execution controls are downstream and unspecified; unsupported escalation or bypass paths were not assumed.", "Potential threats not included: payment execution compromise beyond the described draft path, because no concrete bypass or privilege path is provided."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C4"], "recommendation": "Enforce strict invoice authenticity, schema, and field-level consistency checks at the supplier portal and invoice agent before any billing detail is written into the payment draft; reject or quarantine documents that fail checks, and require manual review for high-risk changes to payee, amount, or bank fields.", "verification": "Submit benign test invoices with altered totals, payee names, or mismatched attachments and confirm they are blocked or routed to review before F3, with no draft created from untrusted fields.", "existing_control_effect": "K1 and K3 indicate validation/authenticity checks and draft generation separation are present, but their effectiveness is unspecified, so they may reduce accidental errors yet do not establish reliable prevention.", "residual_risk": "If validation misses crafted content or authenticity signals are weak, fraudulent draft values can still be generated and forwarded into approval."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Apply integrity protection on stored invoice records, including append-only audit logging, write authorization tied to workflow state, and record hashing or signing so later changes to parsed invoice data are detectable.", "verification": "Write a normal invoice record, then perform an authorized update through the workflow and confirm the system records a new version or integrity event rather than silently overwriting the original.", "existing_control_effect": "K2 shows the agent writes extracted data to the document store for traceability and that storage access is limited to the invoice workflow, which helps constrain who can write, but enforcement strength is unspecified.", "residual_risk": "Stored records may still be corrupted if workflow-scoped writes are accepted without strong integrity checks or if malformed parsed data is persisted as authoritative."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2"], "recommendation": "Harden the invoice agent against attacker-controlled document content by separating raw document text from instructions, using deterministic extraction rules, and rejecting documents that contain anomalous or conflicting control data before extraction and drafting.", "verification": "Provide benign sample invoices containing unusual formatting or embedded non-invoice text and confirm the agent extracts only allowed invoice fields and does not alter drafting behavior based on embedded instructions.", "existing_control_effect": "K1 indicates validation and document-authenticity checks exist at intake, which may filter some malformed submissions, but the control strength is not established.", "residual_risk": "A crafted invoice or attachment may still steer extraction or drafting if the checks miss adversarial content or if parsing logic treats attacker-supplied text as authoritative."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5", "C6"], "recommendation": "Require the draft service to sign or seal payment drafts, enforce immutable review metadata at the approval queue, and validate that only approved, unmodified drafts can enter F5 for execution.", "verification": "Create a benign draft, modify a non-authorized field after draft creation, and confirm the approval queue or executor rejects it and records the rejection without reaching payment execution.", "existing_control_effect": "K4 states that approval and execution are separated into distinct downstream stages with review and execution gating, which provides workflow separation but does not show effective tamper detection.", "residual_risk": "If downstream gating only checks stage presence and not draft integrity, a manipulated request may still reach approval review or be executed after unauthorized changes."}]}
{"id": "E008", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-04-full_workflow", "topology_family_id": "topology-008", "annotation_status": "candidate", "image_path": "images/comparison/E008.png", "title": "Guarded support-assist workflow with external evidence and account lookup", "purpose": "A customer-support assistant drafts replies using externally supplied help documents and a customer-account store, while enforcing authorization, provenance checks, and bounded tool access.", "assumptions": ["The assistant is connected to a ticketing front end that accepts customer messages and uploaded evidence.", "Help documents may originate outside the trusted zone and can be attacker-influenced.", "The account store is inside a separate trusted zone and requires authenticated, least-privilege read access.", "Tool use is bounded to retrieval and response drafting; the model cannot directly modify account records.", "Document provenance checks and content filtering are enforced, but their effectiveness may be incomplete for adversarially crafted text.", "Attacker capability: an external customer or uploader can submit crafted messages and evidence, but has no privileged access to the account store or internal configuration."], "controls": [{"id": "CTL1", "component_ids": ["C3"], "flow_ids": ["F2", "F3"], "fact": "Uploaded evidence is passed through a provenance and content filter before the assistant receives it; untrusted documents are not treated as authoritative without passing this gate."}, {"id": "CTL2", "component_ids": ["C5", "C6"], "flow_ids": ["F4", "F5", "F6", "F7"], "fact": "Account lookup occurs only through a retrieval gate that issues authenticated, least-privilege read queries to the protected store and returns entitlement-scoped results only."}, {"id": "CTL3", "component_ids": ["C4", "C5", "C7"], "flow_ids": ["F4", "F7", "F8"], "fact": "The assistant is limited to read-only retrieval and draft generation; it cannot directly write account records or execute arbitrary backend actions."}, {"id": "CTL4", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "The intake path accepts external messages and evidence, but the control stack treats them as potentially hostile inputs and applies screening before they reach the agent workspace."}], "architecture": {"components": [{"id": "C1", "label": "Ticket front end", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Evidence uploader", "zone": "Z1", "kind": "process"}, {"id": "C3", "label": "Document filter", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Support assistant", "zone": "Z2", "kind": "agent"}, {"id": "C5", "label": "Retrieval gate", "zone": "Z2", "kind": "process"}, {"id": "C6", "label": "Customer account store", "zone": "Z3", "kind": "data_store"}, {"id": "C7", "label": "Draft response service", "zone": "Z2", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "customer message, uploaded evidence", "boundary_ids": []}, {"id": "F2", "source": "C2", "target": "C3", "data": "evidence package and document metadata", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C3", "target": "C4", "data": "filtered help document text, provenance flags", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "account lookup request", "boundary_ids": []}, {"id": "F5", "source": "C5", "target": "C6", "data": "authenticated read query", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C6", "target": "C5", "data": "account attributes, entitlement-scoped results", "boundary_ids": ["B2"]}, {"id": "F7", "source": "C5", "target": "C4", "data": "approved account summary", "boundary_ids": []}, {"id": "F8", "source": "C4", "target": "C7", "data": "draft reply content and cited evidence", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "Untrusted intake"}, {"id": "Z2", "label": "Guarded agent workspace"}, {"id": "Z3", "label": "Protected account store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction injection via uploaded evidence", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "support assistant behavior and reply integrity", "threat_vector": "Attacker submits crafted customer message or evidence that survives filtering and reaches the assistant as document text/provenance context.", "preconditions": ["Attacker can upload or submit evidence through the ticket front end.", "Document filtering/provenance checks may be incomplete for adversarially crafted text."], "impact": "The assistant may follow attacker-supplied instructions embedded in the evidence, leading to unsafe agent-tool use or manipulated draft responses.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTL1", "CTL4"], "evidence_quote": "Uploaded evidence is passed through a provenance and content filter before the assistant receives it; untrusted documents are not treated as authoritative without passing this gate.", "justification": "There is an explicit path from untrusted intake across B1 into the guarded agent workspace, and the architecture states external help documents can be attacker-influenced. The control screens the input but also notes incomplete effectiveness against adversarial text, so this remains a supported exposure rather than a removed path."}, {"id": "T2", "category": "I", "subtype": "unauthorized account attribute disclosure through bounded retrieval", "component_ids": ["C4", "C5", "C6", "C7"], "flow_ids": ["F4", "F5", "F6", "F7", "F8"], "boundary_ids": ["B2"], "asset": "customer account attributes", "threat_vector": "A crafted support request induces the assistant to query account data and expose more account detail in the draft response than intended by the requester’s entitlement.", "preconditions": ["The assistant can issue account lookup requests through the retrieval gate.", "The response drafting step may include cited evidence or summary details from retrieved account attributes."], "impact": "Sensitive account attributes could be disclosed to the ticket requester in the drafted reply.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL2", "CTL3"], "evidence_quote": "Account lookup occurs only through a retrieval gate that issues authenticated, least-privilege read queries to the protected store and returns entitlement-scoped results only.", "justification": "A read path exists across B2 from the agent workspace to the protected store, but the control explicitly scopes and authenticates access. That reduces exposure, yet the exact data-minimization of the returned summary and drafting behavior is not fully specified, so unauthorized disclosure is possible but not established."}, {"id": "T3", "category": "D", "subtype": "resource exhaustion via repeated retrieval and drafting", "component_ids": ["C1", "C2", "C4", "C5", "C7"], "flow_ids": ["F1", "F2", "F4", "F5", "F7", "F8"], "boundary_ids": ["B1", "B2"], "asset": "support workflow availability", "threat_vector": "An external customer repeatedly submits messages/evidence that drive repeated filtering, retrieval, and draft generation work.", "preconditions": ["Attacker can repeatedly use the ticket front end.", "No rate-limit or capacity control is described for intake, retrieval, or drafting."], "impact": "The assistant workflow may be slowed or degraded, delaying legitimate support handling.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL1", "CTL2", "CTL4"], "evidence_quote": "The intake path accepts external messages and evidence, but the control stack treats them as potentially hostile inputs and applies screening before they reach the agent workspace.", "justification": "The architecture clearly accepts untrusted external input and performs multiple downstream processing steps, which creates a plausible workload-amplification path. However, no explicit throttling, quotas, or backpressure controls are described, so effectiveness against exhaustion is unspecified rather than clearly exposed."}], "scope_notes": ["Boundary IDs derived from the supplied zones and flows: B1 covers Z1<->Z2, B2 covers Z2<->Z3.", "Unsupported or incomplete paths not analyzed as separate threats include direct modification of account records, because the design states the model cannot directly write account records.", "Same-zone flows F3, F4, F7, and F8 do not cross a trust boundary; they were considered only as part of end-to-end paths.", "The exact strength of document filtering against prompt injection and the degree of redaction in entitlement-scoped results are unspecified.", "No explicit rate limiting, abuse detection, or auditing controls were provided for repeated intake or retrieval abuse."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C3", "C4"], "recommendation": "Harden the document filter to strip or quarantine instruction-like content, require provenance labels to survive into the workspace, and configure the support assistant to treat all uploaded evidence as untrusted reference material that cannot change system/task policy.", "verification": "Submit benign test documents containing embedded instructions and check that they are flagged, down-ranked, or excluded from the assistant context while ordinary evidence still passes; confirm the assistant cites provenance and does not adopt document-authored instructions.", "existing_control_effect": "CTL1 and CTL4 already place a provenance/content filter in front of the agent workspace, so the attack is screened before reaching the assistant, but the scenario notes incomplete effectiveness against adversarial text.", "residual_risk": "Residual risk remains that carefully crafted text may still evade filtering or influence the assistant's behavior if downstream prompt handling is too permissive."}, {"threat_id": "T2", "enforcement_component_ids": ["C5", "C6", "C7"], "recommendation": "Enforce per-request entitlement checks in the retrieval gate, return only the minimum account fields needed for the specific support task, and redact account identifiers or sensitive attributes from the draft-response input unless explicitly required.", "verification": "Use a benign test account and request paths with different entitlements, then confirm the gate returns only approved fields and the draft service cannot echo suppressed attributes into the reply.", "existing_control_effect": "CTL2 already requires authenticated, least-privilege reads with entitlement-scoped results, and CTL3 limits the assistant to read-only retrieval and draft generation.", "residual_risk": "If the retrieved summary is still too broad or the draft service re-exposes approved data verbatim, some sensitive information may be disclosed within the requester's entitlement scope or through mistaken over-sharing."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C5", "C7"], "recommendation": "Add rate limiting, quota controls, and per-ticket backpressure on intake, retrieval, and draft generation so repeated submissions cannot trigger unbounded processing; log and suppress duplicate or abusive requests before they fan out to the agent workspace and store.", "verification": "Run a benign load test with repeated normal tickets and confirm requests are throttled or queued after configured thresholds while legitimate traffic continues to complete within expected service limits.", "existing_control_effect": "CTL1, CTL2, and CTL4 already screen hostile input and constrain retrieval, which reduces some abuse, but they do not describe capacity or frequency limits.", "residual_risk": "A determined attacker may still cause localized slowdowns within allowed quotas or by distributing requests across many sessions if abuse detection is weak."}]}
{"id": "E009", "collection": "comparison", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-10-no_reviewer", "topology_family_id": "topology-009", "annotation_status": "candidate", "image_path": "images/comparison/E009.png", "title": "Guarded invoice-to-payment agent with approval gate", "purpose": "An invoice-handling AI reads externally supplied invoice documents, extracts payment-relevant fields, prepares payment drafts, and passes them to a separate approval and execution path under constrained tool authority.", "assumptions": ["Invoice documents arrive from outside the trusted application zone and may be attacker-influenced.", "The invoice agent has bounded read-only access to documents and a limited drafting tool, but cannot execute payments directly.", "Approval and payment execution are separated from invoice intake by an explicit authorization step.", "Provenance checks and schema validation exist, but their effectiveness against crafted content is not fully specified.", "No privileged internal access is granted to the attacker; influence is limited to supplied documents/evidence."], "controls": [{"id": "K1", "component_ids": ["C2"], "flow_ids": ["F1", "F2"], "fact": "The invoice agent is configured for read-only document access and can only produce draft output, not settle payments."}, {"id": "K2", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "Incoming invoices are subjected to provenance checks and schema validation before extraction, though resistance to crafted content is unspecified."}, {"id": "K3", "component_ids": ["C4", "C5"], "flow_ids": ["F3", "F4"], "fact": "Payment execution requires a separate approval gate; drafts in the queue do not authorize settlement without review."}, {"id": "K4", "component_ids": ["C2", "C3"], "flow_ids": ["F2", "F3"], "fact": "Drafts are handed off through an internal queue, limiting direct tool authority from the agent to the executor."}], "architecture": {"components": [{"id": "C1", "label": "Invoice Inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Draft Review Queue", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Approval Gate", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "extracted fields and payment draft", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "draft package for review", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Source"}, {"id": "Z2", "label": "Guarded Agent Zone"}, {"id": "Z3", "label": "Approval and Payments"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "crafted-invoice-manipulates-extraction-or-draft", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "invoice intake and draft preparation", "threat_vector": "Attacker-controlled invoice content uses crafted text or embedded evidence to tamper with the agent's extraction and draft generation.", "preconditions": ["Attacker can supply invoice documents from Z1.", "Crafted content is able to influence parsing or extraction logic."], "impact": "Incorrect fields or payment drafts may be produced and queued for later review.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "Incoming invoices are subjected to provenance checks and schema validation before extraction, though resistance to crafted content is unspecified.", "justification": "The path from external invoices to agent extraction and draft storage is explicit, and the control statement leaves crafted-content resistance unspecified rather than limiting the path."}, {"id": "T2", "category": "A", "subtype": "prompt-or-context-injection-via-invoice-content", "component_ids": ["C1", "C2"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "invoice agent instructions and drafting behavior", "threat_vector": "Malicious invoice content attempts to manipulate the agent's context, instructions, or tool-use choices during processing.", "preconditions": ["Attacker controls or influences the supplied invoice content.", "The agent processes document text that can affect its internal reasoning or drafting output."], "impact": "The agent may follow attacker-chosen instructions, alter draft content, or mishandle invoice processing.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "The invoice agent is configured for read-only document access and can only produce draft output, not settle payments.", "justification": "Read-only access prevents direct payment execution, but the architecture still allows external documents to reach the agent for extraction and drafting; the effect of provenance/schema controls on crafted content is unspecified, so the manipulation path remains a supported exposure."}, {"id": "T3", "category": "I", "subtype": "draft-content-exposure-within-internal-hand-off", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F2", "F3"], "boundary_ids": ["B2"], "asset": "payment draft contents and supporting invoice data", "threat_vector": "Sensitive invoice-derived fields in drafts are passed from the agent zone into the approval zone, where unauthorized exposure could occur if review handling is weak.", "preconditions": ["Drafts contain sensitive payment-relevant data.", "An unauthorized recipient or weak handling path exists in or around review."], "impact": "Invoice details or payment terms may be disclosed beyond the intended reviewer set.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K4", "K3"], "evidence_quote": "Drafts are handed off through an internal queue, limiting direct tool authority from the agent to the executor.", "justification": "The architecture shows an internal handoff of draft packages across B2, but it does not specify unauthorized recipients or inadequate review handling; therefore disclosure potential is possible but not established."}, {"id": "T4", "category": "D", "subtype": "review-or-queue-delay-blocks-payment-processing", "component_ids": ["C3", "C4", "C5"], "flow_ids": ["F3", "F4"], "boundary_ids": ["B2"], "asset": "payment execution availability", "threat_vector": "Queued drafts or the approval step may be overwhelmed, delayed, or stalled, preventing timely progression to execution.", "preconditions": ["Draft volume or review workload is high.", "Queue or approval handling can be slowed or blocked."], "impact": "Payments may be delayed or not executed when expected.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Payment execution requires a separate approval gate; drafts in the queue do not authorize settlement without review.", "justification": "The gate prevents direct settlement from drafts, which reduces immediate misuse, but the architecture does not specify queue capacity, review latency, or resilience to overload; denial-of-service is therefore not confirmed but remains a plausible path."}], "scope_notes": ["Boundary IDs assigned from the supplied flow order: B1 for Z1-Z2, B2 for Z2-Z3.", "No direct payment execution path exists from the invoice agent; F4 requires approval via C4.", "Controls mention provenance checks and schema validation, but their effectiveness against crafted content is explicitly unspecified.", "No attacker access to internal zones or privileged internal access is assumed beyond influence over supplied invoices.", "No candidate threat is asserted for repudiation or elevation of privilege because the architecture text does not provide a concrete unsupported mechanism."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2"], "recommendation": "Harden the invoice intake pipeline to reject or neutralize crafted invoice content before extraction: enforce strict canonical schema validation, content sanitization, attachment-type allowlists, and deterministic field parsing with fail-closed handling for malformed or ambiguous inputs.", "verification": "Submit benign test invoices that include harmless but malformed fields, duplicate keys, and embedded metadata; confirm they are rejected or normalized before C2 produces a draft, and that no draft is queued when validation fails.", "existing_control_effect": "K2 already provides provenance checks and schema validation at F1, which helps screen some nonconforming inputs, but resistance to crafted content is unspecified.", "residual_risk": "A sufficiently well-formed malicious invoice may still influence extraction logic if parser assumptions or model behavior remain exploitable."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C1"], "recommendation": "Constrain agent instruction handling by separating invoice content from system prompts, stripping or quoting untrusted text, and using an allowlisted extraction template that ignores imperative language found in documents; keep the agent in draft-only mode with no direct action privileges.", "verification": "Use harmless invoices containing injected instructions such as requests to ignore fields or alter output, then confirm C2 treats them as data only, preserves the extraction template, and never converts document text into tool directives or execution steps.", "existing_control_effect": "K1 already limits C2 to read-only document access and draft generation, which blocks direct payment execution, and K2 adds intake checks; however, those controls do not by themselves guarantee immunity to instruction manipulation.", "residual_risk": "The agent may still be steered into producing incorrect drafts or biased field selections if injected content affects reasoning despite the read-only constraint."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C4"], "recommendation": "Protect the draft handoff with access control and data minimization: restrict C3 read access to the approval role, redact nonessential invoice details in the draft package, and log all queue reads and review actions for authorized reviewers only.", "verification": "Perform a benign review workflow test with a sample draft and confirm only designated approvers can retrieve it from C3, redacted fields are omitted where not needed, and access logs record each read and approval step.", "existing_control_effect": "K4 and K3 already separate the queue from execution and require approval before settlement, which limits direct misuse; they do not explicitly address who can read queued draft contents.", "residual_risk": "Authorized reviewers may still see sensitive invoice data, and misconfiguration of queue permissions could expose drafts to unintended internal readers."}, {"threat_id": "T4", "enforcement_component_ids": ["C3", "C4", "C5"], "recommendation": "Add operational backpressure controls for the queue and approval workflow: set queue depth alerts, reviewer SLA timers, retry limits, and safe degradation rules so stalled or overfull review traffic does not block normal payment execution indefinitely.", "verification": "Run a benign load test with synthetic drafts to confirm alerts fire at configured thresholds, excessive backlog is surfaced to operators, and older drafts are handled according to the defined timeout or escalation policy without freezing C5 indefinitely.", "existing_control_effect": "K3 ensures drafts cannot bypass approval, and K4 keeps execution separated from the agent; this reduces direct abuse but does not specify capacity or latency protections.", "residual_risk": "Large spikes in draft volume or reviewer absence can still delay payments even if the system remains secure, especially if operational monitoring is not actively staffed."}]}
{"id": "E010", "collection": "comparison", "domain": "finance_operations", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-12-no_reviewer", "topology_family_id": "topology-010", "annotation_status": "candidate", "image_path": "images/comparison/E010.png", "title": "Invoice intake to draft payment approval pipeline", "purpose": "Model an invoice-processing flow where an agent ingests externally supplied invoice documents, extracts payment details, prepares payment drafts, and forwards them into an approval and payment-execution path.", "assumptions": ["An external sender can influence invoice contents, attachments, or supporting evidence, but has no privileged internal access.", "The invoice agent can read external documents and produce draft payment artifacts, but the effectiveness of any validation, sandboxing, approval checks, or execution controls is unspecified.", "The approval and payment-execution steps are part of one connected functional path within the finance operation.", "The architecture is limited to the invoice intake, draft preparation, approval handoff, and payment execution flow; no extra supporting services are assumed beyond what is needed for that path."], "controls": [{"id": "CTL1", "component_ids": ["C1", "C2"], "flow_ids": ["F1", "F2"], "fact": "Invoice intake passes externally supplied documents into the finance workflow; document validation and agent-side filtering are present as configured behavior, but their effectiveness is unspecified."}, {"id": "CTL2", "component_ids": ["C3", "C4"], "flow_ids": ["F3", "F4"], "fact": "Draft payment records are handed from the invoice agent to the approval step through the connected finance path; review depth and approval enforcement are unspecified."}, {"id": "CTL3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "Payment execution occurs only after the approval handoff in this workflow; the strength of authorization checks at execution time is unspecified."}], "architecture": {"components": [{"id": "C1", "label": "Invoice intake", "zone": "Z2", "kind": "process"}, {"id": "C2", "label": "Invoice agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Draft payment record", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Approval handoff", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment executor", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "External sender", "zone": "Z1", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C6", "target": "C1", "data": "invoice documents, attachments, supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C2", "data": "ingested invoice content and metadata", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C3", "data": "extracted payment details, draft invoice fields", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "payment draft for approval review", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F6", "source": "C5", "target": "C3", "data": "payment status, execution result", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External sender"}, {"id": "Z2", "label": "Finance operations"}, {"id": "Z3", "label": "Approval and payment execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "invoice-content tampering leading to poisoned draft payment", "component_ids": ["C6", "C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1", "B2"], "asset": "draft payment record and downstream payment instruction", "threat_vector": "Attacker-controlled invoice documents, attachments, or supporting evidence alter the extracted payment details used to create a draft and ultimately influence approval and execution.", "preconditions": ["External sender can influence invoice contents, attachments, or supporting evidence.", "Validation, filtering, and approval enforcement are unspecified in effectiveness."], "impact": "Incorrect or manipulated payment details may be reviewed, approved, and executed, causing financial loss or fraudulent payment.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTL1", "CTL2", "CTL3"], "evidence_quote": "Invoice intake passes externally supplied documents into the finance workflow; document validation and agent-side filtering are present as configured behavior, but their effectiveness is unspecified.", "justification": "The path from external invoice input through agent extraction to draft creation and approval is explicit, and the controls do not specify effective prevention of content manipulation."}, {"id": "T2", "category": "S", "subtype": "spoofed invoice sender identity or legitimacy", "component_ids": ["C6", "C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "invoice intake and draft payment record", "threat_vector": "A malicious external party submits invoice material that appears to originate from a legitimate sender or business context, influencing the agent to process it as trusted input.", "preconditions": ["Attacker can supply invoice documents and related evidence.", "No explicit sender-authentication or provenance check is described."], "impact": "The finance workflow may accept and process fraudulent invoices, producing draft payment records based on false legitimacy.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL1"], "evidence_quote": "Invoice intake passes externally supplied documents into the finance workflow; document validation and agent-side filtering are present as configured behavior, but their effectiveness is unspecified.", "justification": "The architecture clearly accepts external documents across B1, but it does not specify whether provenance or sender authentication is effective enough to block spoofed submissions."}, {"id": "T3", "category": "I", "subtype": "invoice and payment-data disclosure", "component_ids": ["C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F2", "F3", "F4", "F6"], "boundary_ids": ["B2"], "asset": "invoice content, payment details, and execution status", "threat_vector": "Invoice content and payment status travel within the finance and approval path, creating a potential disclosure path if unauthorized internal recipients or weak access controls exist.", "preconditions": ["An attacker can access an internal component or intercept internal workflow data.", "No explicit internal access-control or confidentiality enforcement is described."], "impact": "Sensitive invoice and payment information could be exposed to unauthorized parties inside the workflow.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL2", "CTL3"], "evidence_quote": "Draft payment records are handed from the invoice agent to the approval step through the connected finance path; review depth and approval enforcement are unspecified.", "justification": "The flows carry sensitive data across B2, but the diagram does not identify an unauthorized recipient or specify whether confidentiality controls prevent disclosure."}, {"id": "T4", "category": "D", "subtype": "approval or execution disruption via malformed invoice inputs", "component_ids": ["C6", "C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5", "F6"], "boundary_ids": ["B1", "B2"], "asset": "invoice-processing availability", "threat_vector": "An external sender can submit large, malformed, or repetitive invoice material that burdens intake, extraction, approval handoff, or execution steps.", "preconditions": ["External sender can influence invoice contents, attachments, or supporting evidence.", "Resource limits, validation robustness, and operational throttling are unspecified."], "impact": "The pipeline may slow down, stall, or become unable to process legitimate invoices and payment records.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL1", "CTL2", "CTL3"], "evidence_quote": "Invoice intake passes externally supplied documents into the finance workflow; document validation and agent-side filtering are present as configured behavior, but their effectiveness is unspecified.", "justification": "The architecture supports a continuous workflow from external input through execution, but the effectiveness of controls against load or malformed-input disruption is not stated."}], "scope_notes": ["Boundary IDs inferred from the listed flows only: B1 for Z1↔Z2 and B2 for Z2↔Z3.", "No explicit authentication, authorization, logging, nonrepudiation, or sandboxing mechanism is described; related impacts remain uncertain.", "A data-store read or internal data flow alone does not establish unauthorized disclosure; recipient misuse is not shown.", "No unsupported agent-tool path or memory-injection mechanism is described beyond invoice content influence.", "The architecture does not specify whether the approval handoff is a human review, automated gate, or both, limiting stronger claims about enforcement."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C4", "C5"], "recommendation": "Enforce strict invoice schema, provenance, and attachment validation at intake and again before approval; require a human approval step that compares the draft against the original source document and blocks any payment instruction that lacks a matching approved invoice identifier, payee, amount, and supporting evidence hash.", "verification": "Submit benign test invoices with altered amounts, changed payees, mismatched attachments, and duplicate invoice IDs; confirm they are rejected or quarantined at intake or before approval, and that only unchanged, approved drafts can reach execution.", "existing_control_effect": "CTL1 already provides intake validation/filtering, and CTL2/CTL3 already place a review and execution step in the path, but their effectiveness is unspecified so they may only partially reduce the poisoning path.", "residual_risk": "A valid-looking but fraudulent invoice that satisfies the validation rules may still reach review if the approval process is weak or if business rules allow incorrect but well-formed requests."}, {"threat_id": "T2", "enforcement_component_ids": ["C1", "C2"], "recommendation": "Add sender authentication and provenance checks at invoice intake, such as allowlisted sender identifiers, signed documents, or verified payment-source metadata, and reject or quarantine any invoice that cannot be traced to an approved sender.", "verification": "Use benign invoices from approved and unapproved senders, plus replayed or unsigned samples, and confirm only authenticated, provenance-matched submissions are processed into drafts.", "existing_control_effect": "CTL1 already indicates validation and agent-side filtering at intake, which may reduce obvious spoofing, but no explicit sender-authentication control is described.", "residual_risk": "A spoofed invoice may still pass if the attacker can imitate legitimate formatting or if provenance signals are unavailable or inconsistently checked."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C4", "C5"], "recommendation": "Restrict access to draft payment records and execution status to least-privilege roles, encrypt internal workflow data in transit and at rest, and log all reads and modifications across the approval and execution path with periodic review of access events.", "verification": "Perform benign access checks from authorized and unauthorized test identities, confirm unauthorized reads are denied, and confirm logs record each approved access and status update.", "existing_control_effect": "CTL2 and CTL3 already indicate a connected approval and execution path, but they do not specify confidentiality, authorization, or logging controls, so the disclosure path is only weakly constrained by design.", "residual_risk": "Authorized users or compromised internal components may still view sensitive invoice and payment data, and operational logs may expose limited metadata if not tightly protected."}, {"threat_id": "T4", "enforcement_component_ids": ["C1", "C2", "C4", "C5"], "recommendation": "Apply size limits, content parsing limits, rate limiting, and quarantine rules at intake and pre-execution; reject malformed, oversized, or repetitive invoice submissions before they can consume extraction, approval, or execution resources.", "verification": "Send benign oversized attachments, malformed documents, and repeated low-rate test submissions, then confirm they are throttled, rejected, or isolated without delaying normal invoice processing.", "existing_control_effect": "CTL1 already includes validation/filtering and CTL2/CTL3 preserve a staged workflow, but their ability to stop resource exhaustion is unspecified.", "residual_risk": "A sustained flood of well-formed but high-volume submissions may still consume capacity if throttling and capacity planning are insufficient."}]}
{"id": "E011", "collection": "comparison", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "single_prompt", "trial_id": "slot-09-single_prompt", "topology_family_id": "topology-011", "annotation_status": "candidate", "image_path": "images/comparison/E011.png", "title": "finance_operations", "purpose": "An invoice agent reads externally supplied invoice documents, prepares payment drafts, and hands them to an approval and payment-execution path.", "assumptions": ["Synthetic design; no actual production data.", "The supplier portal is attacker-influenced but not authenticated as trusted by default.", "The invoice agent can invoke only the validation and ledger interfaces shown here; no direct bank or email tool access is assumed.", "Approval remains a separate human step before payment execution."], "controls": [{"id": "CTRL1", "component_ids": ["C3"], "flow_ids": ["F1", "F2", "F3"], "fact": "Document Validation Service rejects unsigned or malformed invoices and records document provenance before returning a normalized invoice record."}, {"id": "CTRL2", "component_ids": ["C2"], "flow_ids": ["F2", "F4"], "fact": "The Invoice Agent is limited to read-only document ingestion, field extraction, and draft creation; it cannot execute payments or edit approval records."}, {"id": "CTRL3", "component_ids": ["C5", "C6"], "flow_ids": ["F5", "F6"], "fact": "Approver Console requires named-user approval before the Payment Executor accepts an instruction."}, {"id": "CTRL4", "component_ids": ["C4"], "flow_ids": ["F4", "F5"], "fact": "Draft Ledger stores payment drafts with immutable timestamps and the originating document identifier."}], "architecture": {"components": [{"id": "C1", "label": "Supplier Portal", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Document Validation Service", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Draft Ledger", "zone": "Z2", "kind": "data_store"}, {"id": "C5", "label": "Approver Console", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "parsed invoice text, extracted fields, and provenance metadata", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C2", "data": "validation result and normalized invoice record", "boundary_ids": []}, {"id": "F4", "source": "C2", "target": "C4", "data": "payment draft and reconciliation notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "draft payment package for human approval", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C5", "target": "C6", "data": "approved payment instruction", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Invoice Processing"}, {"id": "Z3", "label": "Approval and Payment"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation via supplier document", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F4"], "boundary_ids": ["B1"], "asset": "invoice draft integrity and downstream approval quality", "threat_vector": "Attacker-controlled invoice text embeds instructions that influence the Invoice Agent during extraction or drafting.", "preconditions": ["Attacker can influence external documents supplied through the supplier portal.", "The agent processes document content that may contain adversarial instructions.", "No stronger content isolation is stated for the agent's prompt/context handling."], "impact": "The agent may alter the payment draft, misstate payee details, or omit reconciliation notes, causing an incorrect approval request.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "The Invoice Agent is limited to read-only document ingestion, field extraction, and draft creation; it cannot execute payments or edit approval records.", "justification": "The architecture explicitly allows the agent to read externally supplied documents and create drafts, while the stated control constrains execution authority but does not specify robust instruction isolation or anti-prompt-injection handling. This leaves a supported agent-specific manipulation path from external input to draft creation."}, {"id": "T2", "category": "S", "subtype": "spoofed invoice provenance", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "supplier identity and invoice authenticity", "threat_vector": "A forged or replayed document is presented as if it came from a legitimate supplier source.", "preconditions": ["Attacker can supply or modify invoice documents at the external boundary.", "Provenance checks depend on the validation service and are not otherwise independently verified."], "impact": "A fake invoice may be normalized and advanced toward approval, creating a payment fraud risk.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "Document Validation Service rejects unsigned or malformed invoices and records document provenance before returning a normalized invoice record.", "justification": "The validation service explicitly checks signing/malformed conditions and records provenance, which reduces but does not eliminate spoofing risk because the exact trust anchor and provenance strength are unspecified."}, {"id": "T3", "category": "I", "subtype": "draft exposure through approval handoff", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": ["B2"], "asset": "draft payment details and invoice metadata", "threat_vector": "Payment draft content crosses from the invoice-processing zone to the approval zone where it may be observed by an unauthorized or overbroad approver workflow participant.", "preconditions": ["An attacker can cause sensitive but valid invoice details to be placed into a draft.", "The approval workflow may expose draft contents to users beyond the minimum necessary approver set."], "impact": "Sensitive invoice amounts, supplier identifiers, or reconciliation notes may be disclosed to an unintended recipient.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "Approver Console requires named-user approval before the Payment Executor accepts an instruction.", "justification": "Named-user approval constrains execution, but the effect on who can view draft contents is not specified. The path exists, yet unauthorized disclosure is not directly demonstrated by the architecture facts."}], "scope_notes": ["Boundary IDs follow first-seen unordered zone pairs: F1/F2/F3/F4 create B1 between Z1 and Z2; F5 creates B2 between Z2 and Z3.", "Unsupported paths remain possible where agent prompt isolation, document authenticity proofs, or approver visibility rules are not fully specified.", "A data-store write to the Draft Ledger does not by itself prove disclosure to an unauthorized recipient.", "No direct bank connectivity or autonomous payout path is shown; payment execution requires the approval flow."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Harden the Invoice Agent with strict instruction parsing rules, strip or quarantine free-form directives from supplier documents, and require the validation service to pass only structured fields plus provenance metadata into draft creation.", "verification": "Submit benign invoices containing embedded instructions and confirm the agent ignores them while still extracting invoice fields correctly; verify the ledger stores only structured drafts with provenance identifiers.", "existing_control_effect": "The agent is already limited to read-only ingestion and draft creation, and the validation service records document provenance.", "residual_risk": "If document parsing or context separation is incomplete, adversarial text may still influence extracted values or notes."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C2"], "recommendation": "Enforce signed-source verification and supplier allowlisting in the validation service, and have the agent reject any normalized record lacking an accepted provenance marker before drafting.", "verification": "Use test invoices with missing, altered, and valid signatures to confirm only allowlisted signed documents reach draft creation.", "existing_control_effect": "Document Validation Service rejects unsigned or malformed invoices and records document provenance before returning a normalized invoice record.", "residual_risk": "If provenance anchors are weak or not uniquely bound to the supplier, replayed or impersonated documents may still pass some checks."}, {"threat_id": "T3", "enforcement_component_ids": ["C5", "C4"], "recommendation": "Limit approver visibility to minimum-necessary fields, redact nonessential invoice details in the handoff package, and retain the immutable draft record for audit review only by authorized roles.", "verification": "Review the approval screen with sample drafts to confirm only required fields are displayed and that unauthorized users cannot open the ledger record.", "existing_control_effect": "Approver Console requires named-user approval before the Payment Executor accepts an instruction.", "residual_risk": "Even with named-user approval, approved users may still see sensitive draft content unless field-level redaction is added."}]}
{"id": "E012", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-04-no_reviewer", "topology_family_id": "topology-012", "annotation_status": "candidate", "image_path": "images/comparison/E012.png", "title": "Guarded support-assist workflow with external evidence intake and account lookup", "purpose": "A support agent drafts customer replies by combining externally supplied help documents with customer-account data, while enforcing provenance checks, scoped account authorization, and bounded tool use.", "assumptions": ["External help documents may be attacker-influenced or malformed.", "The support agent can read help content and query a customer-account store through controlled tools.", "Account access is limited to the minimum records needed for the active ticket, but some residual risk remains where document provenance or response grounding is imperfect.", "No privileged internal access is available to the attacker; influence is limited to public-facing document or evidence inputs."], "controls": [{"id": "K1", "component_ids": ["C3"], "flow_ids": ["F1", "F2"], "fact": "Inbound help content is checked for source provenance and tagged before the agent can use it."}, {"id": "K2", "component_ids": ["C2", "C4"], "flow_ids": ["F3", "F4", "F6"], "fact": "The account lookup tool accepts only ticket-scoped read queries and returns only approved customer fields."}, {"id": "K3", "component_ids": ["C2", "C6"], "flow_ids": ["F7"], "fact": "The draft step uses the validated evidence summary and the authorized account result set as its only inputs."}, {"id": "K4", "component_ids": ["C4", "C5"], "flow_ids": ["F4", "F5"], "fact": "The record store is read-only to this workflow; the support agent has no direct write path to customer data."}], "architecture": {"components": [{"id": "C1", "label": "Help intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Provenance gate", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Account lookup", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Customer record store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Reply draft", "zone": "Z2", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C3", "data": "help documents and evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C3", "target": "C2", "data": "validated evidence summary and provenance flags", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C4", "data": "scoped account query for active ticket", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C5", "data": "read request for allowed customer fields", "boundary_ids": []}, {"id": "F5", "source": "C5", "target": "C4", "data": "ticket-scoped account data", "boundary_ids": []}, {"id": "F6", "source": "C4", "target": "C2", "data": "authorized account result set", "boundary_ids": ["B2"]}, {"id": "F7", "source": "C2", "target": "C6", "data": "grounded customer response draft", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "Public intake"}, {"id": "Z2", "label": "Support workspace"}, {"id": "Z3", "label": "Account services"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/evidence injection via public intake", "component_ids": ["C1", "C3", "C2", "C6"], "flow_ids": ["F1", "F2", "F7"], "boundary_ids": ["B1"], "asset": "support reply drafting process and customer-facing response quality", "threat_vector": "Attacker supplies malformed or misleading help documents in public intake to influence the provenance gate’s summary and the agent’s downstream drafting.", "preconditions": ["Attacker can submit or influence public-facing help documents or evidence.", "The agent relies on the validated evidence summary when drafting replies."], "impact": "Misleading or attacker-shaped evidence can steer the support agent to generate incorrect, unsafe, or policy-violating responses.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Inbound help content is checked for source provenance and tagged before the agent can use it.", "justification": "The architecture explicitly accepts external help content across the public-to-workspace boundary and then reuses the validated summary in drafting. Provenance checking reduces the path, but the stated residual risk and attacker-influenced input keep this as a supported agent-influence threat."}, {"id": "T2", "category": "I", "subtype": "ticket-scoped account data overexposure", "component_ids": ["C2", "C4", "C5"], "flow_ids": ["F3", "F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "customer account fields returned to the support agent", "threat_vector": "A support workflow query crosses into account services and retrieves customer data for the active ticket.", "preconditions": ["The support agent has access to the lookup tool.", "An active ticket exists that authorizes a scoped query."], "impact": "If the approved field set or ticket scoping is broader than intended, customer record details may be disclosed to the support agent beyond the minimum necessary for the case.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K4"], "evidence_quote": "The account lookup tool accepts only ticket-scoped read queries and returns only approved customer fields.", "justification": "The design clearly supports retrieval of account data across the Z2/Z3 boundary, but the control only states intended scoping and approved fields. Effectiveness against overbroad disclosure is not fully specified, so the exposure is possible but not demonstrated as weak enough to call exposed."}], "scope_notes": ["Only public-facing evidence input and support-tool/account-lookup paths are in scope.", "No direct write path to customer data is described; write or tamper threats against C5 are unsupported by the supplied architecture facts.", "The architecture does not specify authentication, logging, replay resistance, or human-review steps; those gaps prevent stronger conclusions for repudiation or privilege-escalation scenarios.", "Reverse-direction or same-zone flows do not create additional boundaries beyond B1 and B2 as listed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C3", "C2", "C6"], "recommendation": "Keep provenance gating mandatory at C3 and only allow C2 to draft from a fixed, structured evidence summary that excludes raw public text, instructions, and imperative language. Add a refusal rule in C2/C6: if provenance is weak, conflicting, or malformed, the draft must omit the evidence rather than reinterpret it.", "verification": "Submit benign test documents with conflicting claims, embedded instructions, and malformed markup; confirm C3 tags them as untrusted and C2/C6 only receive the bounded summary and produce a cautious draft that does not follow attacker text.", "existing_control_effect": "K1 already checks source provenance and tags inbound content, and K3 limits drafting inputs to the validated evidence summary plus authorized account results.", "residual_risk": "An attacker may still influence reply quality through misleading but provenance-valid public evidence, so some prompt/evidence shaping risk remains even with filtering and structured summarization."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C4", "C5"], "recommendation": "Enforce least-privilege field allowlists at C4 per ticket type and require C2 to present a ticket ID and purpose token for every lookup. Keep C5 read-only to this workflow and reject any request that expands beyond the approved field set or ticket scope.", "verification": "Use a harmless test ticket that requests only one permitted field, then attempt broader field requests and cross-ticket lookups; verify C4 returns only the approved field set and that C5 remains inaccessible except through read-only, ticket-scoped queries.", "existing_control_effect": "K2 already restricts lookup queries to ticket scope and approved fields, and K4 prevents direct writes from this workflow to the record store.", "residual_risk": "If the approved allowlist or ticket scoping is misconfigured, the agent can still receive more customer data than intended, so some over-disclosure risk remains within the supported lookup path."}]}
{"id": "E013", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-15-no_reviewer", "topology_family_id": "topology-013", "annotation_status": "candidate", "image_path": "images/comparison/E013.png", "title": "Guarded multi-agent incident triage with provenance-checked evidence flow", "purpose": "A guarded incident-response workflow where two analyst agents independently process externally supplied incident evidence, submit findings to a coordinator, and persist curated case notes into a case store under explicit authorization and provenance controls.", "assumptions": ["External incident documents, logs, or attachments may be attacker-influenced but arrive without privileged system access.", "Analyst agents can use bounded tools only through a coordinator-enforced policy layer; direct case-store writes are not allowed.", "The coordinator verifies identity/authorization for internal service calls and applies provenance metadata checks before accepting findings.", "The case store is append/update protected by authenticated application writes, but content quality and semantic correctness are not fully guaranteed by the controls stated.", "Attacker capability is limited to supplying crafted external evidence and attempting instruction or content manipulation through that evidence; no internal credentialed access is assumed."], "controls": [{"id": "CTL1", "component_ids": ["C4", "C5"], "flow_ids": ["F5", "F6"], "fact": "Coordinator-only write path to the case store; direct analyst writes are disallowed by design and storage writes require authenticated application calls."}, {"id": "CTL2", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Analyst tools are bounded and policy-enforced through the coordinator; analysts may parse/retrieve evidence but do not have unrestricted shell or store access."}, {"id": "CTL3", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "Internal service calls are checked for identity and authorization before findings are accepted for triage."}, {"id": "CTL4", "component_ids": ["C5", "C6"], "flow_ids": ["F6"], "fact": "Provenance metadata is checked on accepted case-note writes, but the check does not fully guarantee semantic correctness of the content."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Triage Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Provenance Gate", "zone": "Z2", "kind": "process"}, {"id": "C6", "label": "Case Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "findings, extracted indicators, provenance metadata", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "findings, extracted indicators, provenance metadata", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "authorized case note draft and write request", "boundary_ids": []}, {"id": "F6", "source": "C5", "target": "C6", "data": "approved case notes, provenance tags, update record", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Evidence"}, {"id": "Z2", "label": "Internal Triage"}, {"id": "Z3", "label": "Case Storage"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "prompt/content injection via external evidence", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1"], "asset": "triage findings and downstream case notes", "threat_vector": "Crafted incident documents, logs, or attachments inject misleading instructions or content into analyst processing.", "preconditions": ["Attacker can supply or influence external evidence content.", "Analysts process evidence received from the external evidence inbox."], "impact": "Could tamper with analyst findings and extracted indicators, causing incorrect triage outputs to propagate to the coordinator.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTL2"], "evidence_quote": "Analyst tools are bounded and policy-enforced through the coordinator; analysts may parse/retrieve evidence but do not have unrestricted shell or store access.", "justification": "The architecture explicitly allows external evidence to flow into analyst agents across the external/internal boundary, and the stated control only bounds tools; it does not eliminate content-manipulation risk in the evidence itself."}, {"id": "T2", "category": "I", "subtype": "sensitive case-data disclosure through shared triage outputs", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4"], "boundary_ids": [], "asset": "findings, extracted indicators, provenance metadata", "threat_vector": "An analyst discloses sensitive incident details in findings sent to the coordinator.", "preconditions": ["An analyst processes sensitive evidence and includes it in triage outputs.", "The coordinator receives and handles those outputs."], "impact": "Could expose sensitive incident information within the internal triage zone or into later stored notes.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL3"], "evidence_quote": "Internal service calls are checked for identity and authorization before findings are accepted for triage.", "justification": "The internal flow exists, but the only stated control is identity/authorization checking on service calls. That limits who can submit findings, yet the architecture does not specify confidentiality filtering, redaction, or recipient restrictions for the content of those findings."}, {"id": "T3", "category": "T", "subtype": "unauthorized note content manipulation before storage", "component_ids": ["C4", "C5", "C6"], "flow_ids": ["F5", "F6"], "boundary_ids": ["B2"], "asset": "case notes and provenance tags in the case store", "threat_vector": "A forged or manipulated write request alters approved case-note content or provenance tags before persistence.", "preconditions": ["An attacker can influence the coordinator's note draft or write request content.", "The write path reaches the provenance gate and case store."], "impact": "Could tamper with persisted case records, undermining case integrity and provenance fidelity.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTL1", "CTL4"], "evidence_quote": "Coordinator-only write path to the case store; direct analyst writes are disallowed by design and storage writes require authenticated application calls.", "justification": "The write path is constrained to coordinator-mediated authenticated application calls and provenance metadata is checked, which reduces direct tampering opportunities. However, the control explicitly does not fully guarantee semantic correctness, so content manipulation is not eliminated."}, {"id": "T4", "category": "D", "subtype": "triage pipeline disruption via malformed evidence", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1"], "asset": "availability of analyst triage and coordinator processing", "threat_vector": "Large, malformed, or adversarial evidence causes excessive parsing or repeated rework in the analyst/coordinator loop.", "preconditions": ["Attacker can submit crafted external documents or attachments.", "Analyst parsing and triage consume bounded internal resources."], "impact": "Could delay or stall triage operations and reduce availability of the incident-response workflow.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL2"], "evidence_quote": "Analyst tools are bounded and policy-enforced through the coordinator; analysts may parse/retrieve evidence but do not have unrestricted shell or store access.", "justification": "Bounding and policy enforcement suggest mitigation, but the effectiveness against resource-exhaustion or malformed-input attacks is not specified, so availability exposure cannot be confirmed as fully exposed."}], "scope_notes": ["Boundary IDs inferred from the listed flow order: B1 covers Z1<->Z2 and B2 covers Z2<->Z3.", "No direct analyst-to-case-store write flow is present; unsupported paths for that route are out of scope.", "The architecture states provenance checks, but semantic correctness is explicitly not guaranteed, leaving residual integrity risk.", "No evidence of cross-agent spoofing or repudiation handling is stated beyond identity/authorization checks, so those categories are not asserted here."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Add content-safety and provenance validation at the coordinator/provenance-gate boundary: strip or quarantine instruction-like text from external evidence summaries, require evidence-to-finding citations, and reject findings that introduce new unverified directives from C1 inputs.", "verification": "Use a benign test corpus with harmless embedded instructions in attachments and confirm the coordinator flags, quarantines, or strips them before the analyst findings are accepted; verify the original evidence remains readable while injected directives are not propagated into triage notes.", "existing_control_effect": "CTL2 already bounds analyst tools through the coordinator, reducing direct abuse of agent capabilities and store access.", "residual_risk": "Attackers can still influence analyst reasoning through persuasive or ambiguous content; content filtering reduces but does not eliminate manipulation risk."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Enforce output redaction and minimization on analyst-to-coordinator findings: require a coordinator-side schema that separates indicators from narrative text, automatically redact secrets or personal data before acceptance, and block unsafely detailed free-text fields from reaching downstream notes.", "verification": "Submit a benign sample incident containing marked test secrets and personal-data placeholders, then confirm the coordinator accepts the finding only after redaction/minimization and that prohibited fields do not appear in stored case notes.", "existing_control_effect": "CTL3 restricts who can submit findings, which limits unauthorized submitters but not the sensitivity of approved content.", "residual_risk": "Authorized analysts may still include sensitive details that are hard to classify automatically, so some disclosure risk remains within the internal workflow."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5", "C6"], "recommendation": "Require structured, allowlisted write requests at the provenance gate: bind each case-note update to a coordinator-generated signed record, validate field-level schemas for notes and provenance tags, and reject any mutation that changes approved content outside the expected update template.", "verification": "Use a benign update containing only allowed fields and confirm it is stored unchanged; then replay the same request with harmless extra fields or altered provenance tags and confirm the gate rejects it without writing to the case store.", "existing_control_effect": "CTL1 prevents direct analyst writes and forces authenticated coordinator-mediated storage writes, while CTL4 checks provenance metadata on accepted writes.", "residual_risk": "A malicious or compromised coordinator path could still submit semantically misleading but schema-valid content; the stated controls reduce direct tampering but do not guarantee semantic correctness."}, {"threat_id": "T4", "enforcement_component_ids": ["C1", "C2", "C3", "C4"], "recommendation": "Add input size, parsing-time, and rework limits on external evidence processing: cap attachment size and nesting depth, isolate expensive parsing in a bounded worker with timeout/circuit-breaker behavior, and instruct the coordinator to drop or defer evidence that repeatedly triggers excessive processing.", "verification": "Send benign oversized but nonmalicious sample files and confirm the workflow truncates, times out safely, or quarantines them without blocking unrelated cases; confirm normal-sized evidence still completes end-to-end.", "existing_control_effect": "CTL2 already bounds analyst tools and prevents unrestricted shell or store access, which limits some abuse amplification.", "residual_risk": "An attacker may still cause partial slowdown through many moderate-cost submissions or edge-case parsers, so availability impact is reduced but not eliminated."}]}
{"id": "E014", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "single_prompt", "trial_id": "slot-02-single_prompt", "topology_family_id": "topology-014", "annotation_status": "candidate", "image_path": "images/comparison/E014.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "The attacker can influence external help documents but has no privileged access to the support app or customer store.", "The support agent uses the workspace output to compose customer replies."], "controls": [{"id": "K1", "component_ids": ["C3"], "flow_ids": ["F1", "F2"], "fact": "The help document fetcher only accepts HTTPS URLs from an allowlist."}, {"id": "K2", "component_ids": ["C2"], "flow_ids": ["F2", "F3", "F4"], "fact": "The response workspace shows source snippets and account fields in the same pane."}, {"id": "K3", "component_ids": ["C1", "C2"], "flow_ids": ["F4", "F5"], "fact": "The agent can edit the draft before sending it to the customer."}], "architecture": {"components": [{"id": "C1", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C2", "label": "Response Workspace", "zone": "Z2", "kind": "process"}, {"id": "C3", "label": "Help Document Fetcher", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Help Docs Repository", "zone": "Z1", "kind": "external_entity"}, {"id": "C5", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C4", "target": "C3", "data": "help articles and support snippets", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C3", "target": "C2", "data": "retrieved help content", "boundary_ids": []}, {"id": "F3", "source": "C5", "target": "C2", "data": "customer-account details", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C1", "target": "C2", "data": "draft response text and edits", "boundary_ids": []}, {"id": "F5", "source": "C2", "target": "C1", "data": "prepared response and cited facts", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support App"}, {"id": "Z3", "label": "Customer Data"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection via external help docs", "component_ids": ["C4", "C3", "C2", "C1"], "flow_ids": ["F1", "F2", "F5"], "boundary_ids": ["B1"], "asset": "response preparation integrity", "threat_vector": "Attacker supplies help content that embeds instructions intended to steer the support agent or workspace output.", "preconditions": ["Attacker can influence external help documents.", "Agent relies on fetched snippets while drafting responses."], "impact": "Incorrect or attacker-steered replies, including unsafe tool or response content.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "The help document fetcher only accepts HTTPS URLs from an allowlist.", "justification": "The path exists because external help content reaches the workspace used by the agent, but the allowlist limits which sources can be fetched, reducing exposure rather than leaving an unrestricted path."}, {"id": "T2", "category": "I", "subtype": "cross-zone account-data exposure in workspace", "component_ids": ["C5", "C2", "C1"], "flow_ids": ["F3", "F5"], "boundary_ids": ["B2"], "asset": "customer-account details", "threat_vector": "Sensitive account fields are displayed together with response drafting content, allowing the agent to view more data than needed for the reply.", "preconditions": ["The workspace presents account fields alongside reply content.", "The agent can read the workspace output."], "impact": "Unnecessary disclosure of customer-account details to the support agent.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "The response workspace shows source snippets and account fields in the same pane.", "justification": "This design fact directly supports visibility of account data in the agent workspace across the customer-data boundary, with no compensating access separation stated."}, {"id": "T3", "category": "R", "subtype": "draft attribution ambiguity", "component_ids": ["C1", "C2"], "flow_ids": ["F4", "F5"], "boundary_ids": [], "asset": "response accountability", "threat_vector": "Edits made by the support agent are merged into the draft without a distinct audit marker for who changed what.", "preconditions": ["The workspace permits agent edits before send.", "Reviewers later need to attribute response content changes."], "impact": "Disputes over whether a response came from the agent, fetched content, or a later edit.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "The agent can edit the draft before sending it to the customer.", "justification": "The edit capability is explicit, but the available controls do not state whether edits are logged or attributed; effectiveness for accountability is therefore unspecified."}], "scope_notes": ["Only one connected customer-support architecture is analyzed.", "Boundary IDs follow first-seen zone pairs: B1 between External Sources and Support App; B2 between Customer Data and Support App.", "Unsupported paths: the customer store is not shown as directly exposed outside the workspace, so disclosure beyond the agent is not asserted.", "Incomplete information: no logging, redaction, or approval control is specified for draft edits or account-field display."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C3", "C2"], "recommendation": "Add content sanitization and instruction stripping in the help document fetcher, and render fetched snippets as non-executable quoted text in the workspace.", "verification": "Use a benign test document containing obvious directive text and confirm the workspace only displays inert quoted content and no instruction tokens are preserved as actionable UI text.", "existing_control_effect": "The HTTPS allowlist already reduces source risk by limiting where documents can be fetched from.", "residual_risk": "A trusted-but-malicious allowed source can still supply harmful instructions unless content is sanitized and separated from agent guidance."}, {"threat_id": "T2", "enforcement_component_ids": ["C2"], "recommendation": "Split the workspace view so customer-account fields are masked by default and only revealed on explicit need-to-know selection within the response workspace.", "verification": "Open a test account and confirm the default pane hides sensitive fields unless the operator selects a permitted detail view.", "existing_control_effect": "The current workspace already co-locates source snippets and account fields, so no reducing control is evident for this path.", "residual_risk": "The agent may still access some account data when needed for support, but unnecessary ambient exposure is lowered."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C1"], "recommendation": "Add immutable edit logging in the response workspace that records agent identity, timestamp, and before/after draft diffs for each change.", "verification": "Create a sample draft, make two harmless edits, and confirm the audit log shows both edits with unchanged original text preserved.", "existing_control_effect": "The current edit permission enables the workflow, but it does not state any attribution or audit mechanism.", "residual_risk": "Logging improves accountability, but a malicious operator can still make harmful edits if approval review is absent."}]}
{"id": "E015", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "single_prompt", "trial_id": "slot-03-single_prompt", "topology_family_id": "topology-015", "annotation_status": "candidate", "image_path": "images/comparison/E015.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "The attacker can influence external help documents but has no privileged access to the support app or customer-account store.", "Tool calls from the drafting service are bounded to the listed retrieval and lookup flows."], "controls": [{"id": "CTRL1", "component_ids": ["C4"], "flow_ids": ["F1", "F2"], "fact": "Provenance checks reject documents without an approved source label before excerpting."}, {"id": "CTRL2", "component_ids": ["C3"], "flow_ids": ["F3", "F4"], "fact": "The drafting service can call only the documented retrieval and customer-lookup interfaces, and it cannot execute arbitrary tools."}, {"id": "CTRL3", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "Customer-account access requires service authentication and record-level authorization tied to the case owner."}, {"id": "CTRL4", "component_ids": ["C2", "C3"], "flow_ids": ["F3", "F6"], "fact": "Agent-visible drafts are logged with the originating case ID and the retrieved source snippets."}], "architecture": {"components": [{"id": "C1", "label": "Help Document Repository", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Response Drafting Service", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Policy and Retrieval Filter", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C4", "data": "externally supplied help document content and metadata", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C4", "target": "C3", "data": "approved excerpts and retrieval results", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C3", "data": "support case details and response instructions", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C5", "data": "customer-account lookup request", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C3", "data": "customer account attributes and case context", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C3", "target": "C2", "data": "draft customer response", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support App Zone"}, {"id": "Z3", "label": "Customer Data Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction or context manipulation via external help documents", "component_ids": ["C1", "C4", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "response drafting behavior and response quality", "threat_vector": "Attacker-influenced help documents embed misleading instructions that survive ingestion and steer the drafting service.", "preconditions": ["Attacker can influence external document content.", "The filter accepts some externally supplied documents for retrieval review."], "impact": "Unsafe or incorrect support responses may be drafted from manipulated guidance, reducing trust in agent output.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "Provenance checks reject documents without an approved source label before excerpting.", "justification": "The path crosses from External Sources to the Support App Zone, but provenance checks explicitly limit which documents are excerpted, reducing the attack path rather than leaving it fully exposed."}, {"id": "T2", "category": "I", "subtype": "customer data exposure through draft generation", "component_ids": ["C3", "C5", "C2"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "customer account attributes", "threat_vector": "A support agent or manipulated drafting context requests more customer data than needed, causing sensitive account attributes to appear in the draft.", "preconditions": ["The drafting service can query the account store for case preparation.", "The draft may include returned account attributes in agent-visible output."], "impact": "Sensitive customer attributes may be disclosed to the support agent beyond what is necessary for the case.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL3", "CTRL4"], "evidence_quote": "Customer-account access requires service authentication and record-level authorization tied to the case owner.", "justification": "The data path crosses from Support App Zone to Customer Data Zone and back; record-level authorization limits retrieval, and draft logging reduces silent misuse, so the path is constrained rather than fully open."}, {"id": "T3", "category": "R", "subtype": "draft provenance ambiguity", "component_ids": ["C2", "C3"], "flow_ids": ["F3", "F6"], "boundary_ids": [], "asset": "traceability of support response preparation", "threat_vector": "A user or operator disputes which inputs were used to generate a customer response.", "preconditions": ["Drafts may be edited or forwarded after generation.", "Audit review depends on recorded case identifiers and snippets."], "impact": "It may be difficult to prove whether a draft followed approved sources or case input.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL4"], "evidence_quote": "Agent-visible drafts are logged with the originating case ID and the retrieved source snippets.", "justification": "The relevant flows remain inside the Support App Zone, so there is no boundary-crossing exposure; logging improves traceability but does not eliminate all dispute scenarios."}], "scope_notes": ["Flow boundary IDs were assigned from the listed flow order: F1/F2 cross B1 between Z1 and Z2; F4/F5 cross B2 between Z2 and Z3.", "The architecture does not specify whether user-entered case details are separately validated before reaching the drafting service.", "No unsupported tool paths are modeled beyond the documented retrieval and customer-lookup interfaces.", "Threats are limited to the supplied functional path; no claim is made about other support channels or production controls."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C4", "C3"], "recommendation": "Keep provenance enforcement mandatory at C4 and require C3 to reject any excerpt set that lacks an approved source label or contains conflicting instruction-like text.", "verification": "Submit a benign test document with no approved source label and confirm C4 blocks excerpting while C3 receives no content from F2.", "existing_control_effect": "CTRL1 already rejects documents without an approved source label before excerpting.", "residual_risk": "A labeled but misleading external document can still pass provenance checks and influence response quality if human review is weak."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C5"], "recommendation": "Enforce least-privilege lookup scopes in C3 so only case-owned records and minimum necessary fields can be requested from C5, and keep draft logging enabled for every lookup-driven response.", "verification": "Run a case-owned lookup and a non-owned lookup attempt, then confirm the former succeeds only for allowed fields and the latter is denied without draft output.", "existing_control_effect": "CTRL3 requires service authentication and record-level authorization tied to the case owner.", "residual_risk": "Authorized lookups can still return sensitive but necessary account attributes that appear in the draft."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Require every draft exported from C3 to include the case ID and source-snippet metadata, and keep review procedures that check the log against the final customer response.", "verification": "Generate a benign draft, then confirm the case ID and snippet references appear in the log entry and can be matched to the response text.", "existing_control_effect": "CTRL4 logs drafts with the originating case ID and retrieved source snippets.", "residual_risk": "Logs support later review but do not prevent an operator from misrepresenting the draft after it leaves the system."}]}
{"id": "E016", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-05-no_reviewer", "topology_family_id": "topology-016", "annotation_status": "candidate", "image_path": "images/comparison/E016.png", "title": "Support agent with external docs and account lookup", "purpose": "A customer-support AI agent drafts replies by reading externally supplied help documents and querying a customer-account store, then composing a response for the human or customer-facing workflow.", "assumptions": ["The support agent can ingest externally supplied help documents or evidence artifacts during case handling.", "The agent has read access to a customer-account store for account context and history.", "Relevant control implementation and effectiveness are unspecified; no assumption is made that validation, access checks, or prompt hardening are absent or effective.", "The attacker can influence external documents or evidence but has no privileged access to internal systems.", "The architecture is one connected functional flow with no extra padding components."], "controls": [{"id": "K1", "component_ids": ["C2"], "flow_ids": ["F1", "F2", "F3"], "fact": "The support agent is configured to combine retrieved document snippets with account context when drafting case responses; how strongly it resists instruction injection or misleading evidence is unspecified."}, {"id": "K2", "component_ids": ["C4"], "flow_ids": ["F2", "F3"], "fact": "The account store is a separate internal data source that is queried read-only by the support agent for case context; authorization and filtering effectiveness are unspecified."}, {"id": "K3", "component_ids": ["C3", "C5"], "flow_ids": ["F4", "F5"], "fact": "Drafts are passed to the reply queue and then to the customer-facing workflow for review or sending; any human approval gate is not specified."}], "architecture": {"components": [{"id": "C1", "label": "Help document inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Reply draft queue", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Customer account store", "zone": "Z3", "kind": "data_store"}, {"id": "C5", "label": "Human or customer-facing channel", "zone": "Z1", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "externally supplied help documents, evidence artifacts", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C4", "data": "account lookup query", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C4", "target": "C2", "data": "customer account context, history, status", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C2", "target": "C3", "data": "draft reply content", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C5", "data": "prepared support response", "boundary_ids": ["B1"]}], "trust_zones": [{"id": "Z1", "label": "External sources"}, {"id": "Z2", "label": "Support workflow"}, {"id": "Z3", "label": "Customer data store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction injection via external help documents", "component_ids": ["C1", "C2", "C3", "C5"], "flow_ids": ["F1", "F4", "F5"], "boundary_ids": ["B1"], "asset": "support reply drafting and customer communication", "threat_vector": "Attacker-controlled help documents or evidence artifacts are ingested by the support agent and can influence draft generation.", "preconditions": ["Attacker can supply or modify external documents/evidence used in case handling.", "The agent combines retrieved snippets with account context when drafting replies."], "impact": "Injected instructions or misleading evidence may steer the agent to produce unsafe or incorrect replies, causing integrity loss in the support workflow and downstream customer harm.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "The support agent is configured to combine retrieved document snippets with account context when drafting case responses; how strongly it resists instruction injection or misleading evidence is unspecified.", "justification": "This is a supported agent-specific manipulation path from external input in Z1 into the drafting process in Z2, with the drafted output later reaching the customer-facing workflow. The control fact explicitly says resistance to instruction injection is unspecified, so the path is exposed."}, {"id": "T2", "category": "I", "subtype": "inappropriate disclosure of customer account context through drafted reply", "component_ids": ["C2", "C4", "C3", "C5"], "flow_ids": ["F2", "F3", "F4", "F5"], "boundary_ids": ["B2", "B1"], "asset": "customer account context, history, and status", "threat_vector": "The support agent retrieves internal account data and may incorporate sensitive context into a prepared response that is sent outward.", "preconditions": ["The agent can query the customer-account store for case context.", "Draft content may include retrieved account details.", "The prepared response can reach the customer-facing workflow."], "impact": "Sensitive customer account information could be disclosed to an unintended recipient or broader audience than intended through the drafted reply.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The account store is a separate internal data source that is queried read-only by the support agent for case context; authorization and filtering effectiveness are unspecified.", "justification": "The architecture shows internal account data flowing to the agent and then into draft replies that are passed toward the customer-facing channel. However, the documentation does not state that sensitive fields are actually included or that disclosure reaches an unauthorized recipient, so this remains unknown rather than exposed."}, {"id": "T3", "category": "T", "subtype": "misleading evidence alters reply content", "component_ids": ["C1", "C2", "C3", "C5"], "flow_ids": ["F1", "F4", "F5"], "boundary_ids": ["B1"], "asset": "support reply accuracy and case record integrity", "threat_vector": "Externally supplied evidence artifacts are accepted into the drafting workflow and can tamper with the content and conclusions of the reply.", "preconditions": ["Attacker can influence external documents or evidence.", "The agent uses those artifacts when composing the response."], "impact": "The drafted reply may reflect altered, false, or incomplete evidence, leading to incorrect support actions or customer decisions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "The support agent is configured to combine retrieved document snippets with account context when drafting case responses; how strongly it resists instruction injection or misleading evidence is unspecified.", "justification": "This is a concrete tampering path where attacker-controlled evidence enters the support workflow and can change the resulting reply content. The architecture explicitly allows ingestion of external evidence, and the control note leaves resistance to misleading evidence unspecified."}], "scope_notes": ["Boundary IDs were assigned from the listed flows: B1 for Z1-Z2 and B2 for Z2-Z3.", "No standalone repudiation, denial-of-service, or elevation-of-privilege path is supported by the supplied facts.", "The customer-facing channel is described as a workflow destination; whether it is human-reviewed before sending is unspecified, which limits disclosure and integrity conclusions.", "Read-only access to the account store reduces write-impact paths, but authorization and filtering effectiveness are unspecified, so unauthorized read filtering cannot be assumed.", "No extra components or hidden approval gates should be inferred beyond the stated architecture."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Strip or quarantine untrusted instructions from inbound help documents before use, and require the support agent to treat external text as evidence only, not as operating instructions. Add a drafting rule in C2 to ignore any document content that attempts to direct tool use, policy overrides, or reply changes, and keep C3 as the only path for outbound drafts after this filtering.", "verification": "Submit benign test documents that contain obvious instruction-like text and confirm the agent flags them as untrusted content and drafts responses without following those instructions. Confirm the reply queue only receives drafts derived from approved support context, not copied prompt text.", "existing_control_effect": "K1 already combines retrieved snippets with account context, so the attack path exists but is partially constrained by the fact that the output still passes through the support workflow and reply queue. However, resistance to instruction injection is unspecified, so the current control does not reliably prevent manipulation.", "residual_risk": "If filtering or agent behavior is bypassed, a crafted document can still influence draft generation and downstream customer communication."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C5"], "recommendation": "Minimize account fields returned from C4 to only the specific data needed for the case, suppress sensitive history/status values by default, and require a human review step before any draft containing account-derived details is sent from C3 to C5. Enforce a redaction rule in C2 so draft text cannot include full account identifiers, secrets, or unrelated history.", "verification": "Use a harmless test account with known non-sensitive placeholders and confirm the agent only receives the limited approved fields. Then review sample drafts to ensure sensitive values are redacted or omitted before the response leaves the queue.", "existing_control_effect": "K2 keeps the store read-only, which limits write impact, and K3 shows drafts pass through a reply queue before reaching the customer-facing workflow. But authorization and filtering are unspecified, so those controls do not by themselves ensure that only appropriate data is disclosed.", "residual_risk": "If the agent is allowed to pull too much context or redaction fails, sensitive account details may still appear in customer-facing replies."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2"], "recommendation": "Validate externally supplied evidence against trusted case records before it can affect the draft, and require the agent to separate cited evidence from conclusions. Add a rule in C2 to reject or downgrade evidence that conflicts with internal case data unless explicitly confirmed by a trusted reviewer.", "verification": "Provide benign contradictory sample evidence and confirm the agent labels it as unverified, does not adopt its claims into the reply, and instead requests confirmation or uses the trusted record only.", "existing_control_effect": "K1 indicates the agent already mixes external snippets with account context, which means the workflow can incorporate evidence and also has internal context available for comparison. But the current control leaves resistance to misleading evidence unspecified, so tampering remains possible.", "residual_risk": "A crafted artifact can still distort the reply if validation is incomplete or if the agent incorrectly trusts the supplied evidence."}]}
{"id": "E017", "collection": "comparison", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-17-full_workflow", "topology_family_id": "topology-003", "annotation_status": "candidate", "image_path": "images/comparison/E017.png", "title": "Incident Evidence Triage with Dual Analyst Agents and Case Coordinator", "purpose": "Model an incident-response workflow where two analyst agents independently review externally supplied evidence, forward findings to a coordinator, and persist case material in a shared case store, with control effectiveness unspecified.", "assumptions": ["Externally supplied documents, logs, or artifacts may be attacker-influenced before ingestion.", "Two analyst agents are distinct processing components and both can receive the same external evidence stream.", "A coordinator aggregates or reconciles analyst findings before writing to the case store.", "Relevant validation, sandboxing, approval, and access-control implementation details are unspecified.", "No privileged system access is assumed for the attacker beyond influence over external evidence content."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1"], "fact": "Analyst agents ingest the same external evidence stream; validation and sandboxing behavior is unspecified."}, {"id": "K2", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "The coordinator compares and reconciles findings from both analysts before persisting them."}, {"id": "K3", "component_ids": ["C5"], "flow_ids": ["F5"], "fact": "The case store persists coordinator-written case material and references to supporting evidence."}, {"id": "K4", "component_ids": ["C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "fact": "No explicit access-control or approval detail is provided for this workflow."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Case Coordinator", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, artifacts", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, artifacts", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, extracted indicators, notes", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, extracted indicators, notes", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C4", "target": "C5", "data": "case summary, reconciled findings, evidence references", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Evidence Source"}, {"id": "Z2", "label": "Analyst Zone"}, {"id": "Z3", "label": "Coordination and Case Store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction/context manipulation via evidence content", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1", "B2"], "asset": "Analyst reasoning and coordinator reconciliation outputs", "threat_vector": "Attacker-influenced evidence content can inject misleading instructions or context into the analyst agents, then propagate to the coordinator through analyst findings.", "preconditions": ["Attacker can influence documents, logs, or artifacts before ingestion.", "The evidence stream is processed by both analyst agents.", "Validation and sandboxing behavior is unspecified."], "impact": "Misleading analyst conclusions or coordinator reconciliation, leading to incorrect case material and downstream decision errors.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K4"], "evidence_quote": "Analyst agents ingest the same external evidence stream; validation and sandboxing behavior is unspecified.", "justification": "The architecture explicitly allows attacker-influenced external evidence to be ingested by both agents, with unspecified validation/sandboxing and no explicit approval/access-control detail, creating a supported exposure for instruction or context manipulation."}, {"id": "T2", "category": "I", "subtype": "sensitive evidence disclosure across analyst and case-store path", "component_ids": ["C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1", "B2"], "asset": "Incident evidence and case material", "threat_vector": "Sensitive incident documents, logs, artifacts, and derived findings are handled by multiple components and stored in the case store, which may expose them to unintended recipients if access control is weak or absent.", "preconditions": ["The evidence contains sensitive incident details.", "Access control or approval details are unspecified."], "impact": "Unauthorized parties could learn incident contents or derived investigative details through the shared workflow or stored case material.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "The case store persists coordinator-written case material and references to supporting evidence.", "justification": "The workflow clearly persists evidence references and case material, but the architecture does not specify who may read the case store or whether access is restricted. That makes disclosure possible but not confirmed, so the path remains unknown rather than exposed."}, {"id": "T3", "category": "T", "subtype": "tampering with reconciled findings before persistence", "component_ids": ["C2", "C3", "C4", "C5"], "flow_ids": ["F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "Case summary and reconciled findings", "threat_vector": "A compromised or manipulated analyst output can alter the coordinator's reconciled findings that are then written into the case store.", "preconditions": ["An attacker can influence analyst output through malicious evidence content or other unspecified means.", "Coordinator reconciliation does not fully detect the manipulation."], "impact": "Corrupted case records, inaccurate evidence references, or altered incident conclusions in persistent storage.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The coordinator compares and reconciles findings from both analysts before persisting them.", "justification": "There is an explicit reconciliation step, which may reduce but does not eliminate tampering risk; effectiveness is not specified. Because the control exists but its strength is unclear, the assessment is unknown."}], "scope_notes": ["Boundary B1 covers flows between Z1 and Z2; boundary B2 covers flows between Z2 and Z3.", "No supported repudiation, denial-of-service, or elevation-of-privilege path is explicit in the supplied facts.", "The architecture does not specify read permissions for the case store, so disclosure from storage is incomplete information rather than confirmed unauthorized access.", "Validation/sandboxing effectiveness and access-control enforcement are explicitly unspecified, limiting confidence in any stronger or weaker assessment."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Add strict evidence sanitization and tool-privilege isolation at the analyst boundary, and require the coordinator to ignore any instructions embedded in source artifacts unless they come from an authenticated workflow field.", "verification": "Submit benign test artifacts containing instruction-like text, hidden prompts, and malformed metadata; confirm analysts treat them as inert evidence and the coordinator only consumes structured findings from allowed fields.", "existing_control_effect": "K1 already states the analyst agents ingest the same external evidence stream, but validation and sandboxing are unspecified; this means the path is recognized but not yet bounded by a stated control.", "residual_risk": "If malicious content still influences analyst interpretation or structured extraction, the coordinator may receive skewed findings despite sanitization."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Restrict case-store read access to authorized case roles only, and have the coordinator write only the minimum necessary case summary and evidence references, with access checks enforced before any retrieval.", "verification": "Use a benign authorized and unauthorized test account to attempt case-store reads; verify the unauthorized account is denied and the coordinator writes only the approved minimal record set.", "existing_control_effect": "K3 confirms the case store persists coordinator-written material and evidence references, but K4 says no explicit access-control or approval detail is provided, so disclosure is not prevented by any stated control.", "residual_risk": "Sensitive details may still be exposed if authorized readers over-broadly access stored case material or if referenced evidence remains reachable through other channels."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Make coordinator reconciliation write guarded by provenance checks: require both analyst outputs, record their hashes or identifiers, and reject persistence if the inputs are inconsistent or incomplete.", "verification": "Run a benign reconciliation test where one analyst output is modified in a controlled way; confirm the coordinator detects the mismatch and blocks persistence or flags the record for review.", "existing_control_effect": "K2 provides a reconciliation step before persistence, which can reduce casual tampering but does not guarantee integrity because its detection strength is unspecified.", "residual_risk": "A coordinated manipulation that survives reconciliation, or a failure in provenance tracking, could still corrupt the stored case summary and references."}]}
{"id": "E019", "collection": "comparison", "domain": "incident_response", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-14-full_workflow", "topology_family_id": "topology-017", "annotation_status": "candidate", "image_path": "images/comparison/E019.png", "title": "Weakly controlled incident-evidence triage with two analyst agents and coordinator", "purpose": "Incident-response workflow where two analyst agents ingest attacker-influenced external evidence, produce findings, and relay them through a coordinator into a case store under weak control assumptions.", "assumptions": ["External documents/evidence are attacker-influenced and arrive outside the trusted application zone.", "Two analyst agents may share prompt context or retrieval/tooling with weak validation.", "The coordinator accepts analyst findings and writes case records to a case store with permissive or weakly enforced checks.", "No privileged internal access is assumed for the attacker; influence is limited to external evidence content and any downstream agent behavior it induces.", "The case store is a trusted internal asset, while ingestion sources remain untrusted."], "controls": [{"id": "CTRL1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External evidence is accepted into the triage workspace with weak content validation and no strong provenance enforcement."}, {"id": "CTRL2", "component_ids": ["C2", "C3"], "flow_ids": ["F3", "F4"], "fact": "Both analyst agents can consume the same evidence and shared context, with limited isolation between their working state."}, {"id": "CTRL3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "The coordinator can persist findings to the case store using permissive write checks and minimal review gating."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident evidence, attachments, embedded content", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident evidence, attachments, embedded content", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, alerts, extracted indicators", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, alerts, extracted indicators", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case notes, triage decisions, persisted findings", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "Untrusted Evidence Sources"}, {"id": "Z2", "label": "Triage Workspace"}, {"id": "Z3", "label": "Case Records"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction injection via untrusted evidence", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3", "F5"], "boundary_ids": ["B1", "B2"], "asset": "triage findings and case records", "threat_vector": "Attacker-controlled embedded content in incoming evidence steers Analyst A’s interpretation or actions, which then propagates to the coordinator and into persisted case notes.", "preconditions": ["Attacker can influence external evidence content on F1.", "Analyst A consumes the content in the triage workspace and may share context/tooling with other analysts."], "impact": "Unsafe agent behavior can corrupt triage decisions, produce misleading alerts, or persist attacker-shaped conclusions in the case store.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2", "CTRL3"], "evidence_quote": "External evidence is accepted into the triage workspace with weak content validation and no strong provenance enforcement.", "justification": "The architecture explicitly places attacker-influenced evidence across B1 into analyst agents, with weak validation and limited isolation, and then allows coordinator persistence across B2 with minimal review gating. This supports an agent-manipulation path that can alter downstream findings and records."}, {"id": "T2", "category": "I", "subtype": "cross-analyst context leakage", "component_ids": ["C2", "C3"], "flow_ids": ["F3", "F4"], "boundary_ids": [], "asset": "shared triage context and analyst findings", "threat_vector": "Shared context or retrieval state allows one analyst’s processed evidence, alerts, or extracted indicators to be observed or reused by the other analyst without a clear need-to-know control.", "preconditions": ["Both analysts process attacker-influenced evidence in the same workspace.", "Shared context or tooling exposes one analyst’s working state to the other."], "impact": "Sensitive case details or intermediate findings may be disclosed between analysts beyond intended separation.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "Both analyst agents can consume the same evidence and shared context, with limited isolation between their working state.", "justification": "The design facts explicitly describe shared context and limited isolation between the two analyst agents, which is a supported disclosure path inside the triage workspace. The architecture does not specify stronger separation that would reduce this path."}, {"id": "T3", "category": "T", "subtype": "tampering with persisted case records", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": ["B2"], "asset": "case store records", "threat_vector": "Coordinator writes attacker-influenced or malformed findings into the case store, altering the integrity of persisted case notes and triage decisions.", "preconditions": ["An attacker can shape analyst outputs through untrusted evidence or downstream agent behavior.", "The coordinator accepts and writes those outputs with weak review controls."], "impact": "Case records may be corrupted, misleading, or later relied upon as authoritative despite being manipulated upstream.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "The coordinator can persist findings to the case store using permissive write checks and minimal review gating.", "justification": "The write path from the coordinator to the trusted case store crosses B2 and is described as permissive with minimal review gating. That supports a tampering risk to persisted records when upstream findings are attacker-shaped."}], "scope_notes": ["No direct read path from the case store back to the attacker is shown; disclosure from C5 is therefore unsupported here.", "Reverse flows between the same zones are not present, and no additional boundaries beyond B1 and B2 are implied.", "The scenario mentions weak validation and shared context, but does not specify exact exploit mechanics, authentication details, or data retention behavior.", "Attacker access is limited to influencing external evidence content; no privileged internal access is assumed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C4"], "recommendation": "Add strict evidence sanitization and provenance tagging at intake, isolate analyst prompts/context from raw embedded content, and require coordinator-side policy checks before persisting any agent-generated conclusion; block tool use or write-through actions when content is untrusted or policy-unverified.", "verification": "Submit benign evidence samples containing harmless prompt-like text and confirm they are rendered as inert data, provenance labels are preserved end-to-end, and the coordinator refuses to persist outputs that lack required policy metadata.", "existing_control_effect": "CTRL1 already weakly validates incoming evidence, CTRL2 provides shared workspace processing, and CTRL3 gates writes; these controls reduce trivial passthrough but do not reliably stop instruction injection or downstream propagation.", "residual_risk": "A determined attacker may still influence analyst interpretation through subtle content shaping, so some unsafe reasoning or misleading conclusions may remain possible even after filtering and gating."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Separate analyst working state with per-agent context partitions and least-privilege retrieval, and restrict cross-analyst sharing to explicit case artifacts approved for both users; log any context reuse for review.", "verification": "Run a benign shared-case exercise and confirm each analyst can only access their own working notes by default, while approved shared artifacts are visible to both and cross-access attempts are denied or logged.", "existing_control_effect": "CTRL2 already indicates shared context with limited isolation, which supports collaboration but leaves leakage risk; it does not establish a need-to-know boundary between analysts.", "residual_risk": "Some information may still be inferable from common evidence or approved shared outputs, so complete non-disclosure between analysts is not guaranteed."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Require schema validation, authoritative-field signing or approval, and a two-step review before coordinator writes to the case store; reject malformed or unreviewed findings and retain immutable audit logs for every persisted change.", "verification": "Use benign test records to confirm only schema-valid, reviewed entries are accepted, rejected writes leave no partial updates, and audit logs record who approved and what was stored.", "existing_control_effect": "CTRL3 already provides permissive write checks and minimal review gating, which allows persistence but only weakly constrains integrity failures.", "residual_risk": "If upstream analyst outputs are manipulated in a way that still satisfies validation and review, incorrect but well-formed records may still be persisted."}]}
{"id": "E021", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "single_prompt", "trial_id": "slot-16-single_prompt", "topology_family_id": "topology-018", "annotation_status": "candidate", "image_path": "images/comparison/E021.png", "title": "incident_response", "purpose": "Two analyst agents process externally supplied incident evidence and send findings through a coordinator to a case store.", "assumptions": ["Synthetic design; no actual production data.", "Attacker can influence external documents and attachments delivered to the intake gateway, but has no privileged access to workspace or storage.", "Analysts may use bounded tools for parsing and summarization, with coordinator-approved writes to case storage."], "controls": [{"id": "CTL1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External submissions are accepted only through the intake gateway, and analyst tools reject unsigned attachments unless provenance metadata is present."}, {"id": "CTL2", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "fact": "Analyst agents run with read-only access to evidence and cannot directly write to the case store."}, {"id": "CTL3", "component_ids": ["C4"], "flow_ids": ["F3", "F4", "F5", "F6"], "fact": "The coordinator applies role-based authorization before persisting case records and routes uncertain items to a review queue."}, {"id": "CTL4", "component_ids": ["C5", "C6"], "flow_ids": ["F5", "F6"], "fact": "Case updates are append-only and review items are isolated from final case records."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake Gateway", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Review Queue", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "findings, extracted indicators, analyst notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "findings, extracted indicators, analyst notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case records, triage decisions, evidence references", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C4", "target": "C6", "data": "items needing secondary review", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Evidence Sources"}, {"id": "Z2", "label": "Analysis Workspace"}, {"id": "Z3", "label": "Case Management"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction manipulation via external evidence", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "analyst findings and downstream triage integrity", "threat_vector": "Attacker-controlled evidence text may embed instructions that steer Analyst A during parsing and summarization.", "preconditions": ["Attacker can supply crafted incident documents through the evidence intake path.", "Analyst A follows embedded instructions despite the intended analysis task."], "impact": "Incorrect findings may be produced and forwarded to the coordinator, degrading incident triage quality and contaminating case records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTL1"], "evidence_quote": "External submissions are accepted only through the intake gateway, and analyst tools reject unsigned attachments unless provenance metadata is present.", "justification": "The architecture explicitly admits externally supplied evidence to the analysts; provenance checks exist, but the fact does not limit instruction-following inside the content, so the path remains exposed."}, {"id": "T2", "category": "T", "subtype": "evidence tampering before analysis", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "evidence integrity", "threat_vector": "A malicious sender can alter attachments or logs so analysts process falsified incident material.", "preconditions": ["Attacker can influence external documents and attachments delivered to the intake gateway.", "No stronger validation than provenance metadata is available for the supplied artifact."], "impact": "Analysts may base conclusions on altered material, leading to incorrect incident conclusions or missed indicators.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTL1"], "evidence_quote": "External submissions are accepted only through the intake gateway, and analyst tools reject unsigned attachments unless provenance metadata is present.", "justification": "Provenance metadata and unsigned-attachment rejection reduce but do not eliminate tampering risk; the design still relies on externally supplied content reaching analysts."}, {"id": "T3", "category": "R", "subtype": "findings provenance dispute", "component_ids": ["C2", "C3", "C4", "C5"], "flow_ids": ["F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "case record accountability", "threat_vector": "A contributor may later deny authorship or content of submitted findings if attribution is not preserved end to end.", "preconditions": ["Multiple analysts submit findings through a shared coordinator.", "The stored case record may not fully preserve who produced each assertion."], "impact": "Investigators may be unable to attribute conclusions or reconstruct decision history during incident review.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL3", "CTL4"], "evidence_quote": "The coordinator applies role-based authorization before persisting case records and routes uncertain items to a review queue.", "justification": "Authorization and append-only storage improve accountability, but the supplied facts do not state explicit signing, attribution retention, or immutable audit trails for each analyst contribution."}, {"id": "T4", "category": "D", "subtype": "review-queue saturation", "component_ids": ["C1", "C4", "C6"], "flow_ids": ["F1", "F6"], "boundary_ids": ["B1", "B2"], "asset": "availability of incident processing", "threat_vector": "A flood of malformed or ambiguous evidence can force excessive routing into the review queue and slow coordinator processing.", "preconditions": ["Attacker can send many externally supplied documents.", "Coordinator routes uncertain items to review rather than auto-accepting them."], "impact": "Backlog growth can delay triage and case updates for legitimate incidents.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTL3"], "evidence_quote": "The coordinator applies role-based authorization before persisting case records and routes uncertain items to a review queue.", "justification": "Routing uncertain items to review is a deliberate control, but its capacity, rate limits, and queue isolation behavior are unspecified, so availability impact cannot be reduced further from the given facts."}], "scope_notes": ["No direct path from analysts to final case storage is shown; writes are mediated by the coordinator.", "Boundary B1 covers flows from Z1 to Z2; B2 covers flows from Z2 to Z3, based on first-seen zone pairs.", "The supplied controls describe provenance checks, read-only analyst access, coordinator authorization, and append-only storage, but do not specify signature verification, content sanitization, rate limiting, or audit log retention.", "Unsupported paths: direct analyst writes to the case store are not present in the architecture.", "Residual uncertainty remains around how well provenance metadata prevents malicious instructions embedded in otherwise valid evidence."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C4"], "recommendation": "Require content sanitization and instruction-stripping at the intake gateway, and have the coordinator reject analyst outputs that contain quoted or copied control-language from evidence unless separately justified by workflow metadata.", "verification": "Submit benign sample documents containing embedded commands and confirm they are quarantined or transformed without reaching analyst prompts unchanged.", "existing_control_effect": "The current provenance check limits some untrusted inputs, but it does not address embedded instructions.", "residual_risk": "Attackers may still influence analyst judgment through subtle framing or evidence ambiguity, so human review remains necessary."}, {"threat_id": "T2", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Add mandatory artifact hashing and signature validation at intake, and require analysts to flag any mismatched provenance before evidence is forwarded.", "verification": "Replay a harmless test attachment with a modified hash or missing signature and confirm it is rejected before analysis begins.", "existing_control_effect": "Unsigned-attachment rejection already reduces some tampering, but provenance metadata alone may be insufficient against crafted content.", "residual_risk": "If attackers can produce validly signed but misleading evidence, semantic tampering remains possible."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C4", "C5"], "recommendation": "Persist analyst identity, source document references, and coordinator decision stamps with each case update, and ensure append-only records include per-contribution attribution fields.", "verification": "Create a benign case with two analysts contributing and confirm the stored record preserves distinct authorship and decision history end to end.", "existing_control_effect": "Role-based authorization and append-only storage support accountability, but they do not by themselves prove non-repudiation.", "residual_risk": "Disputes can still arise if analysts reuse accounts or if upstream evidence provenance is incomplete."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C6"], "recommendation": "Set bounded review-queue capacity, per-sender throttling, and alerting for repeated ambiguous submissions so the coordinator can shed or defer excess load safely.", "verification": "Use a harmless burst of ambiguous test items and confirm queue limits, throttling, and alerts activate without blocking normal cases.", "existing_control_effect": "Routing uncertain items to a review queue provides a containment step, but the capacity and rate controls are unspecified.", "residual_risk": "A sustained high-volume flood may still consume analyst attention even with queue limits, so operational monitoring is still needed."}]}
{"id": "E022", "collection": "comparison", "domain": "finance_operations", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-08-full_workflow", "topology_family_id": "topology-019", "annotation_status": "candidate", "image_path": "images/comparison/E022.png", "title": "Invoice intake to payment-approval pipeline with weak document controls", "purpose": "Process externally supplied invoices into payment drafts, then route them through human approval and payment execution under weak intake and validation controls.", "assumptions": ["Externally supplied invoice documents and supporting evidence arrive from untrusted senders and may contain attacker-controlled text or attachments.", "The invoice agent can read those documents and generate payment drafts, but it does not itself have privileged payment execution rights.", "A separate approval step exists before payment execution, but the control is permissive/weak and may rely on the draft content and attached evidence.", "The payment-execution component is distinct from the invoice agent and approval step, with no assumption of strong schema validation or strong segregation of duties.", "The attacker can influence document contents and evidence, but has no direct access to internal payment systems or administrative credentials.", "Intake validation and attachment screening are incomplete or inconsistently enforced."], "controls": [{"id": "K1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "Inbound invoice handling accepts external documents with only weak validation; untrusted text and attachments are ingested into the agent workflow."}, {"id": "K2", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "fact": "The agent-generated payment draft is forwarded to the approval queue with permissive review that may rely on the draft and attached evidence rather than independent re-checks."}, {"id": "K3", "component_ids": ["C3", "C4"], "flow_ids": ["F3"], "fact": "Payment execution occurs only after a human or policy approval step, but segregation and schema checks are not assumed strong."}, {"id": "K4", "component_ids": ["C4", "C3"], "flow_ids": ["F4"], "fact": "Execution feedback returns to the same approval queue for reconciliation and exception handling, creating a shared case state across the downstream path."}], "architecture": {"components": [{"id": "C1", "label": "Supplier inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Approval queue", "zone": "Z3", "kind": "process"}, {"id": "C4", "label": "Payment executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "payment draft, extracted fields, attached evidence summary", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C3", "target": "C4", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C3", "data": "payment status and exception result", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External sender boundary"}, {"id": "Z2", "label": "Invoice processing zone"}, {"id": "Z3", "label": "Approval and payment zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "document/attachment tampering leading to payment draft manipulation", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1", "B2"], "asset": "payment instruction and invoice case state", "threat_vector": "Attacker supplies crafted invoice text or attachments that alter extraction results, then the altered draft is reviewed and approved.", "preconditions": ["Attacker can send invoice documents and supporting evidence to the external inbox.", "Intake validation and attachment screening do not reliably block malicious content."], "impact": "Incorrect payment draft can be approved and executed, causing unauthorized or incorrect payment.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K3"], "evidence_quote": "Inbound invoice handling accepts external documents with only weak validation; untrusted text and attachments are ingested into the agent workflow.", "justification": "The external sender boundary is crossed by untrusted documents into the agent workflow, and the draft is then forwarded with permissive review. Weak validation plus reliance on the draft/evidence supports a plausible tampering path to payment execution."}, {"id": "T2", "category": "A", "subtype": "instruction/context manipulation of invoice agent", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1", "B2"], "asset": "agent-produced payment draft", "threat_vector": "Attacker embeds prompt-injection or misleading instructions in invoice content so the invoice agent generates a malicious or distorted payment draft.", "preconditions": ["Attacker controls invoice document contents or attachments.", "The invoice agent processes untrusted text from the sender boundary without strong isolation or robust instruction filtering."], "impact": "The agent may be steered to produce an attacker-favorable payment draft or altered extracted fields, which can then reach approval.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "Inbound invoice handling accepts external documents with only weak validation; untrusted text and attachments are ingested into the agent workflow.", "justification": "This is an agent-specific attack on the invoice agent using externally supplied content. The architecture explicitly allows untrusted text into the agent workflow and forwards the resulting draft to permissive review, so the manipulation path is supported."}, {"id": "T3", "category": "I", "subtype": "sensitive invoice/evidence disclosure through approval feedback state", "component_ids": ["C3", "C4"], "flow_ids": ["F4"], "boundary_ids": [], "asset": "payment status, exception result, and shared case state", "threat_vector": "A sender-induced or malformed case causes detailed execution feedback and exception handling to reveal internal processing details in the shared approval queue state.", "preconditions": ["Attacker can influence invoice contents to trigger unusual execution outcomes.", "The approval queue state is accessible to parties who should not see internal exception detail, or that detail is later exposed outside the intended process."], "impact": "Internal payment status or exception details may be disclosed through shared case handling, aiding follow-on abuse or revealing process information.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Execution feedback returns to the same approval queue for reconciliation and exception handling, creating a shared case state across the downstream path.", "justification": "The architecture shows feedback returning to a shared approval queue, but it does not specify who can read that state or whether disclosure actually leaves the trusted zone. The path exists, but unauthorized recipient exposure is not established."}], "scope_notes": ["Boundary IDs follow the supplied flow list: F1 crosses B1, F2 crosses B2, and F3/F4 are same-zone flows with no boundary.", "No direct path to administrative credentials or internal payment-system access is described.", "The architecture supports a weak review and shared case-state pattern, but it does not specify concrete authorization on the approval queue or the exact content of exception messages.", "Unsupported paths include direct external access to C3 or C4, and direct payment execution without the approval flow."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Enforce strict intake sanitization and attachment quarantine at the invoice agent, require deterministic schema/field validation before drafting, and add an independent approval re-check that compares the draft against normalized source fields rather than trusting attached evidence alone.", "verification": "Submit benign sample invoices with harmless attachments and confirm the agent only accepts allowed file types, strips active content, rejects malformed fields, and that approval flags any draft that does not match the normalized invoice data.", "existing_control_effect": "K1 already provides weak inbound validation and K2/K3 provide a review gate, but they do not reliably prevent crafted content from altering extraction or being approved.", "residual_risk": "A determined sender may still craft invoices that pass basic validation or exploit reviewer error, so some incorrect-payment risk remains if independent re-checks are bypassed or inconsistently used."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Separate untrusted invoice text from agent instructions, treat document content as data only, disable any tool or prompt instructions derived from the invoice body, and require a human or rules-based verifier to approve only normalized extracted fields.", "verification": "Use harmless test invoices containing misleading phrases and confirm the agent ignores them as instructions, produces the same draft as a plain-text equivalent, and that the approval queue receives only normalized fields with no embedded instruction text.", "existing_control_effect": "K1 admits untrusted text into the agent workflow and K2 forwards the generated draft to permissive review, so existing controls reduce but do not eliminate instruction-manipulation risk.", "residual_risk": "If the agent still interprets attacker-controlled text indirectly or the reviewer relies on the draft without independent checks, the attack can still distort the payment draft."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C4"], "recommendation": "Minimize and classify execution feedback in the approval queue, restrict exception visibility to authorized operators only, and store reconciliation details in a separate access-controlled record instead of the shared case state.", "verification": "Trigger benign payment exceptions in a test environment and confirm the approval queue shows only non-sensitive status codes, while detailed diagnostics are visible only to authorized staff in the restricted record.", "existing_control_effect": "K4 already routes execution feedback back to the approval queue for reconciliation, but it does not define strong access restrictions or message minimization, so disclosure control is only partial.", "residual_risk": "Some status information will still be visible to the approval workflow, and any future expansion of exception detail or access scope could reintroduce disclosure."}]}
{"id": "E024", "collection": "comparison", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-09-no_reviewer", "topology_family_id": "topology-020", "annotation_status": "candidate", "image_path": "images/comparison/E024.png", "title": "Guarded invoice-to-payment assistant with approval handoff", "purpose": "A single connected finance workflow where an invoice-reading agent ingests externally supplied documents, extracts payment details, prepares payment drafts, and forwards them into a separate approval and payment-execution path under explicit authorization and provenance controls.", "assumptions": ["Externally supplied invoices and supporting evidence arrive from outside the trusted application zone and may be attacker-influenced.", "The invoice agent can read documents but has bounded tool authority: it can draft, classify, and route work, but cannot directly execute payments.", "Approval and payment execution are separate actions with enforced authorization checks, though exact effectiveness is not fully specified.", "Provenance checks exist for document ingestion, but residual risk remains for forged or manipulated invoice content.", "A single shared case record or workflow state is maintained across the drafting and approval stages."], "controls": [{"id": "K1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "Ingress accepts only provenance-tagged invoice submissions; the agent processes them in read-only mode and cannot call payment tools from the ingest path."}, {"id": "K2", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "fact": "The agent writes extracted fields and draft notes into a shared case record, preserving source notes for later review."}, {"id": "K3", "component_ids": ["C3", "C4"], "flow_ids": ["F3"], "fact": "Approval review consumes the stored draft and case history before any execution request is formed."}, {"id": "K4", "component_ids": ["C4", "C6", "C5"], "flow_ids": ["F4", "F5", "F6"], "fact": "Payment execution requires an approval decision from the review service and a positive authorization response; the executor does not accept direct agent-issued payment commands."}, {"id": "K5", "component_ids": ["C5", "C3"], "flow_ids": ["F7"], "fact": "Execution results are appended back to the case record for traceability and later reconciliation."}], "architecture": {"components": [{"id": "C1", "label": "Invoice intake gateway", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice reading agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Case record store", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Approval review service", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment execution service", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Authorization service", "zone": "Z3", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "parsed invoice fields, provenance notes, draft payment case state", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "payment draft and case history", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C6", "data": "approval request and authorization context", "boundary_ids": []}, {"id": "F5", "source": "C6", "target": "C4", "data": "approval decision or denial", "boundary_ids": []}, {"id": "F6", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F7", "source": "C5", "target": "C3", "data": "payment status and execution record", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "Untrusted document ingress"}, {"id": "Z2", "label": "Trusted invoice processing"}, {"id": "Z3", "label": "Controlled finance execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "invoice content tampering leading to wrong payment draft", "component_ids": ["C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F6"], "boundary_ids": ["B1", "B2"], "asset": "payment case integrity", "threat_vector": "Attacker-supplied invoice documents or supporting evidence are manipulated before ingestion, causing the invoice-reading agent to extract incorrect payment details that propagate into the shared case record and approval path.", "preconditions": ["Attacker can supply or alter invoice documents in the untrusted ingress channel.", "Manipulated content is not rejected by provenance checks or still produces plausible extracted fields."], "impact": "Incorrect payment drafts may be approved and executed, resulting in payment to the wrong amount, recipient, or terms.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K3", "K4"], "evidence_quote": "Ingress accepts only provenance-tagged invoice submissions; the agent processes them in read-only mode and cannot call payment tools from the ingest path.", "justification": "The architecture explicitly accepts externally supplied invoices from the untrusted zone and preserves extracted fields into a shared case record that drives approval and execution. Provenance tagging and read-only ingest reduce direct tool abuse, but residual risk from forged or manipulated invoice content is explicitly acknowledged in the assumptions."}, {"id": "T2", "category": "T", "subtype": "case record tampering propagating to approval", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F2", "F3"], "boundary_ids": ["B2"], "asset": "case record integrity", "threat_vector": "A compromised or faulty invoice-reading agent writes incorrect parsed fields, provenance notes, or draft state into the shared case record, and the approval service consumes that record as the basis for execution decisions.", "preconditions": ["The agent can write to the shared case record.", "The stored record is trusted by the approval review service as input to its decision process."], "impact": "Tampered case content can bias or mislead review, enabling inappropriate approval or denial and corrupting the audit trail.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The agent writes extracted fields and draft notes into a shared case record, preserving source notes for later review.", "justification": "There is a direct write path from the agent to shared state, and the approval service explicitly consumes that state before forming an execution request. This creates a supported integrity exposure on the shared workflow record; the cited controls preserve notes for review but do not state validation strong enough to eliminate the path."}, {"id": "T3", "category": "T", "subtype": "authorization context tampering", "component_ids": ["C4", "C6", "C5"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": [], "asset": "approval and execution authorization integrity", "threat_vector": "An attacker influences the approval request or authorization context so that the review service forms or receives an authorization state inconsistent with the original invoice evidence, leading to an incorrect approval decision.", "preconditions": ["The approval request can be shaped by prior stored state or context fields.", "Authorization-service validation does not fully bind the decision to the intended invoice and review context."], "impact": "Improper authorization context could allow an execution instruction to be issued when it should not be, or block valid payments.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Payment execution requires an approval decision from the review service and a positive authorization response; the executor does not accept direct agent-issued payment commands.", "justification": "The path between review, authorization, and execution is clearly present, but the exact effectiveness of the authorization checks is not fully specified. The control reduces direct agent command injection, yet the architecture does not state how strongly the authorization context is bound to invoice identity or whether tampering is detected."}], "scope_notes": ["No direct cross-boundary flow from the invoice agent to the payment executor exists; direct agent-issued payment commands are explicitly blocked.", "The architecture does not specify whether the provenance-tagging check rejects forged documents or only records provenance.", "The exact authorization logic and its binding to invoice identity, case history, or review context are unspecified.", "The shared case record is a single stateful asset across drafting and approval, so integrity threats are in scope even when no confidentiality breach is shown.", "No supported recipient for confidential data disclosure beyond the case record and review path is described, so standalone disclosure threats are not asserted."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Require the invoice-reading agent to attach an immutable source hash and extracted-field justification to each draft, and make the approval review service verify those fields against the stored original document before any approval can proceed; reject drafts that lack a matching provenance record or that change recipient/amount/terms without a human-reviewed exception.", "verification": "Submit a benign test invoice, then repeat with a modified copy that changes only the payee or amount; confirm the review service blocks the altered draft, logs the mismatch, and preserves the original source record for comparison.", "existing_control_effect": "K1 and K2 already limit the ingest path to provenance-tagged inputs and keep source notes in the shared case record, which helps review but does not by itself ensure the extracted values match the source content.", "residual_risk": "A forged document that passes provenance tagging or a semantically plausible manipulation may still influence extraction quality; the control reduces but does not eliminate malicious invoice-content tampering."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Protect the shared case record with field-level write authorization and append-only change history, and require the approval review service to read only signed, versioned case snapshots rather than mutable live state; any agent-written update must be validated against the prior snapshot before it can affect approval.", "verification": "Create a normal draft update and confirm it appears as a new signed version; then attempt a direct overwrite of an existing payment field and confirm the store rejects it or records it as a non-authoritative change that the review service ignores.", "existing_control_effect": "K2 and K3 already keep source notes in shared state and have the review service consume stored draft/history before execution, which supports traceability but does not prevent unauthorized or faulty writes to the record.", "residual_risk": "A compromised agent can still generate misleading but syntactically valid updates, so integrity risk remains if validation rules or signing are bypassed."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C6", "C5"], "recommendation": "Bind the approval decision and authorization request to a unique invoice/case identifier, approval timestamp, and immutable review digest, and require the authorization service to verify that the digest matches the exact reviewed case before returning approval for execution; reject any request whose context does not match the stored reviewed snapshot.", "verification": "Run a normal approval flow and confirm the authorization response is accepted only when the case digest and identifier match; then resend the same approval request with a changed case identifier or altered context and confirm the authorization service denies it.", "existing_control_effect": "K4 already prevents direct agent-issued payment commands and requires both an approval decision and a positive authorization response, which blocks the simplest bypass path.", "residual_risk": "If the authorization binding is only partially enforced or the digesting is weak, a malformed or replayed context may still cause incorrect approval or denial, so some authorization-context tampering risk remains."}]}
{"id": "E025", "collection": "comparison", "domain": "finance_operations", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-12-full_workflow", "topology_family_id": "topology-021", "annotation_status": "candidate", "image_path": "images/comparison/E025.png", "title": "Invoice-to-Approval Payment Draft Workflow", "purpose": "Model a finance operations flow where an invoice-handling agent ingests externally supplied invoice documents, extracts payment-relevant details, drafts payment instructions, and passes them into an approval and payment-execution path.", "assumptions": ["Externally supplied invoice documents and supporting evidence are outside the trusted application zone.", "The invoice agent can read documents and generate payment drafts, but the implementation and effectiveness of controls around parsing, validation, approval, and execution are unspecified.", "The attacker has no privileged internal access but can influence the content of incoming invoices or supporting evidence.", "Approval and execution remain distinct downstream functions in the same connected workflow."], "controls": [{"id": "CTRL1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Invoice intake and draft generation are separated from approval; the effectiveness of parsing and validation checks is unspecified."}, {"id": "CTRL2", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "A downstream approval review step exists before payment execution; its exact review rules and enforcement strength are unspecified."}, {"id": "CTRL3", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "Payment execution is handled by a separate process that receives only approved instructions; rejection or status feedback returns to approval."}], "architecture": {"components": [{"id": "C1", "label": "Invoice Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Draft Builder", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Approval Review", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "extracted invoice fields and payment details", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "payment draft for review", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F5", "source": "C5", "target": "C4", "data": "execution status or rejection feedback", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Invoice Processing"}, {"id": "Z3", "label": "Approval and Payment Ops"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "invoice-content tampering / prompt-injection into extraction", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1", "B2"], "asset": "payment draft integrity", "threat_vector": "Attacker-controlled invoice documents or supporting evidence are crafted to alter extracted fields and steer the agent into producing a manipulated payment draft.", "preconditions": ["Attacker can influence the content of incoming invoices or supporting evidence.", "The invoice agent accepts and processes external documents.", "Parsing and validation effectiveness is unspecified."], "impact": "Incorrect payment details may be carried into the draft and forwarded for approval, enabling fraudulent or misdirected payment instructions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "Invoice intake and draft generation are separated from approval; the effectiveness of parsing and validation checks is unspecified.", "justification": "The architecture explicitly allows external invoice content to reach the agent and draft builder across B1 and then B2, while the control fact leaves parsing/validation effectiveness unspecified, so the manipulation path is supported."}, {"id": "T2", "category": "S", "subtype": "spoofed invoice source / fraudulent payment request", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1", "B2"], "asset": "approval decision integrity", "threat_vector": "An external attacker submits invoice material that impersonates a legitimate business request, causing the workflow to treat it as a valid payment candidate.", "preconditions": ["Attacker can influence inbound invoice documents.", "The system does not have specified source authentication or authenticity checks on incoming invoices.", "Approval review may rely on the drafted payment package."], "impact": "A fraudulent payment request can be approved and passed to execution if reviewers accept the spoofed material.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2"], "evidence_quote": "A downstream approval review step exists before payment execution; its exact review rules and enforcement strength are unspecified.", "justification": "There is a clear approval path from draft to payment execution, but the strength of source-authenticity checks and reviewer enforcement is unspecified, so spoofing is plausible but not fully established as exposed."}, {"id": "T3", "category": "D", "subtype": "workflow overload via malicious invoices", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "invoice processing availability", "threat_vector": "The attacker floods intake with large or pathological invoice payloads to consume agent and draft-generation resources.", "preconditions": ["Attacker can send or influence many incoming invoice documents.", "Resource limits, rate limits, and parsing safeguards are unspecified."], "impact": "Invoice processing may slow, stall, or fail, delaying draft creation and downstream approval.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "Invoice intake and draft generation are separated from approval; the effectiveness of parsing and validation checks is unspecified.", "justification": "The architecture shows a processing path for external invoices, but it does not specify capacity controls; availability impact is plausible yet control strength is unknown."}, {"id": "T4", "category": "I", "subtype": "invoice data exposure through draft/feedback handling", "component_ids": ["C2", "C3", "C4", "C5"], "flow_ids": ["F2", "F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "invoice and payment details", "threat_vector": "Sensitive invoice fields and payment details are propagated into drafts and status feedback, potentially exposing financial data within the workflow.", "preconditions": ["The draft and approval participants handle sensitive payment details.", "Access controls for internal viewing of drafts and execution feedback are unspecified."], "impact": "Internal parties or components may learn more payment information than necessary.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL3"], "evidence_quote": "Payment execution is handled by a separate process that receives only approved instructions; rejection or status feedback returns to approval.", "justification": "The flows show payment data moving through approval and execution, and status feedback returning, but there is no explicit disclosure control described, so disclosure risk is not proven exposed."}], "scope_notes": ["No candidate was added for unauthorized payment execution because the workflow requires an approval step before C5 and the enforcement strength is unspecified rather than clearly absent.", "No same-zone flow boundary exists for F2, F4, or F5, so those paths were not treated as crossing trust boundaries.", "The evidence quote had to match a supplied control fact exactly; controls describing unspecified effectiveness were used as evidence without extrapolating beyond them.", "Unsupported paths such as direct external-to-execution access are not shown in the architecture and were not assumed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Enforce strict document authentication, schema validation, and field-level sanity checks before extracted values can be promoted into a payment draft; reject or quarantine invoices that contain conflicting totals, unusual payees, or embedded instructions that are not part of the expected invoice schema.", "verification": "Submit benign test invoices with malformed fields, duplicate totals, and text-only instruction blocks; confirm they are rejected or held for review and never appear in the draft output.", "existing_control_effect": "Invoice intake and draft generation are separated from approval; the effectiveness of parsing and validation checks is unspecified.", "residual_risk": "Well-formed but fraudulent invoices may still pass syntactic checks and require human review to catch business-logic fraud."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C4"], "recommendation": "Require source authentication and provenance checks for incoming invoices, and have approval review verify the originating vendor identity and payment context against an allowlist or purchase-order record before approval is granted.", "verification": "Use benign test submissions from an unregistered sender and from a registered sender with altered vendor details; confirm the former is blocked and the latter is flagged for review before approval can proceed.", "existing_control_effect": "A downstream approval review step exists before payment execution; its exact review rules and enforcement strength are unspecified.", "residual_risk": "A sophisticated attacker may still imitate legitimate business context closely enough that a reviewer approves a fraudulent request."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Add rate limiting, payload-size limits, timeout controls, and queue backpressure on invoice intake and parsing so the agent and draft builder can shed abusive load without exhausting resources.", "verification": "Send a small burst of oversized and repeated benign test invoices and confirm they are throttled, deferred, or rejected while normal-volume invoices continue to process.", "existing_control_effect": "Invoice intake and draft generation are separated from approval; the effectiveness of parsing and validation checks is unspecified.", "residual_risk": "Moderate-volume abuse or computationally expensive but valid documents may still degrade throughput before limits trigger."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Minimize and partition payment data shown in drafts and execution feedback, and restrict approval and payment components to role-based access with masking for nonessential fields such as bank details and supporting evidence.", "verification": "Use benign test records and confirm that approval and feedback views show only the minimum required fields, with sensitive values masked or omitted for unauthorized roles.", "existing_control_effect": "Payment execution is handled by a separate process that receives only approved instructions; rejection or status feedback returns to approval.", "residual_risk": "Authorized users still need some payment context, so limited internal disclosure remains possible through legitimate workflow access."}]}
{"id": "E026", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-15-full_workflow", "topology_family_id": "topology-022", "annotation_status": "candidate", "image_path": "images/comparison/E026.png", "title": "Guarded multi-analyst incident triage with coordinator and case store", "purpose": "Two analyst agents examine externally supplied incident evidence, reconcile findings through a coordinator, and persist case records and evidence summaries with enforced authorization, provenance checks, and bounded tool use.", "assumptions": ["Externally supplied incident documents may be attacker-influenced and are untrusted at ingress.", "Analyst agents can read evidence and produce findings, but only the coordinator can commit records to the case store.", "Tool access is scoped to specific retrieval and write actions; no unrestricted shell or network authority is assumed.", "Authorization and provenance checks are enforced, but their completeness against prompt-injection or malformed evidence is not guaranteed from the given facts.", "Attacker capability is limited to supplying crafted incident evidence and embedded instructions; no privileged access to internal agents, coordinator, or case store is assumed."], "controls": [{"id": "CTRL1", "component_ids": ["C5", "C2", "C3"], "flow_ids": ["F2", "F3"], "fact": "Tool gateway restricts analyst reads to scoped retrieval actions and passes provenance metadata with each evidence package; analysts do not receive direct shell or network access."}, {"id": "CTRL2", "component_ids": ["C4", "C6"], "flow_ids": ["F6"], "fact": "Only the coordinator is authorized to write case records to the repository, and writes are gated by an approval step before persistence."}, {"id": "CTRL3", "component_ids": ["C1", "C5"], "flow_ids": ["F1"], "fact": "Ingress performs provenance and authorization checks on incoming documents before the evidence is distributed to analysts, but the design does not state full filtering of embedded instructions."}, {"id": "CTRL4", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F4", "F5", "F6"], "fact": "Findings from both analysts are reconciled by the coordinator before commit, providing a cross-check on inconsistent or low-confidence evidence interpretations."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "agent"}, {"id": "C5", "label": "Tool Gateway", "zone": "Z2", "kind": "process"}, {"id": "C6", "label": "Case Repository", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C5", "data": "incident documents, attachments, metadata", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C5", "target": "C2", "data": "scoped evidence retrieval, provenance metadata", "boundary_ids": []}, {"id": "F3", "source": "C5", "target": "C3", "data": "scoped evidence retrieval, provenance metadata", "boundary_ids": []}, {"id": "F4", "source": "C2", "target": "C4", "data": "findings, confidence notes, cited evidence", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C4", "data": "findings, confidence notes, cited evidence", "boundary_ids": []}, {"id": "F6", "source": "C4", "target": "C6", "data": "approved case record, evidence summary, decision log", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "Untrusted Evidence Sources"}, {"id": "Z2", "label": "Triage Workspace"}, {"id": "Z3", "label": "Case Store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt-injection via untrusted evidence", "component_ids": ["C1", "C5", "C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "analyst reasoning and scoped evidence interpretation", "threat_vector": "Crafted incident documents include embedded instructions that are ingested as evidence and then delivered to analysts through the tool gateway.", "preconditions": ["Attacker can supply incident documents into Evidence Intake.", "Analysts rely on retrieved evidence content during triage."], "impact": "Untrusted instructions may alter analyst interpretation or tool-use decisions, leading to unsafe triage conclusions or manipulated findings.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "Ingress performs provenance and authorization checks on incoming documents before the evidence is distributed to analysts, but the design does not state full filtering of embedded instructions.", "justification": "The architecture explicitly allows externally supplied, attacker-influenced documents to cross into the triage workspace, and the control fact says embedded instructions are not fully filtered. This supports a feasible instruction-manipulation path even though direct shell/network access is removed."}, {"id": "T2", "category": "I", "subtype": "evidence content disclosure to analysts", "component_ids": ["C1", "C5", "C2", "C3"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1"], "asset": "incident evidence and provenance metadata", "threat_vector": "Sensitive or over-broad evidence from an untrusted submission is distributed to analysts for scoped retrieval and review.", "preconditions": ["The submitted package contains sensitive data beyond the intended audience.", "Analysts can view retrieved evidence and provenance metadata."], "impact": "Information present in the uploaded evidence may be exposed to analysts who are not intended recipients of all content in the submission.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL3"], "evidence_quote": "Tool gateway restricts analyst reads to scoped retrieval actions and passes provenance metadata with each evidence package; analysts do not receive direct shell or network access.", "justification": "The path from intake to analyst review is explicit, but the gateway limits reads to scoped retrieval and adds provenance metadata. That control reduces, but does not eliminate, exposure of evidence content to analysts; the exact retrieval scope is unspecified."}, {"id": "T3", "category": "T", "subtype": "tampering with case record content", "component_ids": ["C2", "C3", "C4", "C6"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "case record integrity", "threat_vector": "A malicious or mistaken analyst finding can influence the coordinator’s approved record before persistence to the case store.", "preconditions": ["At least one analyst produces misleading or altered findings.", "Coordinator approval is based on the submitted findings and reconciliation process."], "impact": "Case records may be written with incorrect evidence summaries or decision logs, corrupting downstream case integrity.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL4"], "evidence_quote": "Findings from both analysts are reconciled by the coordinator before commit, providing a cross-check on inconsistent or low-confidence evidence interpretations.", "justification": "The write path to the case store exists only through the coordinator, and reconciliation is an explicit integrity control. That makes direct tampering harder, but not impossible if both submitted findings are misleading or the reconciliation step misses them."}], "scope_notes": ["No direct attacker path to C6 is shown; only coordinator-mediated writes cross B2.", "No direct analyst-to-case-store write exists in the architecture; unsupported paths should not be inferred.", "Effectiveness of provenance and authorization checks against malformed evidence is unspecified beyond the stated limitation on embedded instructions.", "The model does not show whether analysts can modify evidence packages after retrieval, so post-retrieval tampering is out of scope from the given facts."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C5", "C2", "C3"], "recommendation": "Add explicit instruction-content sanitization and safe-rendering at the Tool Gateway before analyst delivery, and have analysts treat evidence as data-only unless provenance-approved. Preserve scoped retrieval, but strip or neutralize embedded directives, links, and tool-call patterns in packages sent over F2/F3.", "verification": "Feed benign test documents containing harmless pseudo-instructions and confirm they are rendered as inert text, not surfaced as actionable guidance or tool prompts; also confirm provenance metadata is still attached and retrieval scope remains unchanged.", "existing_control_effect": "CTRL3 already performs provenance and authorization checks on ingress, and CTRL1 limits analysts to scoped retrieval with no shell or network access, reducing direct abuse paths.", "residual_risk": "If sanitization misses a crafted instruction pattern or an analyst follows out-of-band guidance, prompt-injection influence may still affect triage decisions."}, {"threat_id": "T2", "enforcement_component_ids": ["C5", "C2", "C3"], "recommendation": "Tighten retrieval scope at the Tool Gateway so analysts receive only the minimum evidence subset needed for the triage task, and keep provenance metadata attached to each item. Review scope rules for over-broad returns from F2/F3 and redact unrelated attachments or fields before delivery.", "verification": "Use a harmless sample case with clearly separable fields and confirm analysts receive only the allowed subset, with excluded fields absent from the delivered package and provenance preserved for each returned item.", "existing_control_effect": "CTRL1 already restricts analyst reads to scoped retrieval actions and passes provenance metadata, which lowers exposure compared with unrestricted access.", "residual_risk": "Some sensitive content may remain present within the authorized subset, and the architecture does not specify field-level redaction completeness."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C6"], "recommendation": "Require the Coordinator to persist only signed or otherwise attributable approved case records after a checklist-based reconciliation of both analyst inputs, and reject writes when findings conflict, lack citations, or exceed expected confidence bounds. Keep the repository write gate exclusive to C4.", "verification": "Submit a benign pair of example findings with a deliberate discrepancy and confirm the Coordinator blocks or flags the write before C6 persists the record; then submit aligned findings and confirm only the approved record is written.", "existing_control_effect": "CTRL2 already gates repository writes through coordinator approval, and CTRL4 adds cross-check reconciliation of analyst findings before commit.", "residual_risk": "If both analysts are misled by the same bad evidence or the coordinator approves an incorrect synthesis, an inaccurate but authorized case record can still be persisted."}]}
{"id": "E027", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "no_reviewer", "trial_id": "slot-02-no_reviewer", "topology_family_id": "topology-023", "annotation_status": "candidate", "image_path": "images/comparison/E027.png", "title": "Agent-assisted support drafting with external docs and account lookup", "purpose": "A support agent reads externally supplied help documents and queries an internal customer-account store to draft a reply for review or sending.", "assumptions": ["The help-document source is attacker-influenced and outside the trusted application zone.", "The customer-account store is an internal trusted data source accessed by the support agent through a normal read API.", "Control strength is weak: document authenticity checks are limited or absent, and the workflow relies mainly on prompt content and operator judgment.", "No privileged attacker access is assumed; influence is limited to supplying or altering external documents/evidence presented to the support workflow.", "The support agent can use account data and external documents together when composing a response, but the final reply is still subject to human review or direct sending."], "controls": [{"id": "K1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "External documents are accepted into the agent workspace with limited authenticity validation; the operator mainly judges plausibility before use."}, {"id": "K2", "component_ids": ["C2", "C4"], "flow_ids": ["F2", "F3"], "fact": "The account lookup path is a normal read-only API call from the support agent to the internal store, with access scoped to the current support case."}, {"id": "K3", "component_ids": ["C2", "C3"], "flow_ids": ["F4"], "fact": "Draft text is passed from the agent to the drafting tool for formatting and review, but the tool does not independently verify external evidence."}, {"id": "K4", "component_ids": ["C3", "C5"], "flow_ids": ["F5"], "fact": "Outgoing responses are visible to a customer-facing channel; sending is permitted after operator review, though some replies may be sent directly in routine cases."}], "architecture": {"components": [{"id": "C1", "label": "Help document source", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Drafting tool", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Account store", "zone": "Z3", "kind": "data_store"}, {"id": "C5", "label": "Customer reply channel", "zone": "Z4", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "help documents / evidence text", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C4", "data": "account lookup query", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C4", "target": "C2", "data": "account details / case context", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C2", "target": "C3", "data": "reply draft content", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C5", "data": "reviewed or sent customer response", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External sources"}, {"id": "Z2", "label": "Support workspace"}, {"id": "Z3", "label": "Internal data"}, {"id": "Z4", "label": "Customer channel"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "external-doc injection alters reply draft", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F4"], "boundary_ids": ["B1"], "asset": "Support reply draft and operator decision-making", "threat_vector": "Attacker-controlled help documents are supplied through the external source and read by the support agent, influencing the drafted response content.", "preconditions": ["Attacker can supply or alter external help documents that enter the support workflow.", "The agent relies on document content and operator judgment because authenticity checks are limited."], "impact": "Incorrect or attacker-shaped reply content may be produced and forwarded for review or sending, causing misinformation or unsafe support actions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "External documents are accepted into the agent workspace with limited authenticity validation; the operator mainly judges plausibility before use.", "justification": "The architecture explicitly allows attacker-influenced external documents into the workspace and says verification is limited, so document content can affect the reply draft. This is a supported tampering path on the drafted response, not just generic external input."}, {"id": "T2", "category": "I", "subtype": "account details combined into outbound response", "component_ids": ["C2", "C4", "C3", "C5"], "flow_ids": ["F2", "F3", "F4", "F5"], "boundary_ids": ["B2", "B3"], "asset": "Account details / case context", "threat_vector": "Support-case account details read from the internal store are incorporated into a customer-facing response and may disclose more than intended if the draft is over-shared.", "preconditions": ["The agent queries the internal account store for the current case.", "The drafted response includes account-derived information.", "The response is sent or reviewed into the customer channel."], "impact": "Sensitive account or case context could be exposed to the customer-facing channel.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K4"], "evidence_quote": "Outgoing responses are visible to a customer-facing channel; sending is permitted after operator review, though some replies may be sent directly in routine cases.", "justification": "There is a real path from internal account data to the customer channel, but the architecture does not specify what data is actually included, nor whether review reliably filters sensitive details. That makes disclosure plausible but not confirmed as exposed."}, {"id": "T3", "category": "A", "subtype": "instruction manipulation via external documents", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F4"], "boundary_ids": ["B1"], "asset": "Agent drafting behavior and reply quality", "threat_vector": "Malicious external documents provide misleading instructions or framing that steer the agent’s drafting process.", "preconditions": ["Attacker can influence the external documents consumed by the agent.", "The agent incorporates document text into its drafting process without strong authenticity checks."], "impact": "The support agent may follow attacker-shaped guidance, producing incorrect, policy-bending, or unsafe replies.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "External documents are accepted into the agent workspace with limited authenticity validation; the operator mainly judges plausibility before use.", "justification": "This is an agent-specific manipulation of the drafting context, which fits category A. The weak authenticity control and reliance on prompt/document content make the instruction-shaping path explicitly supported."}], "scope_notes": ["Boundary IDs were derived from the first-seen zone pairs in flow order: B1 for Z1-Z2, B2 for Z2-Z3, B3 for Z2-Z4.", "No direct threat was added for the account store alone because a read path does not by itself show unauthorized disclosure.", "No denial-of-service or elevation path is supported by the supplied controls and flows.", "The review step on outgoing responses may reduce harm, but its effectiveness is not specified; direct sending is also allowed in routine cases."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Require the support agent to use only help documents that pass source authenticity checks and quarantine unverified external text before it can enter the drafting workflow; the drafting tool should preserve citations to the original source and flag any document-derived claims that lack a trusted source match for human review.", "verification": "Provide a benign test document with an untrusted origin marker and confirm it is blocked or quarantined; then use a trusted document and confirm the draft includes source references and the reviewer sees the provenance flags before sending.", "existing_control_effect": "K1 and K3 already add limited operator judgment and keep the drafting tool from independently verifying evidence, which can slow casual misuse but does not stop document-shaped content from influencing the reply.", "residual_risk": "If a document is trusted by the intake process but still misleading, the agent may still draft an incorrect reply; human review remains a last line of defense."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C5"], "recommendation": "Minimize account data in support drafts by restricting the lookup result to case-necessary fields, then add a response-time redaction check in the drafting tool to block sensitive account details from being included in customer-facing output unless explicitly approved for the case.", "verification": "Use a test account record with both necessary and sensitive fields and confirm the agent can retrieve the limited case fields, the drafting tool flags any attempt to include sensitive fields, and the outgoing channel suppresses unapproved redactions.", "existing_control_effect": "K2 scopes the read API to the current support case, and K4 keeps outgoing responses under operator review in many cases; these controls reduce casual over-sharing but do not ensure the final message excludes sensitive details.", "residual_risk": "If a review step misses an over-shared detail or a routine case is sent directly, some account context may still reach the customer channel."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Treat external help documents as untrusted inputs: separate them from instructions, sanitize or strip imperative text before agent use, and require the support workflow to label any externally sourced directive as non-authoritative unless manually confirmed by the operator.", "verification": "Submit a benign external document containing a harmless imperative statement and confirm the agent workspace labels it as untrusted, the drafting flow does not treat it as operative guidance, and the operator must explicitly confirm any use of it.", "existing_control_effect": "K1 already limits authenticity validation and relies on operator judgment, while K3 prevents the drafting tool from independently verifying evidence; these controls help review but do not prevent instruction-shaped content from steering the agent.", "residual_risk": "A sophisticated external document may still influence drafting if it appears plausible and the operator accepts it, so some manipulation risk remains."}]}
{"id": "E028", "collection": "comparison", "domain": "incident_response", "control_profile": "weak", "generation_condition": "no_reviewer", "trial_id": "slot-13-no_reviewer", "topology_family_id": "topology-005", "annotation_status": "candidate", "image_path": "images/comparison/E028.png", "title": "Weakly controlled incident triage pipeline with two analyst agents", "purpose": "Model an incident-response workflow where two analyst agents review attacker-influenced evidence, exchange findings through a coordinator, and persist case state in a shared case store under permissive controls.", "assumptions": ["Externally supplied incident documents, logs, or attachments can be attacker-crafted and enter outside the trusted application zone.", "Two analyst agents have tool access to read evidence and write findings to an internal coordinator, but they do not have privileged administrative access.", "The coordinator accepts analyst outputs and forwards normalized case updates to a case store.", "Control posture is weak or permissive: validation, authorization checks, and content sanitization are incomplete or inconsistently enforced, but not assumed absent everywhere.", "The case store is an internal state repository reachable only through the coordinator path, not directly from the external source.", "Attacker can influence evidence content and embedded instructions, but has no privileged access to analyst tools or the coordinator beyond what weak validation might allow."], "controls": [{"id": "CTRL1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Analyst inbox parsing allows rich text and attachments; basic file-type checks exist, but sanitization and prompt-injection filtering are incomplete and inconsistently applied."}, {"id": "CTRL2", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4"], "fact": "Analyst findings are accepted over an authenticated internal channel, but the coordinator performs only light schema validation and weak provenance checks before aggregation."}, {"id": "CTRL3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "The coordinator is the only configured writer to the case store; write access is restricted, but update authorization is coarse-grained and does not fully separate analyst-originated content from coordinator state."}, {"id": "CTRL4", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "fact": "External evidence is treated as untrusted input, but the workflow permits analysts to inspect and reference raw content during triage because content redaction is not enforced end-to-end."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Triage Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z2", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, extracted indicators, case notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, extracted indicators, case notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "normalized case updates, triage status, assignments", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Source"}, {"id": "Z2", "label": "Incident Response Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt-injection via attacker-crafted evidence", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1"], "asset": "analyst triage reasoning and downstream case updates", "threat_vector": "Attacker-crafted incident documents/logs/attachments contain embedded instructions that influence analyst agent behavior during evidence review and are then propagated into findings and coordinator aggregation.", "preconditions": ["Attacker can supply or modify external evidence content entering the pipeline.", "Analyst agents inspect raw evidence during triage.", "Weak validation and incomplete prompt-injection filtering are in place."], "impact": "Agent instruction/context manipulation can distort findings, lead to incorrect triage decisions, and poison normalized case updates sent to the coordinator.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL4"], "evidence_quote": "Analyst inbox parsing allows rich text and attachments; basic file-type checks exist, but sanitization and prompt-injection filtering are incomplete and inconsistently applied.", "justification": "The architecture explicitly allows attacker-controlled evidence across B1 into both analyst agents, and the stated controls leave content sanitization incomplete. That supports a feasible A-type manipulation path affecting analyst reasoning and downstream outputs."}, {"id": "T2", "category": "T", "subtype": "tampering of analyst findings in coordinator aggregation", "component_ids": ["C2", "C3", "C4", "C5"], "flow_ids": ["F3", "F4", "F5"], "boundary_ids": [], "asset": "case state in the shared case store", "threat_vector": "Malformed or adversarially shaped analyst outputs are accepted by the coordinator, which then aggregates them into normalized case updates and writes them to the case store.", "preconditions": ["An attacker can influence the contents of analyst findings through crafted evidence or induced analyst output.", "The coordinator relies on light schema validation and weak provenance checks.", "Coordinator-originated writes to the case store are the only configured write path."], "impact": "Integrity of case records may be altered, causing incorrect assignments, status, or indicators to be persisted as authoritative state.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL3"], "evidence_quote": "Analyst findings are accepted over an authenticated internal channel, but the coordinator performs only light schema validation and weak provenance checks before aggregation.", "justification": "The internal path from analysts to coordinator to case store is explicit, and the controls describe weak verification before state is written. That is a supported tampering opportunity on the persisted case state."}, {"id": "T3", "category": "I", "subtype": "exposure of raw evidence through analyst inspection and reference", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1"], "asset": "sensitive incident documents, logs, and attachments", "threat_vector": "Sensitive content from external evidence is made available to analysts for inspection and may be referenced in outputs without end-to-end redaction.", "preconditions": ["Evidence contains sensitive data that should not be broadly exposed.", "Analysts can inspect raw content during triage.", "Referenced content can be carried into analyst findings or case notes."], "impact": "Sensitive data may be disclosed to more users or persisted in internal findings, expanding exposure beyond the original evidence intake path.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL4"], "evidence_quote": "External evidence is treated as untrusted input, but the workflow permits analysts to inspect and reference raw content during triage because content redaction is not enforced end-to-end.", "justification": "The design explicitly allows raw evidence inspection and reference, and the missing end-to-end redaction is a direct exposure for sensitive data handled across B1."}, {"id": "T4", "category": "D", "subtype": "coordinator or case-store overload via excessive evidence-driven updates", "component_ids": ["C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1"], "asset": "availability of the triage pipeline and case store update path", "threat_vector": "Attacker-supplied evidence can trigger large, repeated, or complex analyst outputs that burden the coordinator’s aggregation and the restricted writer path to the case store.", "preconditions": ["Attacker can submit crafted evidence over the external intake boundary.", "Analyst processing and coordinator aggregation are exercised by the supplied content.", "No explicit rate limits or backpressure controls are described."], "impact": "Triage throughput may degrade, delaying or blocking case updates and analyst coordination.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2", "CTRL3", "CTRL4"], "evidence_quote": "The coordinator is the only configured writer to the case store; write access is restricted, but update authorization is coarse-grained and does not fully separate analyst-originated content from coordinator state.", "justification": "The architecture shows a shared coordinator-mediated write path and permissive content handling, but it does not specify rate limiting, queue limits, or overload protection. Availability impact is plausible but the effectiveness of controls against this path is unspecified."}], "scope_notes": ["No direct external path to the case store is described; only the coordinator writes to it, so direct-store disclosure or tampering paths are unsupported.", "No explicit privilege escalation mechanism is described for agents beyond influencing their outputs; E is not separately supported.", "Reverse flows from the coordinator back to analysts are not present, so threats are limited to the listed evidence-to-analysis and analysis-to-store paths.", "Controls are described as weak or incomplete, but not entirely absent; unsupported assumptions about total control failure were not used."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Enforce untrusted-content isolation for analyst review: strip active content from evidence before agent consumption, separate raw evidence from agent instructions, and require the coordinator to reject findings that quote or react to hidden instructions unless they are explicitly tagged as evidence-derived and human-verified.", "verification": "Use a benign test document containing inert canary instructions and confirm the agents treat it as data only, that redacted views are shown by default, and that the coordinator rejects untagged instruction-like text in submitted findings.", "existing_control_effect": "CTRL1 and CTRL4 already provide basic file-type checks and acknowledge evidence as untrusted, which reduces simple injection attempts but leaves incomplete sanitization and end-to-end redaction gaps.", "residual_risk": "A determined attacker may still influence agent reasoning through subtle content shaping or novel formatting that bypasses filters, so manual review and periodic filter updates remain necessary."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Require coordinator-side provenance enforcement before any case-store write: bind each analyst finding to a source identity, evidence reference, and immutable checksum, then only persist updates that pass strict schema and origin checks with rejection logging for malformed or conflicting inputs.", "verification": "Submit a harmless malformed finding and a properly signed/structured finding in a test environment; confirm the malformed update is rejected, the valid update is stored, and the audit trail records the decision and source metadata.", "existing_control_effect": "CTRL2 and CTRL3 already keep writes behind the coordinator and use an authenticated internal channel, which limits direct tampering, but weak provenance and coarse authorization still allow polluted data to be aggregated.", "residual_risk": "If analysts are compromised or can produce misleading but well-formed outputs, the coordinator can still persist deceptive content unless higher-fidelity provenance and human review are added."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Apply end-to-end redaction and minimum-necessary views for external evidence: mask sensitive fields before analyst inspection where possible, block copying of raw secrets into findings by policy checks, and have the coordinator redact persisted notes unless retention of the exact snippet is explicitly required.", "verification": "Use a benign test case with labeled sensitive placeholders and confirm analysts see masked content by default, that attempts to paste raw placeholders into findings are flagged, and that stored case notes preserve only approved redacted excerpts.", "existing_control_effect": "CTRL4 already classifies evidence as untrusted and permits analyst inspection, which supports triage, but it does not enforce redaction end-to-end and therefore does not fully prevent disclosure through copied content.", "residual_risk": "Some sensitive context may still need to be viewed during investigation, so limited exposure remains possible whenever analysts must inspect raw artifacts for accuracy."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Add backpressure and update throttling on the coordinator-to-store path: cap per-case write frequency, queue size, and aggregate payload size, and drop or defer excessive analyst-driven updates with clear retry handling and alerts.", "verification": "Run a benign load test with many small, valid updates and confirm the coordinator starts queueing or deferring after configured thresholds while the case store remains responsive and alerts are emitted.", "existing_control_effect": "CTRL2 and CTRL3 already constrain writes to a single coordinator path, which reduces uncontrolled fan-out, but they do not describe rate limits or overload protection.", "residual_risk": "A sustained high-volume or unusually expensive analysis workload can still slow the pipeline, so operational monitoring and capacity tuning remain important."}]}
{"id": "E029", "collection": "comparison", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-10-full_workflow", "topology_family_id": "topology-024", "annotation_status": "candidate", "image_path": "images/comparison/E029.png", "title": "Guarded invoice-to-payment assistant", "purpose": "Process externally supplied invoices, draft payment requests, and route them through separate approval and execution steps with bounded tool use and authorization controls.", "assumptions": ["Invoice inputs originate outside the trusted application zone and may be attacker-influenced.", "The invoice agent can read documents and generate drafts, but cannot directly execute payments.", "Approval is performed by a separate step with role-based authorization and audit logging.", "Tool use is bounded to predefined document parsing and payment-draft creation functions; payment execution requires a distinct authorized action.", "Provenance checks exist for incoming documents, but their effectiveness may be incomplete against forged or manipulated content."], "controls": [{"id": "K1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F2", "F4"], "fact": "The invoice agent is configured for read-only document ingestion plus predefined parsing and draft-generation tools; payment execution is not available in this zone."}, {"id": "K2", "component_ids": ["C4", "C5"], "flow_ids": ["F6"], "fact": "Payment execution requires a separate authorized approval step before the execution service accepts an instruction."}, {"id": "K3", "component_ids": ["C1", "C4"], "flow_ids": ["F1", "F5"], "fact": "Incoming documents carry provenance checks, and approval reviews the draft against the invoice summary before release, though forged content may still pass imperfect checks."}, {"id": "K4", "component_ids": ["C4", "C5", "C7"], "flow_ids": ["F6", "F7", "F8"], "fact": "Approval and execution actions are logged to the audit store with reviewer identity and payment reference for later traceability."}], "architecture": {"components": [{"id": "C1", "label": "Invoice intake agent", "zone": "Z2", "kind": "agent"}, {"id": "C2", "label": "Document parsing service", "zone": "Z2", "kind": "process"}, {"id": "C3", "label": "Payment draft generator", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Approval service", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment execution service", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Supplier invoice portal", "zone": "Z1", "kind": "external_entity"}, {"id": "C7", "label": "Audit log store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C6", "target": "C1", "data": "invoice document and attached evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C2", "data": "parse request with bounded document fields", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C1", "data": "extracted invoice fields and provenance metadata", "boundary_ids": []}, {"id": "F4", "source": "C1", "target": "C3", "data": "normalized invoice details for draft creation", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C4", "data": "payment draft and supporting invoice summary", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F7", "source": "C4", "target": "C7", "data": "approval decision and reviewer identity", "boundary_ids": []}, {"id": "F8", "source": "C5", "target": "C7", "data": "execution result and payment reference", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Suppliers"}, {"id": "Z2", "label": "Invoice Processing App"}, {"id": "Z3", "label": "Approval & Payments Ops"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction manipulation via invoice content", "component_ids": ["C6", "C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "boundary_ids": ["B1"], "asset": "invoice processing workflow and generated payment draft", "threat_vector": "Attacker-supplied invoice document or attached evidence contains crafted content that manipulates the invoice intake agent or downstream parsing/draft creation behavior.", "preconditions": ["Attacker can submit invoice documents through the supplier portal.", "The attacker can influence document text/attachments that the agent reads.", "Provenance checks may not reliably reject forged or manipulated content."], "impact": "Incorrect extraction or draft generation could steer the agent to create a malicious or erroneous payment draft, affecting integrity of the payment workflow.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "The invoice agent is configured for read-only document ingestion plus predefined parsing and draft-generation tools; payment execution is not available in this zone.", "justification": "This is a supported agent-specific manipulation path: external invoice content crosses B1 into the agent zone and is processed by bounded parsing and draft-generation tools. Controls reduce direct execution but do not eliminate content-driven manipulation, and the architecture explicitly notes imperfect provenance checks."}, {"id": "T2", "category": "T", "subtype": "invoice/document tampering before approval", "component_ids": ["C6", "C1", "C3", "C4"], "flow_ids": ["F1", "F5"], "boundary_ids": ["B1", "B2"], "asset": "payment draft and approval decision", "threat_vector": "A forged or modified invoice document reaches the intake agent and is transformed into a draft that is then reviewed for approval.", "preconditions": ["Attacker can supply or alter the invoice document before ingestion.", "Forged content is plausible enough to pass provenance checks and reach review."], "impact": "Tampered invoice content may cause approval of an incorrect payment draft, leading to unauthorized or mistaken payment processing.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "Incoming documents carry provenance checks, and approval reviews the draft against the invoice summary before release, though forged content may still pass imperfect checks.", "justification": "The architecture acknowledges an end-to-end path from external invoice input to approval review, and also states that forged content may still pass imperfect checks. That supports a tampering candidate on the document-to-draft path, despite the review step."}, {"id": "T3", "category": "R", "subtype": "approval or execution repudiation", "component_ids": ["C4", "C5", "C7"], "flow_ids": ["F6", "F7", "F8"], "boundary_ids": [], "asset": "auditability of approval and payment actions", "threat_vector": "An actor involved in approval or execution disputes having approved or executed a payment, or disputes the content of the recorded action.", "preconditions": ["An approval or execution action occurred.", "Audit records are needed to attribute the action after the fact."], "impact": "If logs are incomplete, altered, or not consulted, the system may be unable to reliably attribute approvals or payment executions to a reviewer identity.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Approval and execution actions are logged to the audit store with reviewer identity and payment reference for later traceability.", "justification": "This is a repudiation candidate on the approval/execution path, but the explicit audit logging control reduces the risk by preserving reviewer identity and payment reference. The remaining weakness is not shown in the supplied facts."}, {"id": "T4", "category": "D", "subtype": "document volume or parsing overload", "component_ids": ["C6", "C1", "C2", "C3"], "flow_ids": ["F1", "F2", "F4"], "boundary_ids": ["B1"], "asset": "invoice intake and draft-generation availability", "threat_vector": "An attacker submits excessively large, numerous, or complex invoice documents to overload ingestion, parsing, or draft creation.", "preconditions": ["Attacker can repeatedly submit invoices through the external portal.", "The parsing/draft pipeline has finite capacity."], "impact": "Processing delays or backlog could block normal invoice intake and draft creation, reducing availability of the invoice workflow.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "The invoice agent is configured for read-only document ingestion plus predefined parsing and draft-generation tools; payment execution is not available in this zone.", "justification": "The architecture exposes a processing path for external documents into bounded parsing and draft-generation functions, but no capacity, rate-limiting, or queueing effectiveness is specified. That supports only an unknown availability assessment, not a confirmed denial-of-service exposure."}], "scope_notes": ["The model shows no direct path from the invoice agent zone to payment execution; execution requires the separate approval path.", "No explicit rate limiting, schema validation depth, or parser hardening is provided for external invoice inputs.", "Audit logging is present, but the integrity and retention properties of the log store are not specified.", "No direct unauthorized read of the audit store is shown; a data-store flow alone does not establish disclosure to an attacker."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Enforce strict content isolation for invoice inputs: sanitize and tokenize inbound documents before agent consumption, reject control-like directives in text/attachments, and keep parsing/draft generation constrained to fixed schemas and allowlisted fields only.", "verification": "Submit benign invoices containing harmless embedded text, metadata, and attachments; confirm the agent only extracts declared invoice fields, ignores non-schema directives, and produces the same draft as a clean equivalent document.", "existing_control_effect": "K1 already limits the zone to read-only ingestion plus predefined parsing and draft tools, and K3 adds provenance checks. These reduce direct tool abuse and block payment execution from this zone, but do not fully stop content-driven manipulation.", "residual_risk": "A crafted invoice may still influence extraction or draft content if the document parser or provenance checks misclassify maliciously formatted content."}, {"threat_id": "T2", "enforcement_component_ids": ["C1", "C3", "C4"], "recommendation": "Require approval-time field-level comparison between the original invoice evidence and the generated draft, with hard stop conditions for mismatched supplier, amount, account, or line-item totals before any approval can proceed.", "verification": "Use a benign test invoice and a second version with only noncritical formatting differences; confirm approval proceeds only when the parsed draft matches the source fields and blocks when key values differ.", "existing_control_effect": "K3 already says approval reviews the draft against the invoice summary, which helps catch some tampering. However, the supplied facts also state forged content may still pass imperfect checks.", "residual_risk": "If forged or altered invoice content remains plausible enough to pass provenance and review, an incorrect payment draft can still reach approval."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5", "C7"], "recommendation": "Bind each approval and execution event to immutable audit records containing reviewer identity, action timestamp, payment reference, and request hash, and require audit write success before the action is considered complete.", "verification": "Perform a normal approval and a normal execution in a test environment; confirm both create audit entries that can be cross-referenced with the payment reference and reviewer identity, and confirm missing audit writes prevent completion.", "existing_control_effect": "K4 already logs approval and execution actions with reviewer identity and payment reference, which reduces repudiation by preserving traceability.", "residual_risk": "Repudiation remains possible if audit log integrity, retention, or access controls are weak, or if an actor disputes involvement despite the recorded trail."}, {"threat_id": "T4", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Add bounded intake controls for the external portal and parsing pipeline: per-supplier rate limits, document size caps, concurrency limits, and queue backpressure with fail-closed behavior on overload.", "verification": "Send a small set of benign large-but-valid invoices under controlled load and confirm the system enforces limits, queues safely, and preserves service for normal-sized documents without resource exhaustion.", "existing_control_effect": "K1 limits the workflow to predefined tools, which constrains processing behavior, but no capacity or throttling control is described.", "residual_risk": "An attacker may still cause temporary backlog or delayed processing if sustained traffic stays within any configured limits or exploits parser complexity."}]}
{"id": "E030", "collection": "comparison", "domain": "finance_operations", "control_profile": "unspecified", "generation_condition": "single_prompt", "trial_id": "slot-12-single_prompt", "topology_family_id": "topology-025", "annotation_status": "candidate", "image_path": "images/comparison/E030.png", "title": "finance_operations", "purpose": "An invoice agent reads externally supplied invoice documents, prepares payment drafts, and hands them to an approval and payment-execution path.", "assumptions": ["Synthetic design; no actual production data.", "The attacker can influence externally supplied invoice documents and attached evidence, but has no privileged access to internal systems.", "Control implementation and effectiveness are unspecified unless explicitly stated in a control fact."], "controls": [{"id": "CTRL1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "The agent and parser perform MIME and schema validation before accepting invoice content."}, {"id": "CTRL2", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "Approval Service requires a human reviewer to confirm the payment draft before release."}, {"id": "CTRL3", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "Draft Ledger stores an append-only record of submitted drafts with timestamps and submitter identifiers."}], "architecture": {"components": [{"id": "C1", "label": "Supplier Document Inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Invoice Parser", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Draft Ledger", "zone": "Z2", "kind": "data_store"}, {"id": "C5", "label": "Approval Service", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "extracted invoice fields and classification hints", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "payment draft records", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "draft invoice payment proposal", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C6", "data": "approved payment instruction", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Invoice Processing"}, {"id": "Z3", "label": "Approval and Payment"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation via invoice content", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "invoice drafting workflow and payment draft integrity", "threat_vector": "Attacker places malicious instructions in an externally supplied invoice document to steer the agent into extracting attacker-chosen values or unsafe actions.", "preconditions": ["Attacker can influence invoice documents or attachments sent to the inbox.", "The agent uses document content as input to downstream extraction or drafting."], "impact": "Incorrect payment draft contents may be produced, including altered payee, amount, or approval text.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "The agent and parser perform MIME and schema validation before accepting invoice content.", "justification": "The path from external documents to the invoice agent and parser is explicit, and the attacker can shape that input. Validation is present, but the quoted control only filters content format; it does not state protection against instruction-style manipulation, so the path remains supported and exposed."}, {"id": "T2", "category": "T", "subtype": "draft tampering before approval", "component_ids": ["C4", "C5"], "flow_ids": ["F4"], "boundary_ids": ["B2"], "asset": "payment draft integrity", "threat_vector": "An attacker who can influence draft content through upstream document manipulation causes modified draft fields to be stored and forwarded for approval.", "preconditions": ["The attacker can induce malformed or deceptive invoice data upstream.", "Draft fields are accepted into the ledger without a stated semantic integrity check."], "impact": "The approval service may review and approve a tampered payment proposal.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "Draft Ledger stores an append-only record of submitted drafts with timestamps and submitter identifiers.", "justification": "The ledger records submissions, which helps auditability, but the control does not specify validation of field correctness or prevention of semantically malicious draft content. Because effectiveness against tampering is unspecified, the path is analyzed as unknown rather than exposed."}, {"id": "T3", "category": "R", "subtype": "payment draft repudiation", "component_ids": ["C4", "C5"], "flow_ids": ["F4", "F5"], "boundary_ids": ["B2"], "asset": "traceability of draft approval and release", "threat_vector": "A participant later disputes whether a specific draft was submitted or approved because the record is not shown to bind the exact content to the submitter and approver identity.", "preconditions": ["A dispute arises over a specific payment draft or approval event.", "The system relies on logs or ledger entries for accountability."], "impact": "Investigations may be unable to attribute a draft or approval decision with confidence.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL3", "CTRL2"], "evidence_quote": "Draft Ledger stores an append-only record of submitted drafts with timestamps and submitter identifiers.", "justification": "The ledger and human approval create explicit accountability signals. The quote shows submitter identifiers are recorded, reducing straightforward repudiation, though it does not fully describe cryptographic binding or immutable approval evidence."}], "scope_notes": ["Boundary IDs follow the listed flow order: B1 covers the first external-to-internal zone crossing on F1; B2 covers the first cross-zone transfer between processing and approval on F4 and is shared by reverse-direction or later flows between those same zones.", "The architecture does not state whether validation blocks prompt injection, whether approvals are authenticated with non-repudiation controls, or whether the ledger contents are semantically checked before approval.", "No unsupported disclosure path is asserted from the Draft Ledger alone because a data-store read does not by itself show an unauthorized recipient or data release."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Constrain the agent and parser to treat invoice documents as data only: separate extracted fields from free text, reject instruction-like markup in invoice bodies, and require a fixed extraction template before any draft is created.", "verification": "Submit a benign invoice containing harmless embedded prose and confirm the system extracts only allowed fields while ignoring non-field instructions; then confirm the output draft matches the template.", "existing_control_effect": "The MIME and schema validation already filters accepted content formats, which lowers malformed-input risk but does not specifically address instruction manipulation.", "residual_risk": "Prompt-style or adversarial text embedded in otherwise valid documents may still influence extraction unless content-separation controls are added and maintained."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Add a semantic consistency check on draft records before approval, comparing supplier, amount, invoice number, and tax fields against allowed business rules and source-document references.", "verification": "Use a benign test invoice with consistent fields and confirm it passes the consistency check; use a mismatched but harmless sample and confirm it is flagged before approval.", "existing_control_effect": "Append-only logging preserves the draft history and helps detect later modification, but it does not prevent an initially malformed draft from being recorded.", "residual_risk": "If upstream extraction is wrong yet structurally valid, the approval path can still process a misleading draft unless semantic checks are enforced."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Bind each draft and approval event to a durable identifier and signer record, and require the approval service to reference the exact ledger entry ID before release.", "verification": "Create a benign draft, approve it, and verify the stored entry ID, submitter ID, approver identity, and timestamp are all retrievable and linked in audit logs.", "existing_control_effect": "The append-only ledger with submitter identifiers and human review already improves traceability and reduces simple repudiation claims.", "residual_risk": "Disputes can still arise if identity assurance, reviewer attribution, or record integrity is weak outside the described controls."}]}
{"id": "E031", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "full_workflow", "trial_id": "slot-03-full_workflow", "topology_family_id": "topology-026", "annotation_status": "candidate", "image_path": "images/comparison/E031.png", "title": "Guarded support-assist workflow with external document retrieval and account lookup", "purpose": "Model a support agent workflow that drafts customer replies from retrieved help content and customer-account data, while enforcing provenance checks, role-based access, and bounded tool use.", "assumptions": ["The support agent operates inside a trusted application zone with access only to approved tools.", "External help documents and customer-submitted evidence may be attacker-influenced and are treated as untrusted until provenance checks pass.", "The account store enforces authorization for read access and returns only fields allowed by the support role.", "Tool use is bounded to retrieval, lookup, and draft-generation actions; no direct write-back to customer accounts is assumed.", "A human or downstream approval step may review the draft before sending, but its effectiveness is not guaranteed by the scenario facts.", "The attacker can influence external documents or evidence but has no privileged access to the internal account store or support zone."], "controls": [{"id": "K1", "component_ids": ["C3", "C6"], "flow_ids": ["F5", "F6"], "fact": "Account lookup is role-authorized and returns only support-approved fields from the internal store."}, {"id": "K2", "component_ids": ["C2", "C5"], "flow_ids": ["F2", "F3", "F7"], "fact": "Retrieved help content is subject to provenance checks before it is forwarded into draft generation."}, {"id": "K3", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F4", "F7", "F8"], "fact": "Tool use is bounded to retrieval, lookup, and draft generation; the agent has no direct account write path."}, {"id": "K4", "component_ids": ["C4", "C7"], "flow_ids": ["F9", "F10"], "fact": "A review step may inspect the draft before sending, but the scenario does not guarantee that review always blocks unsafe content."}], "architecture": {"components": [{"id": "C1", "label": "Support agent", "zone": "Z1", "kind": "agent"}, {"id": "C2", "label": "Retrieval tool", "zone": "Z1", "kind": "process"}, {"id": "C3", "label": "Account lookup service", "zone": "Z1", "kind": "process"}, {"id": "C4", "label": "Draft composer", "zone": "Z1", "kind": "process"}, {"id": "C5", "label": "Help documents", "zone": "Z2", "kind": "external_entity"}, {"id": "C6", "label": "Account records", "zone": "Z3", "kind": "data_store"}, {"id": "C7", "label": "Review step", "zone": "Z4", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "search request and customer-supplied query context", "boundary_ids": []}, {"id": "F2", "source": "C2", "target": "C5", "data": "retrieval request", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C5", "target": "C2", "data": "help document content", "boundary_ids": ["B1"]}, {"id": "F4", "source": "C1", "target": "C3", "data": "account lookup request with ticket/customer identifier", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C6", "data": "authorized read query", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C6", "target": "C3", "data": "role-limited account fields", "boundary_ids": ["B2"]}, {"id": "F7", "source": "C2", "target": "C4", "data": "retrieved help content after provenance checks", "boundary_ids": []}, {"id": "F8", "source": "C3", "target": "C4", "data": "validated account facts", "boundary_ids": []}, {"id": "F9", "source": "C4", "target": "C7", "data": "draft response for review", "boundary_ids": ["B3"]}, {"id": "F10", "source": "C7", "target": "C1", "data": "approval or edits", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "Trusted support application"}, {"id": "Z2", "label": "External help source"}, {"id": "Z3", "label": "Customer account store"}, {"id": "Z4", "label": "Review/send channel"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z1", "Z3"]}, {"id": "B3", "zone_ids": ["Z1", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt-injection via retrieved help content", "component_ids": ["C2", "C4"], "flow_ids": ["F2", "F3", "F7"], "boundary_ids": ["B1"], "asset": "support reply drafting behavior", "threat_vector": "Attacker-controlled external help content is retrieved from Z2 and forwarded into draft generation in Z1, where it can manipulate the agent or composer instructions.", "preconditions": ["Attacker can influence external help documents.", "Retrieval returns document content that is consumed by draft generation."], "impact": "Unsafe or misleading draft content, including instruction/context manipulation that changes the support agent’s behavior.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "Retrieved help content is subject to provenance checks before it is forwarded into draft generation.", "justification": "The architecture explicitly shows untrusted external content crossing B1 into the trusted zone and then reaching draft generation. Provenance checks are present, but the workflow still forwards retrieved content into composition, so this remains a supported exposure for agent-context manipulation."}, {"id": "T2", "category": "I", "subtype": "account-data overexposure through drafting", "component_ids": ["C3", "C4", "C7"], "flow_ids": ["F5", "F6", "F8", "F9", "F10"], "boundary_ids": ["B2", "B3"], "asset": "customer account data", "threat_vector": "Support-approved account fields are read from Z3 into the trusted zone and incorporated into a draft that is sent to a review/send channel; if the draft is over-shared, account details may reach an unintended recipient.", "preconditions": ["The draft includes account facts from the lookup path.", "The review/send step does not remove all sensitive details before release."], "impact": "Disclosure of customer account information in a response beyond the intended support context.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1", "K4"], "evidence_quote": "Account lookup is role-authorized and returns only support-approved fields from the internal store.", "justification": "The lookup path is controlled and limited, which reduces direct store exposure. However, the scenario does not specify what the draft or final response contains, and review effectiveness is not guaranteed, so unauthorized disclosure is possible but not established from the facts."}, {"id": "T3", "category": "A", "subtype": "customer-supplied query context manipulation", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F7"], "boundary_ids": [], "asset": "retrieval and drafting behavior", "threat_vector": "Customer-supplied query context in the support agent’s request can steer retrieval and downstream drafting toward attacker-chosen framing or instructions inside Z1.", "preconditions": ["The attacker can influence the customer-supplied query context.", "The retrieval tool or composer treats that context as actionable input."], "impact": "Misleading or unsafe draft generation caused by manipulated internal query context.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "Tool use is bounded to retrieval, lookup, and draft generation; the agent has no direct account write path.", "justification": "The internal workflow clearly accepts customer-supplied context, but the controls only bound tool use and do not state how such context is sanitized or isolated. Because instruction-handling effectiveness is unspecified, this is a plausible but not proven exposure."}], "scope_notes": ["Boundary IDs assigned from the listed flow order: B1 for Z1<->Z2, B2 for Z1<->Z3, B3 for Z1<->Z4.", "No direct write-back to customer accounts is assumed, which limits tampering and privilege-escalation paths through the account store.", "The review step exists, but its blocking strength is unspecified; do not assume it prevents all unsafe drafts.", "The feedback mentions an unknown or empty control component reference, but the supplied architecture facts list valid control component IDs; this analysis uses the provided architecture only.", "Unsupported or incomplete paths: direct compromise of the account store, bypass of role authorization, and guaranteed final-send blocking are not established by the scenario facts."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4", "C7"], "recommendation": "Enforce strict content-isolation for retrieved help text at C2 before it reaches C4: strip or neutralize instruction-like text, tag provenance, and allow C4 to consume only a constrained evidence schema rather than raw document prose; require C7 to block drafts that quote or execute retrieved instructions.", "verification": "Use benign test documents containing harmless imperative phrasing and check that C2 marks them as untrusted evidence, C4 only receives structured excerpts, and C7 rejects drafts that attempt to treat external text as instructions.", "existing_control_effect": "K2 already adds provenance checks before forwarding help content into draft generation, which reduces but does not eliminate instruction-manipulation risk.", "residual_risk": "Attacker-crafted content may still influence summarization or wording if isolation is incomplete, and review may miss subtle instruction-following behavior."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C4", "C7"], "recommendation": "Minimize account facts at C3 to the least necessary fields and enforce a response-composition policy at C4 that redacts account identifiers, balances, and other sensitive attributes unless explicitly required; have C7 verify the outgoing draft against a support-field allowlist before approval or send.", "verification": "Run benign account-lookup cases with non-sensitive and sensitive fields and confirm C4 omits disallowed details while C7 flags any draft that contains fields outside the allowlist.", "existing_control_effect": "K1 limits lookup to role-authorized support fields, which reduces direct overexposure from the store, and K4 adds a review step before sending.", "residual_risk": "If the allowed field set is still too broad or review is bypassed or inattentive, a drafted reply can still disclose more than intended."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2", "C4"], "recommendation": "Treat customer-supplied query context from C1 as untrusted metadata: separate it from operational instructions at C2, apply a fixed retrieval template, and require C4 to ignore any instruction-like text originating from user context unless explicitly revalidated by a support-approved rule set.", "verification": "Use benign customer messages containing arbitrary but non-malicious instructions and confirm the retrieval request remains template-driven and the composer does not alter behavior based on those instructions.", "existing_control_effect": "K3 bounds tool use to retrieval, lookup, and draft generation, which limits the blast radius of manipulated context but does not by itself sanitize or sandbox it.", "residual_risk": "Context steering may still affect search terms or draft framing if the boundary between user-provided text and internal instructions is not enforced consistently."}]}
{"id": "E032", "collection": "comparison", "domain": "finance_operations", "control_profile": "weak", "generation_condition": "no_reviewer", "trial_id": "slot-07-no_reviewer", "topology_family_id": "topology-027", "annotation_status": "candidate", "image_path": "images/comparison/E032.png", "title": "Invoice-to-Payment Assistant with Weak Review Gates", "purpose": "Ingest externally supplied invoices, extract payment details, draft payment requests, and pass them through approval into payment execution under weak content and tool controls.", "assumptions": ["External invoice files and supporting evidence are attacker-influenceable but arrive through ordinary intake channels, not privileged system access.", "The invoice processor can read document contents and may follow embedded instructions or malformed prompts because document sanitization and instruction separation are weak.", "Approval is human-in-the-loop but uses permissive review cues from the draft and supporting evidence; payment execution consumes approved drafts through a connected internal path.", "At least one payment/tooling action is reachable from the draft handoff, but downstream validation strength is unspecified or inconsistent."], "controls": [{"id": "CTRL1", "component_ids": ["C1", "C3"], "flow_ids": ["F1"], "fact": "Inbound invoice files are accepted through a normal intake channel with only limited sanitization before parsing."}, {"id": "CTRL2", "component_ids": ["C2", "C3"], "flow_ids": ["F3"], "fact": "The evidence store is readable by the invoice processor, but instruction separation and attachment filtering are weak or inconsistently enforced."}, {"id": "CTRL3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "The draft presented to the approver is formatted as a review aid, but the review cues are permissive and may mirror agent-extracted fields with limited cross-checking."}, {"id": "CTRL4", "component_ids": ["C5", "C6"], "flow_ids": ["F6"], "fact": "Payment execution requires an approval action, yet the exact downstream validation on approved drafts is unspecified and may rely on the prior review."}], "architecture": {"components": [{"id": "C1", "label": "Invoice Inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Evidence Store", "zone": "Z1", "kind": "data_store"}, {"id": "C3", "label": "Invoice Processor", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Draft Builder", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Approver", "zone": "Z3", "kind": "agent"}, {"id": "C6", "label": "Payment Executor", "zone": "Z3", "kind": "process"}, {"id": "C7", "label": "Ledger", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C3", "data": "invoice files, attachments, embedded text", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C2", "data": "stored invoice evidence and metadata", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C3", "data": "retrieved evidence for extraction and review", "boundary_ids": ["B1"]}, {"id": "F4", "source": "C3", "target": "C4", "data": "parsed payee, amount, due date, notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "payment draft and supporting summary", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C5", "target": "C6", "data": "approved payment release", "boundary_ids": []}, {"id": "F7", "source": "C6", "target": "C7", "data": "posted payment record", "boundary_ids": []}, {"id": "F8", "source": "C6", "target": "C2", "data": "execution receipt and status update", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External Intake"}, {"id": "Z2", "label": "Invoice Agent"}, {"id": "Z3", "label": "Approval & Payment"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z1", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction/content manipulation via invoice payload", "component_ids": ["C1", "C3", "C4"], "flow_ids": ["F1", "F4"], "boundary_ids": ["B1"], "asset": "invoice processing and draft creation", "threat_vector": "Attacker supplies crafted invoice text/attachments that the invoice processor may follow as embedded instructions during parsing and extraction.", "preconditions": ["Attacker can submit invoice files through the ordinary intake channel.", "Processor behavior can be influenced by document content because sanitization and instruction separation are weak."], "impact": "Could alter extracted payee, amount, due date, or notes and propagate a manipulated payment draft.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "Inbound invoice files are accepted through a normal intake channel with only limited sanitization before parsing.", "justification": "The architecture explicitly allows attacker-influenceable invoices into the processor across B1, and the control fact shows only limited sanitization before parsing. That supports a feasible content/instruction manipulation path into draft creation."}, {"id": "T2", "category": "I", "subtype": "evidence leakage across intake and processing", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F2", "F3"], "boundary_ids": ["B1"], "asset": "invoice evidence and metadata", "threat_vector": "An attacker-controlled invoice may place sensitive or misleading material into the evidence store and have it retrieved by the processor for review/extraction.", "preconditions": ["Attacker can influence the contents of invoices and attachments.", "The processor reads evidence from the store as part of extraction/review."], "impact": "May expose invoice-related sensitive data to the processing agent or retain attacker-supplied material in evidence artifacts for downstream use.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "The evidence store is readable by the invoice processor, but instruction separation and attachment filtering are weak or inconsistently enforced.", "justification": "There is a direct read path from the evidence store to the processor, but the quoted control indicates the path is constrained by weak separation/filtering rather than wholly unguarded. That reduces but does not eliminate disclosure/misuse risk."}, {"id": "T3", "category": "A", "subtype": "review-cue manipulation in draft handoff", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": ["B2"], "asset": "approval decision quality", "threat_vector": "A manipulated draft can exploit permissive review cues so the human approver relies on agent-extracted fields with limited cross-checking.", "preconditions": ["Processor or draft builder has already produced a misleading draft.", "The approver relies on the draft’s review cues and supporting summary."], "impact": "Could cause approval of a fraudulent or altered payment request, leading to incorrect downstream execution.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "The draft presented to the approver is formatted as a review aid, but the review cues are permissive and may mirror agent-extracted fields with limited cross-checking.", "justification": "The quoted control directly supports a weak review-gate path from draft builder to approver across B2. Because the cues may mirror extracted fields with limited cross-checking, the manipulation path remains feasible."}, {"id": "T4", "category": "T", "subtype": "approved draft relies on unspecified downstream validation", "component_ids": ["C5", "C6", "C7"], "flow_ids": ["F6", "F7"], "boundary_ids": [], "asset": "payment execution and ledger integrity", "threat_vector": "A compromised or misled approval action can pass a bad payment request into execution where downstream validation is unspecified or inconsistent.", "preconditions": ["An approval action is obtained for a manipulated draft.", "Payment execution accepts the approval with validation that may rely on prior review."], "impact": "Could result in an incorrect payment record being posted to the ledger.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL4"], "evidence_quote": "Payment execution requires an approval action, yet the exact downstream validation on approved drafts is unspecified and may rely on the prior review.", "justification": "Approval is required, which limits direct unauthorized execution, but the quoted control leaves the validation strength unspecified. That makes the tampering path plausible but not well enough specified to mark exposed."}], "scope_notes": ["Boundary IDs were assigned from the listed flows: B1 for Z1-Z2, B2 for Z2-Z3, and B3 for Z1-Z3; no separate threat was added for F8 because the control fact does not specify a concrete exploit path from the execution receipt back into payment control.", "The architecture mentions attacker-influenceable external invoice files and weak instruction separation; this supports agent-manipulation threats, but not a separate privilege-escalation claim without a clearer tool-authentication or authorization bypass path.", "A read path alone from the evidence store does not prove disclosure to an unauthorized external recipient; the evidence-related finding is limited to feasible misuse/exposure within the processing path.", "Downstream validation on approved drafts and the effect of execution receipts on the evidence store are incomplete in scope, so unsupported paths were not upgraded beyond the specified assessments."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C3", "C4"], "recommendation": "Harden intake and parsing so invoice files are treated as untrusted data only: strip or quarantine active content, separate instructions from document text, and require the processor to ignore any embedded directives before extraction or draft generation.", "verification": "Submit benign test invoices containing harmless embedded directives and malformed formatting; confirm the processor logs them as data only, extracts fields without following directives, and produces unchanged drafts when the same invoice is replayed.", "existing_control_effect": "Current intake allows the normal channel with limited sanitization, so the path is reachable but partially filtered rather than blocked.", "residual_risk": "Well-crafted content may still influence extraction if sanitization or instruction separation regresses or misses a document type."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Restrict evidence retrieval to minimum necessary fields, apply attachment/content filtering before the processor reads from the evidence store, and log each retrieval with a stable record of what was exposed to the processor.", "verification": "Use a harmless test record with mixed metadata and attachment markers, then confirm the processor receives only the permitted fields and filtered content, and that retrieval logs show the exact subset returned.", "existing_control_effect": "The evidence store is readable by the processor, but instruction separation and attachment filtering are weak or inconsistently enforced.", "residual_risk": "Some non-sensitive metadata may still be available for extraction workflows, and filtering gaps can reintroduce misuse or leakage inside the processing path."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Make approval review cues independent from agent-extracted fields by requiring side-by-side source evidence, explicit discrepancy highlights, and a mandatory second check for payee, amount, and due date before approval.", "verification": "Run a benign invoice through the draft and approval flow, then confirm the approver sees source-backed comparisons, discrepancy flags, and cannot approve without reviewing the referenced evidence fields.", "existing_control_effect": "The draft is already presented as a review aid, but the cues are permissive and may mirror agent-extracted fields with limited cross-checking.", "residual_risk": "A vigilant approver can still miss a subtle manipulation if the source evidence itself is misleading or if review is rushed."}, {"threat_id": "T4", "enforcement_component_ids": ["C5", "C6", "C7"], "recommendation": "Add an explicit downstream validation gate in payment execution that rechecks approval provenance, draft integrity, recipient identity, amount limits, and ledger-entry consistency before any posting occurs.", "verification": "Use a benign approved draft in a test environment and confirm execution only proceeds after the independent validation checks pass, with failures blocked and recorded when fields are altered.", "existing_control_effect": "Payment execution requires an approval action, yet the exact downstream validation on approved drafts is unspecified and may rely on the prior review.", "residual_risk": "If approval is compromised, some incorrect requests may still reach execution unless the new validation is independent of the review step and resistant to forged approval context."}]}
{"id": "E033", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "single_prompt", "trial_id": "slot-04-single_prompt", "topology_family_id": "topology-028", "annotation_status": "candidate", "image_path": "images/comparison/E033.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "Attacker can influence externally supplied help documents but has no privileged access to the support workspace or customer store.", "Response drafting uses bounded retrieval and policy review before a human sends the final reply."], "controls": [{"id": "K1", "component_ids": ["C2"], "flow_ids": ["F1", "F2"], "fact": "The ingestion step strips active content and preserves source metadata before forwarding excerpts."}, {"id": "K2", "component_ids": ["C3"], "flow_ids": ["F3", "F4", "F7"], "fact": "The drafting tool only accepts account lookups for the current case and forwards drafts to policy review before display."}, {"id": "K3", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "The customer store exposes read-only, case-scoped account fields through an authenticated service account."}, {"id": "K4", "component_ids": ["C6"], "flow_ids": ["F7", "F8"], "fact": "Policy review flags drafts that request unsupported disclosures or non-workflow actions."}], "architecture": {"components": [{"id": "C1", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C2", "label": "Doc Ingest Service", "zone": "Z2", "kind": "process"}, {"id": "C3", "label": "Response Drafting Tool", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Help Document Repository", "zone": "Z1", "kind": "external_entity"}, {"id": "C5", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Policy Check Service", "zone": "Z2", "kind": "process"}], "flows": [{"id": "F1", "source": "C4", "target": "C2", "data": "help documents and referenced snippets", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "sanitized excerpts and metadata", "boundary_ids": []}, {"id": "F3", "source": "C1", "target": "C3", "data": "response prompts and case notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C5", "data": "account lookup request", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C3", "data": "account details and status fields", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C3", "target": "C1", "data": "draft response and citations", "boundary_ids": []}, {"id": "F7", "source": "C3", "target": "C6", "data": "draft for policy review", "boundary_ids": []}, {"id": "F8", "source": "C6", "target": "C1", "data": "approval or rejection guidance", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support Workspace"}, {"id": "Z3", "label": "Customer Data Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection via help docs", "component_ids": ["C4", "C2", "C3", "C1"], "flow_ids": ["F1", "F2", "F3", "F6"], "boundary_ids": ["B1"], "asset": "support response quality and agent tool authority", "threat_vector": "Attacker-controlled help documents can embed instructions that survive ingestion and steer the drafting flow.", "preconditions": ["Attacker can influence external documents.", "The agent follows retrieved text as guidance during drafting."], "impact": "Misleading or unsafe responses, including unauthorized operational steps or incorrect customer guidance.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "The ingestion step strips active content and preserves source metadata before forwarding excerpts.", "justification": "The path exists across the external-source to workspace boundary, but the design explicitly strips active content and routes drafts through policy review, reducing direct instruction-following risk."}, {"id": "T2", "category": "I", "subtype": "account-data exposure in response drafting", "component_ids": ["C3", "C5", "C1"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B2"], "asset": "customer account details", "threat_vector": "A support draft may surface more account data than needed if the lookup is too broad or the human uses the draft verbatim.", "preconditions": ["A case is opened for a real customer record.", "The drafting workflow requests account fields relevant to the case."], "impact": "Sensitive account information could be disclosed to the wrong recipient in a customer response.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3", "K4"], "evidence_quote": "The customer store exposes read-only, case-scoped account fields through an authenticated service account.", "justification": "The lookup path crosses into the customer-data zone, but access is case-scoped and the draft is checked for unsupported disclosures, so the exposure is constrained rather than open-ended."}, {"id": "T3", "category": "R", "subtype": "draft provenance dispute", "component_ids": ["C1", "C2", "C3", "C6"], "flow_ids": ["F1", "F2", "F6", "F7", "F8"], "boundary_ids": ["B1"], "asset": "response provenance and auditability", "threat_vector": "If a customer disputes advice, the workflow may not fully prove which sourced document fragment or policy decision drove the final draft.", "preconditions": ["The support agent sends a response based on generated draft text.", "Audit records are incomplete or not consulted after the fact."], "impact": "Difficulty attributing the response to a specific source, reviewer, or policy decision.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1", "K4"], "evidence_quote": "Policy review flags drafts that request unsupported disclosures or non-workflow actions.", "justification": "There is a review step, but the facts do not specify durable audit logging or immutable provenance records, so repudiation resistance is not fully determined."}], "scope_notes": ["Boundary B1 covers traffic between External Sources (Z1) and Support Workspace (Z2) via F1/F2; B2 covers traffic between Support Workspace (Z2) and Customer Data Zone (Z3) via F4/F5.", "Unsupported paths remain possible if the human agent copies external text outside the drafting workflow; that behavior is outside the explicit control facts.", "The architecture does not specify retention, logging, or ticketing-system integration, so audit completeness cannot be assumed.", "No direct write path to the customer store is shown; disclosure analysis is limited to read-only lookup and drafted output."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C6"], "recommendation": "Keep the ingestion sanitizer mandatory for all external help documents, require the drafting tool to ignore instruction-like text from retrieved excerpts, and enforce policy review before any response is shown to the agent.", "verification": "Use a harmless test document containing fake directives and confirm C2 strips active content, C3 treats the text as citation material only, and C6 blocks any draft that attempts unsupported actions.", "existing_control_effect": "The ingestion step strips active content and preserves source metadata before forwarding excerpts; drafts are already routed to policy review.", "residual_risk": "A persuasive but non-executable instruction may still influence phrasing or prioritization before review."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C5", "C6"], "recommendation": "Limit C3 to a minimal case-scoped field set, keep C5 read-only for support lookups, and require C6 to reject drafts that include account data outside the case need.", "verification": "Run a benign lookup for a sample case and confirm only the approved fields are returned and that an overbroad draft is rejected by policy review.", "existing_control_effect": "The customer store exposes read-only, case-scoped account fields through an authenticated service account.", "residual_risk": "A valid case may still justify some sensitive fields, so human handling remains a disclosure risk if the draft is copied outside process."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C6"], "recommendation": "Add durable provenance capture at ingestion, draft generation, and policy review so each response records source snippets, lookup scope, and reviewer outcome before send.", "verification": "For a routine case, confirm the saved record links the selected help excerpt, the account lookup scope, and the review decision without altering the response content.", "existing_control_effect": "Policy review flags drafts that request unsupported disclosures or non-workflow actions.", "residual_risk": "Even with provenance notes, disputes may persist if supporting logs are incomplete or if the final human message deviates from the recorded draft."}]}
{"id": "E034", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-01-full_workflow", "topology_family_id": "topology-029", "annotation_status": "candidate", "image_path": "images/comparison/E034.png", "title": "Support Agent with External Docs and Account Lookup", "purpose": "A customer-support assistant drafts replies using retrieved help articles and customer-account data, while accepting externally supplied documents or evidence from customers and public sources.", "assumptions": ["The support agent operates with weakly enforced prompt/tool controls and can be influenced by retrieved external text.", "The customer-account store is read-only for the assistant, but its access policy is permissive enough to return customer records for lookup.", "Externally supplied documents and evidence are attacker-influenced inputs outside the trusted application zone unless validated by the assistant workflow.", "The architecture is a single connected support workflow with no additional parallel subsystems."], "controls": [{"id": "K1", "component_ids": ["C2", "C4"], "flow_ids": ["F3", "F4"], "fact": "Help retrieval is allowed from a preconfigured article source, but query filtering is weak and retrieved text is passed back into the agent context with limited instruction sanitization."}, {"id": "K2", "component_ids": ["C2", "C6"], "flow_ids": ["F5", "F6"], "fact": "The account lookup API is read-only for the assistant and returns only customer records needed for case handling; access is permissive for lookup but does not allow write actions."}, {"id": "K3", "component_ids": ["C5", "C2"], "flow_ids": ["F2"], "fact": "Uploaded evidence is accepted into the workflow as untrusted external input, with validation left to the agent workflow rather than enforced at the boundary."}, {"id": "K4", "component_ids": ["C2", "C3"], "flow_ids": ["F7", "F8"], "fact": "Drafting is performed inside the support workspace, but the response generator applies only basic policy checks and does not strongly separate external evidence from internal instructions."}], "architecture": {"components": [{"id": "C1", "label": "Customer Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Drafting Workflow", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Help Article Retriever", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Uploaded Evidence", "zone": "Z3", "kind": "external_entity"}, {"id": "C6", "label": "Account Lookup API", "zone": "Z4", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "customer message and issue details", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C5", "target": "C2", "data": "external documents and evidence", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "search query for help articles", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C2", "data": "retrieved help content", "boundary_ids": []}, {"id": "F5", "source": "C2", "target": "C6", "data": "account lookup request", "boundary_ids": ["B3"]}, {"id": "F6", "source": "C6", "target": "C2", "data": "customer account record", "boundary_ids": ["B3"]}, {"id": "F7", "source": "C2", "target": "C3", "data": "case context, retrieved text, and account data", "boundary_ids": []}, {"id": "F8", "source": "C3", "target": "C1", "data": "draft support reply", "boundary_ids": ["B1"]}], "trust_zones": [{"id": "Z1", "label": "Customer Channel"}, {"id": "Z2", "label": "Support Workspace"}, {"id": "Z3", "label": "External Docs"}, {"id": "Z4", "label": "Account Store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction/context manipulation via external docs into drafting", "component_ids": ["C5", "C2", "C3"], "flow_ids": ["F2", "F7", "F8"], "boundary_ids": ["B2", "B1"], "asset": "support reply drafting workflow", "threat_vector": "Attacker-supplied evidence or documents are ingested by the support agent and then forwarded into draft generation, where weak separation can let external text steer the agent or override intended instructions.", "preconditions": ["Attacker can submit or influence uploaded evidence or public-source documents.", "Agent workflow accepts the content as part of case handling and passes it into drafting context."], "impact": "Malformed or instruction-bearing evidence can manipulate the agent's response content, causing unsafe tool-use decisions or incorrect support replies.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Uploaded evidence is accepted into the workflow as untrusted external input, with validation left to the agent workflow rather than enforced at the boundary.", "justification": "The design explicitly admits attacker-influenced external input and states that the response generator does not strongly separate external evidence from internal instructions, supporting a feasible agent-context manipulation path."}, {"id": "T2", "category": "I", "subtype": "account data overexposure through support lookup", "component_ids": ["C2", "C6"], "flow_ids": ["F5", "F6"], "boundary_ids": ["B3"], "asset": "customer account record", "threat_vector": "A support agent can request account records over a permissive read-only lookup API, potentially exposing more account detail to the workspace than needed for the case.", "preconditions": ["The agent is allowed to issue lookup requests for a customer case.", "The returned record contains sensitive fields beyond the minimum needed for handling."], "impact": "Sensitive customer account information may be disclosed within the support workspace and used in downstream drafting or seen by inappropriate operators.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "The account lookup API is read-only for the assistant and returns only customer records needed for case handling; access is permissive for lookup but does not allow write actions.", "justification": "There is a clear cross-boundary account lookup path, but the control says the API returns only needed records, so overexposure is not established; the exact data minimization effectiveness is unspecified."}], "scope_notes": ["No threat is claimed for help-article retrieval alone because the provided control describes weak filtering but still confines retrieval to a preconfigured article source; impact depends on how retrieved text is used.", "The architecture does not specify whether account records are filtered by field or redacted before reaching the agent, so disclosure severity cannot be raised beyond unknown.", "No unsupported boundary IDs were added; flows within the support workspace have no boundary traversal and were not treated as separate external attacks."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Enforce boundary-level content separation in the support workflow: treat uploaded evidence as data-only, strip or quarantine instruction-like text before it reaches drafting, and require the drafting component to ignore external directives embedded in case artifacts.", "verification": "Submit benign sample evidence containing harmless pseudo-instructions and confirm the workflow tags it as untrusted, excludes it from the drafting instruction channel, and still produces a normal draft from case facts alone.", "existing_control_effect": "K3 already marks uploaded evidence as untrusted, and K4 keeps drafting inside the support workspace with basic policy checks; this reduces risk but does not reliably prevent external text from influencing the agent context.", "residual_risk": "A determined attacker may still shape the content of the case narrative or exploit weak policy checks if sanitization is incomplete or bypassed."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C6"], "recommendation": "Apply strict field-level minimization on account lookups: allow only case-specific attributes to return to the support workspace, redact unnecessary sensitive fields at the API response, and log each lookup against an explicit support-case purpose.", "verification": "Use a test account and request a lookup containing optional or sensitive fields; confirm the API returns only the approved minimal field set and redacts everything else while preserving normal case handling.", "existing_control_effect": "K2 already makes the lookup read-only and intended for customer records needed for handling, which lowers tampering risk and limits write impact, but it does not by itself prove strong field minimization.", "residual_risk": "If the permitted field set is broader than intended or the case context is misclassified, some sensitive account data may still be exposed inside the workspace."}]}
{"id": "E035", "collection": "comparison", "domain": "customer_support", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-03-no_reviewer", "topology_family_id": "topology-030", "annotation_status": "candidate", "image_path": "images/comparison/E035.png", "title": "Guarded Support Agent with Provenanced Knowledge Retrieval and Scoped Account Lookup", "purpose": "Customer-support workflow where an AI agent drafts replies using externally supplied help documents and a customer-account store under explicit authorization, provenance checks, and bounded read-only tool use.", "assumptions": ["The support agent can retrieve from a help-document repository and a customer-account store under explicit service permissions.", "External help documents, attachments, or evidence can be attacker-influenced but are not trusted by default.", "The customer-account store enforces authenticated access and record-level authorization for the support workflow.", "Tool use is bounded to read-only retrieval and draft generation; no direct write-back to customer accounts is assumed.", "Provenance metadata is available for retrieved documents, but its completeness or strength may vary by source."], "controls": [{"id": "K1", "component_ids": ["C2", "C4", "C6"], "flow_ids": ["F2", "F3", "F6"], "fact": "The orchestrator only issues read-only retrieval calls and scoped account lookups; write operations are not exposed in this workflow."}, {"id": "K2", "component_ids": ["C2", "C6"], "flow_ids": ["F6", "F7"], "fact": "Account lookup is authenticated and limited to case-relevant records by record-level authorization on the customer-account store."}, {"id": "K3", "component_ids": ["C4", "C5", "C2"], "flow_ids": ["F3", "F4", "F8"], "fact": "Retrieved help documents carry provenance metadata and the support app checks source metadata before forwarding text to the drafting agent."}, {"id": "K4", "component_ids": ["C7", "C2", "C3"], "flow_ids": ["F5", "F8"], "fact": "Externally supplied attachments are treated as untrusted input and are sanitized before being included in the drafting context, though content authenticity may remain uncertain."}], "architecture": {"components": [{"id": "C1", "label": "Customer Portal", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support Orchestrator", "zone": "Z2", "kind": "process"}, {"id": "C3", "label": "Drafting Agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Document Retriever", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Help Document Repository", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}, {"id": "C7", "label": "Uploaded Evidence Inbox", "zone": "Z4", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "support request, customer details, case context", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C4", "data": "bounded retrieval query, case topic, source constraints", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C4", "target": "C5", "data": "help-document lookup request", "boundary_ids": []}, {"id": "F4", "source": "C5", "target": "C4", "data": "help article text, provenance metadata", "boundary_ids": []}, {"id": "F5", "source": "C7", "target": "C2", "data": "attachments, screenshots, evidence files", "boundary_ids": ["B3"]}, {"id": "F6", "source": "C2", "target": "C6", "data": "authenticated account lookup request, record scope", "boundary_ids": ["B2"]}, {"id": "F7", "source": "C6", "target": "C2", "data": "authorized account facts, case-relevant record data", "boundary_ids": ["B2"]}, {"id": "F8", "source": "C2", "target": "C3", "data": "sanitized context, retrieved help text, authorized account facts", "boundary_ids": []}, {"id": "F9", "source": "C3", "target": "C2", "data": "draft response, cited source references", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "Customer Input Zone"}, {"id": "Z2", "label": "Support App Zone"}, {"id": "Z3", "label": "Trusted Data Services Zone"}, {"id": "Z4", "label": "Untrusted Document Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt-injection via untrusted evidence", "component_ids": ["C7", "C2", "C3"], "flow_ids": ["F5", "F8", "F9"], "boundary_ids": ["B3"], "asset": "support drafting context and response quality", "threat_vector": "Attacker-controlled attachments or screenshots in the Untrusted Document Zone are uploaded to the support app and then forwarded into the drafting context.", "preconditions": ["Attacker can influence content in uploaded evidence files.", "The draft generator may process sanitized content that still contains adversarial instructions or misleading context."], "impact": "Manipulation of the agent’s instructions/context can lead to unsafe or incorrect drafted replies, including misuse of customer-support workflow.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Externally supplied attachments are treated as untrusted input and are sanitized before being included in the drafting context, though content authenticity may remain uncertain.", "justification": "The architecture explicitly moves untrusted external evidence into the support app and drafting context, and the control only sanitizes content while stating authenticity may remain uncertain. That supports a feasible agent-context manipulation path despite sanitization."}, {"id": "T2", "category": "I", "subtype": "sensitive account-data exposure to draft context", "component_ids": ["C2", "C6", "C3"], "flow_ids": ["F6", "F7", "F8", "F9"], "boundary_ids": ["B2"], "asset": "customer account facts", "threat_vector": "The support app retrieves authorized account facts from the customer-account store and passes them into the drafting agent.", "preconditions": ["The drafted response or cited references may expose more account detail than intended to the customer or other recipients.", "Authorization scope may still allow retrieval of data broader than the minimal reply need."], "impact": "Sensitive customer record details could be disclosed in generated responses or citations if the drafting step includes unnecessary account facts.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K1"], "evidence_quote": "Account lookup is authenticated and limited to case-relevant records by record-level authorization on the customer-account store.", "justification": "There is an explicit account-data path from the trusted store into the drafting agent. Record-level authorization reduces unauthorized access, but the architecture does not specify how well the drafting step prevents oversharing in the final response, so disclosure risk is possible but not demonstrated as exposed."}, {"id": "T3", "category": "I", "subtype": "help-content provenance weakness", "component_ids": ["C4", "C5", "C2", "C3"], "flow_ids": ["F2", "F3", "F4", "F8", "F9"], "boundary_ids": ["B2"], "asset": "help-document text and provenance metadata", "threat_vector": "The document retriever depends on provenance metadata before forwarding help text, but provenance completeness or strength may vary by source.", "preconditions": ["An attacker can influence repository content or provenance quality in the help-document repository.", "A weak or incomplete provenance record may be accepted as sufficient by the metadata check."], "impact": "Misleading or untrusted help content could be incorporated into the draft, causing incorrect or manipulated support responses.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "Retrieved help documents carry provenance metadata and the support app checks source metadata before forwarding text to the drafting agent.", "justification": "The flow from the trusted repository into the drafting agent is explicit, and provenance is checked. However, the supplied assumptions also say provenance completeness or strength may vary, so control effectiveness against malicious or low-integrity source content is unspecified."}, {"id": "T4", "category": "D", "subtype": "support workflow overload via large untrusted uploads", "component_ids": ["C7", "C2", "C3"], "flow_ids": ["F5", "F8", "F9"], "boundary_ids": ["B3"], "asset": "support agent availability", "threat_vector": "Large or numerous untrusted attachments are submitted through the Uploaded Evidence Inbox and passed through sanitization and drafting.", "preconditions": ["An attacker can submit high-volume or oversized evidence files.", "Sanitization and drafting consume shared resources without explicit rate limits in the supplied architecture."], "impact": "Processing overhead could delay or block support-case handling and response generation.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Externally supplied attachments are treated as untrusted input and are sanitized before being included in the drafting context, though content authenticity may remain uncertain.", "justification": "The workflow accepts external evidence across the boundary and processes it before drafting, which creates a plausible availability pressure point. The architecture does not specify size limits, throttling, or queue isolation, so denial impact is possible but not established as exposed."}], "scope_notes": ["Boundary IDs were assigned from the listed flows: B1 for Z1<->Z2, B2 for Z2<->Z3, and B3 for Z2<->Z4.", "No direct write-back to customer accounts is in scope; the design states read-only retrieval and draft generation only.", "The architecture does not specify whether the draft output is automatically sent to customers or reviewed by a human before release.", "Control effectiveness for provenance metadata quality, sanitization depth, and any rate limiting is not specified beyond the quoted facts.", "No separate unsupported path for privilege escalation is evident from the supplied flows and controls."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Enforce strict evidence-to-context filtering at the Support Orchestrator before passing F8 to the Drafting Agent: strip or neutralize instructions, keep only case-relevant factual excerpts, and require the agent to treat uploaded evidence as untrusted reference material rather than instructions.", "verification": "Use benign test attachments that contain obvious instruction-like text and confirm the draft pipeline preserves only allowed factual fields, logs the rejected segments, and produces the same behavior when the evidence is resent with harmless content.", "existing_control_effect": "K4 already sanitizes uploaded evidence before inclusion in the drafting context, which reduces direct injection content but does not by itself guarantee instruction removal or authenticity.", "residual_risk": "A crafted attachment may still contain subtle or obfuscated manipulations that survive sanitization or influence the draft through retained context or misleading facts."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Minimize account data forwarded on F8 by projecting only the specific fields needed for the current case and suppressing identifiers, historical notes, and unrelated account attributes from the drafting context and cited references.", "verification": "Run a benign case with extra nonessential account fields available in C6 and confirm the orchestrator passes only the minimal field set to the Drafting Agent and omits restricted or unused data from the generated draft.", "existing_control_effect": "K2 and K1 ensure authenticated, record-scoped read access and read-only tool use, which limits unauthorized retrieval and prevents direct write-back, but they do not alone prevent oversharing in the draft.", "residual_risk": "Even with field minimization, the drafted reply may still reveal sensitive details that are legitimately case-relevant or become inferable from cited facts."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C2"], "recommendation": "Require provenance policy enforcement on F3/F4/F8: accept help-text only when provenance metadata meets a defined trust threshold, reject ambiguous or incomplete source metadata, and keep a fallback allowlist of approved repository sources for the retriever.", "verification": "Populate the repository with benign documents carrying complete, incomplete, and mismatched provenance metadata and confirm the retriever forwards only the documents that meet the configured provenance policy while logging rejections.", "existing_control_effect": "K3 already checks source metadata before forwarding help text, which reduces the chance of forwarding unvetted content, but the supplied design says provenance strength may vary and does not define a strict acceptance threshold.", "residual_risk": "Trusted-looking but incorrect repository content can still be forwarded if it satisfies the metadata check, so incorrect or manipulated help text may still influence the draft."}, {"threat_id": "T4", "enforcement_component_ids": ["C2"], "recommendation": "Add bounded intake controls on F5 and the sanitization queue: enforce per-upload size limits, attachment-count limits, and rate limiting before evidence enters the drafting path, with overflow rejection instead of deferred processing.", "verification": "Submit benign oversized and high-count test attachments and confirm the intake layer rejects or throttles them before sanitization or drafting, while normal-sized evidence continues through the workflow.", "existing_control_effect": "K4 already sanitizes external attachments before drafting, which can reduce malformed-input cost, but it does not specify any capacity controls or queue isolation.", "residual_risk": "A sustained burst of moderate-sized uploads may still consume resources or increase latency if the upstream intake limits are too permissive or shared infrastructure remains constrained."}]}
{"id": "E036", "collection": "comparison", "domain": "finance_operations", "control_profile": "weak", "generation_condition": "single_prompt", "trial_id": "slot-07-single_prompt", "topology_family_id": "topology-031", "annotation_status": "candidate", "image_path": "images/comparison/E036.png", "title": "finance_operations", "purpose": "An invoice agent reads externally supplied invoice documents, prepares payment drafts, and hands them to an approval and payment-execution path.", "assumptions": ["Synthetic design; no actual production data.", "The external sender can influence invoice content and attached evidence but has no privileged internal access.", "The approver portal is intended to require human review, but the strength of content validation is unspecified beyond basic workflow handling."], "controls": [{"id": "CTRL1", "component_ids": ["C2"], "flow_ids": ["F1", "F3"], "fact": "The invoice agent uses a permissive prompt template that instructs it to extract invoice fields and continue when fields are missing if the document appears otherwise readable."}, {"id": "CTRL2", "component_ids": ["C3"], "flow_ids": ["F2"], "fact": "The document store keeps a copy of ingested invoice attachments for later review."}, {"id": "CTRL3", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "The approver portal requires a logged-in reviewer to click approve before payment instructions are forwarded."}, {"id": "CTRL4", "component_ids": ["C6"], "flow_ids": ["F5"], "fact": "The payment executor accepts only instructions that match the approved draft reference."}], "architecture": {"components": [{"id": "C1", "label": "Vendor Invoice Source", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Document Store", "zone": "Z2", "kind": "data_store"}, {"id": "C4", "label": "Payment Draft Service", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Approver Portal", "zone": "Z3", "kind": "process"}, {"id": "C6", "label": "Payment Executor", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice documents and attached evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "parsed invoice records and stored evidence", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C4", "data": "payment draft details", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "draft payment request for review", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C6", "data": "approved payment instruction", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Party Zone"}, {"id": "Z2", "label": "Invoice Processing Zone"}, {"id": "Z3", "label": "Approval and Payment Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation through invoice content", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "payment draft accuracy and downstream payment handling", "threat_vector": "Attacker-controlled invoice text or attached evidence contains instructions that steer the invoice agent into extracting wrong values or generating a fraudulent payment draft.", "preconditions": ["Attacker can supply or modify invoice documents or evidence.", "The agent follows document text as operational input during extraction or drafting."], "impact": "Incorrect draft amounts, payees, or references can propagate to approval and payment execution.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1"], "evidence_quote": "The invoice agent uses a permissive prompt template that instructs it to extract invoice fields and continue when fields are missing if the document appears otherwise readable.", "justification": "The architecture explicitly lets externally supplied document content drive the agent's extraction and drafting path, and the stated prompt is permissive enough that manipulated content can influence the draft."}, {"id": "T2", "category": "T", "subtype": "invoice content tampering before ingestion", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "stored invoice record integrity", "threat_vector": "A malicious sender tampers with invoice fields, totals, or attachments so the stored record reflects altered payment facts.", "preconditions": ["Attacker can influence the incoming invoice document or evidence.", "The system stores ingested content for later review without an independent origin check."], "impact": "The stored record may preserve altered fields and support an incorrect payment draft or audit trail.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "The document store keeps a copy of ingested invoice attachments for later review.", "justification": "The path preserves attacker-supplied content into the internal store, and no integrity check or origin verification is stated for the stored document."}, {"id": "T3", "category": "R", "subtype": "approval trace ambiguity", "component_ids": ["C5", "C6"], "flow_ids": ["F4", "F5"], "boundary_ids": ["B2"], "asset": "payment approval accountability", "threat_vector": "A reviewer or operator later denies having approved a payment instruction when the workflow only records a click-through approval step.", "preconditions": ["A payment reaches the approval portal.", "The workflow record does not specify tamper-evident approval logging or signer binding."], "impact": "Disputes about who approved a payment can block incident review or reconciliation.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "The approver portal requires a logged-in reviewer to click approve before payment instructions are forwarded.", "justification": "The architecture shows a manual approval step, but it does not specify whether approval is bound to a durable audit record; the control is present but its evidentiary strength is unspecified."}, {"id": "T4", "category": "I", "subtype": "stored invoice exposure", "component_ids": ["C3"], "flow_ids": ["F2"], "boundary_ids": [], "asset": "invoice attachments and parsed records", "threat_vector": "A user or operator with access to the document store can read sensitive invoice attachments or parsed payment details.", "preconditions": ["The document store is readable by an internal user or service beyond the immediate processing need.", "Sensitive invoice data is retained in the store."], "impact": "Vendor banking details, invoice amounts, or other business data may be disclosed.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "The document store keeps a copy of ingested invoice attachments for later review.", "justification": "Retention is explicit, but the architecture does not specify access restrictions or redaction, so disclosure is possible but not established as exposed."}], "scope_notes": ["Boundary B1 is the external-to-internal crossing between Z1 and Z2 via F1.", "Boundary B2 is the internal-to-approval-zone crossing between Z2 and Z3 via F4 and F5.", "Unsupported paths remain incomplete for the document store access model and the exact approval audit trail.", "The architecture does not state whether any content validation, malware scanning, or OCR confidence checks are enforced before drafting.", "No privileged attacker access is assumed; analysis only considers externally influenced documents and normal workflow interactions."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4"], "recommendation": "Constrain the invoice agent to treat invoice text as data only, require structured field extraction against a strict schema, and block draft generation when the document contains instruction-like language or missing critical fields.", "verification": "Submit benign sample invoices with injected instruction text and confirm the agent extracts fields without following the embedded instructions and that incomplete invoices are routed to manual review.", "existing_control_effect": "CTRL1 currently allows the agent to continue on incomplete or loosely formed documents, so the path is only weakly constrained.", "residual_risk": "A sophisticated document can still bias extraction unless stronger validation and human exception handling are added."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Add integrity tagging and source verification at ingestion, and store attachments in an append-only record with hashes tied to the original submission metadata.", "verification": "Ingest a test invoice, alter a copy, and confirm the stored hash or provenance check detects the mismatch and prevents the altered version from replacing the original record.", "existing_control_effect": "CTRL2 preserves the incoming attachments for later review, which helps traceability but does not by itself prevent tampering.", "residual_risk": "If the sender is malicious, the system can still store a forged but internally consistent document unless external authenticity checks are added."}, {"threat_id": "T3", "enforcement_component_ids": ["C5", "C6"], "recommendation": "Bind each approval action to a unique approver identity, timestamp, and immutable audit entry before forwarding to payment execution.", "verification": "Perform a normal approval in a test environment and confirm the resulting record contains the approver identity and cannot be edited without detection.", "existing_control_effect": "CTRL3 already requires a logged-in reviewer to approve, which reduces accidental forwarding but does not prove durable non-repudiation.", "residual_risk": "A legitimate approver may still dispute context or intent if the audit trail is incomplete or alterable."}, {"threat_id": "T4", "enforcement_component_ids": ["C3"], "recommendation": "Restrict document-store access to the invoice agent and authorized reviewers only, and redact or segregate sensitive payment fields from general review copies.", "verification": "Attempt access from an unauthorized internal role in a test setup and confirm the store denies reads and exposes only redacted review copies.", "existing_control_effect": "CTRL2 retains invoice attachments for later review, which supports operations but does not state any read restriction or redaction.", "residual_risk": "Operational reviewers may still see necessary sensitive details unless finer-grained field protection is introduced."}]}
{"id": "E037", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "no_reviewer", "trial_id": "slot-01-no_reviewer", "topology_family_id": "topology-032", "annotation_status": "candidate", "image_path": "images/comparison/E037.png", "title": "Support drafting agent with external docs and account lookup", "purpose": "Assist support staff by drafting customer replies from externally supplied help documents and customer-account records, under weak control over untrusted inputs and tool outputs.", "assumptions": ["Support agents can ingest externally supplied help documents from outside the trusted application zone.", "The agent has read access to a customer-account store for case context.", "Control is weak or permissive: input sanitization, provenance checks, and instruction filtering are limited or inconsistently enforced.", "The attacker can influence external documents or evidence but has no privileged internal access.", "The architecture is one connected functional flow with no extra padding components."], "controls": [{"id": "K1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "External documents are accepted with only basic format checks; provenance labeling is present but not consistently enforced before drafting."}, {"id": "K2", "component_ids": ["C2", "C4"], "flow_ids": ["F2", "F3"], "fact": "The drafting agent has read-only account lookup access for case context; access is scoped to support cases but query filtering is weak and relies on user-provided identifiers."}, {"id": "K3", "component_ids": ["C2", "C3", "C5"], "flow_ids": ["F4", "F5"], "fact": "Drafts are reviewed by support staff before sending, but the review is advisory and does not reliably block unsafe or misleading content."}, {"id": "K4", "component_ids": ["C5", "C6"], "flow_ids": ["F6"], "fact": "Outbound replies are sent through the normal customer-support channel without additional content scanning or approval gates."}], "architecture": {"components": [{"id": "C1", "label": "Help Document Feed", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support Drafting Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Reply Composer", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Case Notes Store", "zone": "Z3", "kind": "data_store"}, {"id": "C5", "label": "Support Staff Console", "zone": "Z2", "kind": "process"}, {"id": "C6", "label": "Customer Reply Queue", "zone": "Z4", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "Help articles, pasted evidence, and other externally supplied support documents", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C4", "data": "Account lookup request using case or customer identifiers", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C4", "target": "C2", "data": "Customer history, plan status, recent tickets, and account context", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C2", "target": "C3", "data": "Draft reply text, cited snippets, and suggested next actions", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C5", "data": "Prepared response for review and edit", "boundary_ids": []}, {"id": "F6", "source": "C5", "target": "C6", "data": "Final outbound customer reply", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support App"}, {"id": "Z3", "label": "Customer Data"}, {"id": "Z4", "label": "Outbound Channel"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction injection via external docs", "component_ids": ["C1", "C2", "C3", "C5", "C6"], "flow_ids": ["F1", "F4", "F5", "F6"], "boundary_ids": ["B1", "B3"], "asset": "Support drafting workflow and outbound customer reply quality", "threat_vector": "Attacker places malicious instructions or hidden directives in externally supplied help documents that the drafting agent ingests and propagates into draft and final replies.", "preconditions": ["Attacker can influence external documents or evidence outside the trusted application zone.", "Agent processes externally supplied help documents during drafting."], "impact": "Injected instructions can steer the agent to produce unsafe, misleading, or attacker-chosen customer replies, affecting what is reviewed and sent outbound.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3", "K4"], "evidence_quote": "External documents are accepted with only basic format checks; provenance labeling is present but not consistently enforced before drafting.", "justification": "The architecture explicitly accepts untrusted external documents into the agent path, provenance is inconsistently enforced, review is only advisory, and outbound sending has no extra scan or gate, so the manipulation path is directly supported."}, {"id": "T2", "category": "I", "subtype": "customer-account data overexposure via weak lookup filtering", "component_ids": ["C2", "C4", "C3", "C5"], "flow_ids": ["F2", "F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "Customer account history and related case context", "threat_vector": "The drafting agent queries the customer-data store using user-provided identifiers and may retrieve more context than intended for the active case.", "preconditions": ["Attacker can influence case or customer identifiers used in lookup requests.", "The agent issues read queries against the customer-account store."], "impact": "Sensitive customer history, plan status, or recent tickets may be included in internal drafts and reviewed by staff beyond the intended need-to-know context.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The drafting agent has read-only account lookup access for case context; access is scoped to support cases but query filtering is weak and relies on user-provided identifiers.", "justification": "This supports a feasible disclosure path from customer-data store to the drafting workflow because the lookup is weakly filtered and the returned account context is fed into drafts and staff review, creating an overexposure risk."}, {"id": "T3", "category": "I", "subtype": "draft-to-outbound sensitive content leakage", "component_ids": ["C2", "C3", "C5", "C6"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": ["B3"], "asset": "Sensitive account context embedded in replies", "threat_vector": "Sensitive details from account context or cited snippets are carried into the prepared response and sent through the normal outbound channel.", "preconditions": ["The draft contains sensitive customer context or internal notes.", "Support staff forwards the draft through the normal outbound path."], "impact": "Confidential case details may be disclosed to the customer or another external recipient via the final reply.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Drafts are reviewed by support staff before sending, but the review is advisory and does not reliably block unsafe or misleading content.", "justification": "The outbound path exists, but the architecture does not specify content-loss prevention or a reliable blocking control; the review step may reduce risk but effectiveness is explicitly limited, so disclosure is possible but not firmly demonstrated as exposed."}], "scope_notes": ["Boundary B2 covers both C2<->C4 flows; the exact query shape and returned fields are not specified.", "The architecture does not specify whether drafts may contain raw customer-data excerpts, redactions, or citations only.", "No direct evidence shows unauthorized access to the case store beyond the agent's scoped read access; unsupported escalation paths are out of scope.", "The outbound channel is normal customer support messaging; whether recipients other than the customer can receive it is unspecified."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C5", "C6"], "recommendation": "Add enforced prompt-injection controls in the drafting pipeline: treat C1 content as untrusted, strip or quarantine instruction-like text before C2 uses it, require provenance tags to be preserved into C3, and block outbound sends at C5 unless the reply is marked as user-approved and free of untrusted directives.", "verification": "Use benign test documents containing fake directives and hidden instructions, then confirm C2 ignores them, C3 preserves provenance labels, and C5 prevents sending until a human rechecks the draft.", "existing_control_effect": "K1 provides only basic format checks and inconsistent provenance enforcement; K3 review is advisory; K4 sends without extra gating, so current controls reduce but do not reliably stop injection-driven reply manipulation.", "residual_risk": "Residual risk remains from subtle prompt manipulation, model misinterpretation, and reviewer error, especially if malicious content is blended into legitimate help text."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C4", "C3"], "recommendation": "Constrain C2 lookup requests to server-side validated case identifiers only, ignore user-supplied free-form customer identifiers unless they match the active case, and return only the minimum fields needed for drafting; redact nonessential history before passing context to C3.", "verification": "Run benign case-lookup tests with mismatched or extra identifiers and confirm C2 only receives the active case record and a minimal field set, with no unrelated ticket history exposed in the draft path.", "existing_control_effect": "K2 already makes the store read-only and case-scoped, but weak filtering and user-provided identifiers still allow overbroad retrieval; K3 may expose retrieved data to staff review, so current controls only partially limit disclosure.", "residual_risk": "Some overexposure remains if the active case itself legitimately contains sensitive details needed for support, or if downstream drafting reintroduces more context than intended."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C5", "C6"], "recommendation": "Add an outbound redaction and approval gate before C5 sends messages: detect account numbers, internal notes, and other sensitive snippets in C3 output, require explicit human confirmation for any redacted content, and prevent delivery through C6 unless the message passes the check.", "verification": "Use benign sample drafts containing mock sensitive placeholders and confirm the gate flags them, forces review, and blocks sending until the content is edited or explicitly approved.", "existing_control_effect": "K3 provides advisory review and K4 uses the normal outbound channel without extra scanning, so some leakage is already reviewed but not reliably prevented.", "residual_risk": "Residual risk remains for unrecognized sensitive text, context-dependent disclosures, and reviewer approval of content that should have been redacted."}]}
{"id": "E038", "collection": "comparison", "domain": "incident_response", "control_profile": "weak", "generation_condition": "single_prompt", "trial_id": "slot-13-single_prompt", "topology_family_id": "topology-017", "annotation_status": "candidate", "image_path": "images/comparison/E038.png", "title": "incident_response", "purpose": "Two analyst agents process externally supplied incident evidence and send findings through a coordinator to a case store.", "assumptions": ["Synthetic design; no actual production data.", "Attacker can supply crafted external documents or evidence, but has no privileged access to internal systems.", "Analyst agents may follow embedded instructions in received evidence unless constrained by local handling rules."], "controls": [{"id": "CL1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External evidence is routed to the analyst agents without a malware-sandbox requirement."}, {"id": "CL2", "component_ids": ["C4"], "flow_ids": ["F3", "F4", "F5"], "fact": "The coordinator appends analyst findings to the case store after a brief human review step."}, {"id": "CL3", "component_ids": ["C5"], "flow_ids": ["F5"], "fact": "Case records are stored with project-scoped access control in the case store."}], "architecture": {"components": [{"id": "C1", "label": "External Evidence Inbox", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Malware Analyst Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Log Triage Agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Finding Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Record Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, attachments, embedded instructions", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident logs, packet captures, embedded instructions", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analysis notes, extracted indicators, recommended actions", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "triage summary, indicators, confidence notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case findings, status updates, evidence references", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Analysis Workspace"}, {"id": "Z3", "label": "Case Management"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation via evidence payload", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "analysis integrity and downstream case findings", "threat_vector": "Crafted external evidence embeds instructions that the malware analyst agent may follow, influencing summaries or recommended actions sent to the coordinator.", "preconditions": ["Attacker can submit the external evidence document.", "The analyst agent treats evidence content as potentially actionable text."], "impact": "Misleading findings or unsafe analyst actions can be written into case records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CL1"], "evidence_quote": "External evidence is routed to the analyst agents without a malware-sandbox requirement.", "justification": "The flow from the external source into the analyst agent is permissive, and the architecture states embedded instructions may be present in received evidence."}, {"id": "T2", "category": "I", "subtype": "sensitive finding disclosure through case store", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": ["B2"], "asset": "case findings and evidence references", "threat_vector": "The coordinator writes detailed findings and evidence references into the case store, creating a disclosure path if project-scoped access is broader than intended or if records are over-detailed.", "preconditions": ["An unauthorized recipient can access the case store or a broader project scope than intended exists.", "Case records include sensitive evidence references."], "impact": "Incident details may be exposed to unauthorized readers of the case store.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CL3"], "evidence_quote": "Case records are stored with project-scoped access control in the case store.", "justification": "Access control is present, but the effectiveness and scope of the project boundary are not specified, so the path is not clearly reduced or exposed."}, {"id": "T3", "category": "T", "subtype": "finding tampering before persistence", "component_ids": ["C3", "C4", "C5"], "flow_ids": ["F4", "F5"], "boundary_ids": ["B2"], "asset": "case integrity", "threat_vector": "A compromised or misled log triage agent can submit altered indicators or confidence notes that the coordinator persists into the case store.", "preconditions": ["The triage agent can generate incorrect output from manipulated logs or evidence.", "The coordinator trusts agent output without independent validation."], "impact": "Incorrect case records could affect incident response decisions and audit quality.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CL2"], "evidence_quote": "The coordinator appends analyst findings to the case store after a brief human review step.", "justification": "A review step exists, but its rigor and validation depth are unspecified, so tampering resistance is uncertain."}], "scope_notes": ["Boundary B1 covers traffic from Z1 to Z2 for F1 and F2; B2 covers traffic from Z2 to Z3 for F5 and also the same zone handoff inside the analysis workspace is not a boundary.", "No privileged internal access is assumed; threats rely on attacker-controlled external content rather than direct system compromise.", "The architecture does not specify sandboxing, prompt-hardening, or content validation for analyst agents, so instruction-manipulation resistance remains unsupported.", "The case store access scope is only described as project-scoped; the exact authorization model and recipient set are incomplete."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4"], "recommendation": "Add strict content-isolation rules at the analyst agents and require the coordinator to reject or redact any analyst output that contains executable instructions, tool directives, or requests to alter system behavior.", "verification": "Submit a benign evidence sample containing fake embedded instructions and confirm the agents summarize it as data only, while the coordinator strips directive-like text before persistence.", "existing_control_effect": "The existing routing control is weak: external evidence reaches the analyst agents without a malware-sandbox requirement.", "residual_risk": "Residual risk remains from subtle prompt-injection patterns or analyst over-interpretation of malicious evidence text."}, {"threat_id": "T2", "enforcement_component_ids": ["C5", "C4"], "recommendation": "Constrain case-store readers to least-privilege project roles and have the coordinator store minimized findings with evidence references separated from free-text summaries.", "verification": "Use a normal case entry and confirm a non-project account cannot open it, while a project account sees only the minimum required fields.", "existing_control_effect": "Project-scoped access control is already present, but its scope is not fully specified.", "residual_risk": "Residual risk remains if project membership is mismanaged or if sensitive details are copied into free-text fields."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C4", "C5"], "recommendation": "Require the coordinator to validate key indicators against the original evidence bundle before writing to the case store and keep a change log of human-reviewed edits.", "verification": "Inject a harmless mismatch between a sample log and the triage summary, then confirm the coordinator flags the discrepancy before persistence.", "existing_control_effect": "A brief human review step exists before appending findings to storage.", "residual_risk": "Residual risk remains because the depth and consistency of the review step are unspecified and may miss subtle tampering."}]}
{"id": "E039", "collection": "comparison", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-18-no_reviewer", "topology_family_id": "topology-033", "annotation_status": "candidate", "image_path": "images/comparison/E039.png", "title": "Incident Evidence Triage with Dual Analysts and Case Coordinator", "purpose": "Model an incident-response workflow where two analyst agents review externally supplied evidence, consolidate findings through a coordinator, and persist case data in a shared store.", "assumptions": ["Two analyst agents independently process attacker-influenced incident documents, logs, or attachments.", "A coordinator receives analyst findings and writes case records to a case store.", "Relevant control implementation and effectiveness are unspecified; no assumption is made that validation, sandboxing, or authorization checks are present or absent.", "The attacker has no privileged internal access and can only influence external evidence inputs.", "The architecture is one connected functional flow with no extra padding components."], "controls": [{"id": "CTRL1", "component_ids": ["C1"], "flow_ids": ["F1", "F2", "F3"], "fact": "External evidence is routed through a dedicated intake component before reaching either analyst; the exact validation and filtering behavior is unspecified."}, {"id": "CTRL2", "component_ids": ["C2", "C3"], "flow_ids": ["F2", "F3", "F4", "F5"], "fact": "Two independent analyst agents review the same evidence in parallel and send findings to the coordinator; authorization and prompt/tool safeguards are unspecified."}, {"id": "CTRL3", "component_ids": ["C4", "C5"], "flow_ids": ["F6"], "fact": "Only the coordinator writes to the shared case store, creating a single aggregation path for persisted case data; write controls are unspecified."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z2", "kind": "process"}, {"id": "C2", "label": "Analyst A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Case Coordinator", "zone": "Z2", "kind": "agent"}, {"id": "C5", "label": "Case Store", "zone": "Z2", "kind": "data_store"}, {"id": "C6", "label": "External Evidence Feed", "zone": "Z1", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C6", "target": "C1", "data": "incident documents, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C2", "data": "filtered evidence bundle", "boundary_ids": []}, {"id": "F3", "source": "C1", "target": "C3", "data": "filtered evidence bundle", "boundary_ids": []}, {"id": "F4", "source": "C2", "target": "C4", "data": "analysis findings, extracted indicators", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C4", "data": "analysis findings, extracted indicators", "boundary_ids": []}, {"id": "F6", "source": "C4", "target": "C5", "data": "case record updates, merged findings", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Evidence Source"}, {"id": "Z2", "label": "Incident-Response App Boundary"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "evidence tampering via external intake", "component_ids": ["C6", "C1", "C2", "C3", "C4", "C5"], "flow_ids": ["F1", "F2", "F3", "F4", "F5", "F6"], "boundary_ids": ["B1"], "asset": "incident evidence and resulting case record", "threat_vector": "Attacker-controlled external documents, logs, or attachments are supplied over F1 into Evidence Intake and can influence the filtered bundles delivered to both analysts, then downstream findings and stored case records.", "preconditions": ["Attacker can influence the external evidence input at C6.", "The intake filter/validation behavior is unspecified, so its resistance to malformed or deceptive content is unknown."], "impact": "Corrupted or misleading evidence can alter analyst findings and persist tainted case records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2", "CTRL3"], "evidence_quote": "External evidence is routed through a dedicated intake component before reaching either analyst; the exact validation and filtering behavior is unspecified.", "justification": "The architecture explicitly accepts attacker-influenced evidence from Z1 across B1 into the intake path and forwards filtered bundles to both analysts, with unspecified validation and downstream write controls. This supports a feasible tampering path affecting the case record."}, {"id": "T2", "category": "I", "subtype": "sensitive evidence disclosure to dual analysts", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F2", "F3"], "boundary_ids": [], "asset": "incident evidence content", "threat_vector": "Evidence flowing from intake to two analyst agents may be exposed to both recipients during parallel review, increasing internal dissemination of sensitive documents, logs, or attachments.", "preconditions": ["The evidence bundle contains sensitive material.", "Analysts are not all intended recipients for every item in the bundle."], "impact": "Sensitive evidence may be seen by more internal agents than necessary.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2"], "evidence_quote": "Two independent analyst agents review the same evidence in parallel and send findings to the coordinator; authorization and prompt/tool safeguards are unspecified.", "justification": "The design clearly duplicates the evidence path to two analysts, but it does not specify confidentiality rules, filtering scope, or per-item access restrictions. That makes unauthorized disclosure plausible but not established from the architecture alone."}, {"id": "T3", "category": "A", "subtype": "prompt or context manipulation of analyst workflow", "component_ids": ["C6", "C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1"], "asset": "analyst decision quality and coordinator findings", "threat_vector": "Attacker-crafted external incident content can be carried through intake into analyst review, where it may manipulate analyst interpretation or unsafe tool/context use if the agents consume evidence as instructions or embedded directives.", "preconditions": ["Analysts process attacker-influenced evidence content as part of their working context.", "No effective prompt/tool safeguards are specified."], "impact": "Misleading instructions or context can distort analysis outcomes and propagate incorrect findings to the coordinator.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL2"], "evidence_quote": "Two independent analyst agents review the same evidence in parallel and send findings to the coordinator; authorization and prompt/tool safeguards are unspecified.", "justification": "This is an agent-specific attack path: untrusted external evidence crosses B1, enters the intake path, and is then presented to analyst agents with no specified prompt or tool safeguards. The architecture therefore supports a feasible context-manipulation risk."}, {"id": "T4", "category": "D", "subtype": "case store write amplification or contention", "component_ids": ["C4", "C5"], "flow_ids": ["F6"], "boundary_ids": [], "asset": "case store availability", "threat_vector": "Repeated or oversized merged findings from the coordinator can pressure the single write path into the shared case store, potentially slowing or blocking case updates.", "preconditions": ["The coordinator accepts large or frequent updates from analysts.", "Write limits, throttling, or validation are unspecified."], "impact": "Case record updates may be delayed or disrupted.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3"], "evidence_quote": "Only the coordinator writes to the shared case store, creating a single aggregation path for persisted case data; write controls are unspecified.", "justification": "The architecture shows a single persisted write path, which creates a possible availability bottleneck, but it does not specify whether backpressure, limits, or validation exist. This supports only an unknown assessment."}], "scope_notes": ["Boundary B1 is the only cross-zone boundary; flows F2-F6 stay within Z2 and have no boundary IDs.", "No unsupported direct access from the external source to the case store is shown; any such path would be speculative.", "The architecture does not specify whether analyst outputs are validated, whether the coordinator deduplicates or reviews conflicts, or whether the case store enforces access or integrity controls.", "No additional trust boundaries or hidden components should be inferred beyond the supplied diagram and control facts."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C4", "C5"], "recommendation": "Enforce content validation and normalization at Evidence Intake before forwarding any bundle; reject malformed or unexpected document types, strip active content, and require integrity-preserving handling of attachments. Add coordinator-side record validation so only schema-conformant, traceable findings can be written to the case store.", "verification": "Submit benign test samples with malformed headers, embedded scripts/macros, oversized attachments, and inconsistent metadata; confirm they are rejected or sanitized at C1 and that only schema-valid records reach C5 via C4.", "existing_control_effect": "CTRL1 already places all external evidence through C1, and CTRL3 keeps writes to C5 behind the coordinator, which limits direct tampering paths but does not specify validation strength.", "residual_risk": "If filtering is incomplete or downstream validation is weak, deceptive but syntactically valid evidence can still bias findings and persist into case records."}, {"threat_id": "T2", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Apply data minimization and need-to-know redaction in Evidence Intake so each analyst receives only the evidence items required for their task; enforce per-item access rules and separate sensitive attachments from routine case material.", "verification": "Use a non-sensitive test case containing labeled sub-items with different access needs; confirm each analyst receives only the permitted subset and that omitted items are not present in their bundle.", "existing_control_effect": "CTRL2 already separates the review into two analysts, but it duplicates the same evidence path to both and does not define confidentiality filtering or authorization, so exposure is reduced only by process separation, not by enforced restriction.", "residual_risk": "If both analysts must legitimately see the same evidence for triage, internal disclosure remains possible; any unfiltered sensitive content sent to both agents will still be visible to both."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Treat external evidence as untrusted data, not instructions: strip or neutralize embedded directives at intake, isolate analyst prompts from raw evidence text, and require analyst-side handling rules that prevent tool actions or workflow changes from evidence content alone.", "verification": "Use harmless sample evidence containing obvious instruction-like text and embedded command-style phrases; confirm the intake component neutralizes them and the analysts continue to process the case using only approved workflow prompts.", "existing_control_effect": "CTRL1 provides a dedicated intake point before analysts see the evidence, and CTRL2 already uses two independent analysts, but neither control specifies prompt isolation, sanitization, or tool-use guardrails.", "residual_risk": "Sophisticated prompt injection or indirect context manipulation may still influence analyst judgment if the content is not fully isolated from their working context or if workflow guardrails are incomplete."}, {"threat_id": "T4", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Add coordinator-side rate limits, bounded write sizes, and schema checks before each case-store update; queue or split large updates so repeated merged findings cannot monopolize the single write path.", "verification": "Generate benign bursts of normal-sized updates and one oversized update in a test environment; confirm C4 throttles or queues excess writes and that C5 remains responsive with valid records only.", "existing_control_effect": "CTRL3 already centralizes all writes through C4, which simplifies enforcement and blocks direct multi-writer contention, but write controls themselves are unspecified.", "residual_risk": "A sustained high volume of legitimate findings can still create backlog or delay even with throttling, especially if storage or queue capacity is limited."}]}
{"id": "E040", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "no_reviewer", "trial_id": "slot-16-no_reviewer", "topology_family_id": "topology-034", "annotation_status": "candidate", "image_path": "images/comparison/E040.png", "title": "Guarded multi-agent incident triage with provenance-filtered evidence handling", "purpose": "Model an incident response workflow where two analyst agents review externally supplied evidence, relay findings through a coordinator, and persist case notes in a shared store under enforced authorization and bounded tool access.", "assumptions": ["Externally supplied incident documents, logs, or attachments may be attacker-influenced but arrive outside the trusted application zone.", "Analyst agents can read evidence and produce findings, but only the coordinator can approve writes to the case store.", "Tool use is bounded to predefined retrieval and note-taking actions; no unrestricted shell or arbitrary network access is assumed.", "Provenance checks and authorization are enforced, but their completeness against malformed or misleading evidence is not guaranteed by the scenario facts."], "controls": [{"id": "K1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Evidence intake applies provenance tagging and content filtering before analyst consumption; analysts only receive the filtered bundle from the external zone."}, {"id": "K2", "component_ids": ["C2", "C3", "C6"], "flow_ids": ["F7", "F8", "F9", "F10"], "fact": "Analyst tool use is constrained to scoped retrieval and note-taking actions through the shared tooling component; no direct shell or network capability is modeled."}, {"id": "K3", "component_ids": ["C4", "C5"], "flow_ids": ["F5", "F6"], "fact": "Only the coordinator is authorized to write to the case store; analysts have read-only visibility through coordinator-mediated review of stored case history."}, {"id": "K4", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4", "F5"], "fact": "Findings are merged by the coordinator before persistence, creating an approval gate between independent analyst outputs and stored case notes."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake Filter", "zone": "Z1", "kind": "process"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Triage Coordinator", "zone": "Z2", "kind": "agent"}, {"id": "C5", "label": "Case Notes Store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Retrieval and Note Tooling", "zone": "Z2", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "filtered evidence bundle and provenance tags", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "filtered evidence bundle and provenance tags", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings and confidence notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings and confidence notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "approved case notes and disposition update", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C5", "target": "C4", "data": "existing case history for review", "boundary_ids": ["B2"]}, {"id": "F7", "source": "C2", "target": "C6", "data": "bounded retrieval request", "boundary_ids": []}, {"id": "F8", "source": "C3", "target": "C6", "data": "bounded retrieval request", "boundary_ids": []}, {"id": "F9", "source": "C6", "target": "C2", "data": "scoped evidence excerpt or case snippet", "boundary_ids": []}, {"id": "F10", "source": "C6", "target": "C3", "data": "scoped evidence excerpt or case snippet", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Evidence Zone"}, {"id": "Z2", "label": "Trusted Triage Zone"}, {"id": "Z3", "label": "Case Store Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/evidence injection via filtered bundle", "component_ids": ["C1", "C2", "C3", "C6", "C4"], "flow_ids": ["F1", "F2", "F7", "F8", "F9", "F10", "F3", "F4", "F5"], "boundary_ids": ["B1", "B2"], "asset": "Analyst reasoning and downstream case notes", "threat_vector": "Attacker-influenced external evidence is filtered and then consumed by analysts; malicious content can manipulate analyst conclusions and bounded retrieval use, which may propagate into coordinator-approved notes.", "preconditions": ["Attacker can supply or influence external incident documents, logs, or attachments.", "The malicious content survives intake filtering/provenance tagging.", "Analysts rely on evidence content when forming findings."], "impact": "Misleading or unsafe analyst outputs can be merged by the coordinator and persisted as incorrect case notes or dispositions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K4"], "evidence_quote": "Evidence intake applies provenance tagging and content filtering before analyst consumption; analysts only receive the filtered bundle from the external zone.", "justification": "The architecture explicitly accepts externally supplied, attacker-influenced evidence and forwards filtered content to analysts. Scoped tools do not prevent content-driven manipulation, and coordinator merging can propagate compromised findings. Control effectiveness against malformed or misleading evidence is not guaranteed."}, {"id": "T2", "category": "T", "subtype": "tampering with case notes through approved write path", "component_ids": ["C4", "C5"], "flow_ids": ["F5", "F6"], "boundary_ids": ["B2"], "asset": "Case notes and disposition history", "threat_vector": "A malicious or compromised coordinator can submit altered case notes or disposition updates to the shared store through the only write path.", "preconditions": ["Attacker gains control of or can impersonate the coordinator role.", "The coordinator accepts or generates incorrect writes before store persistence."], "impact": "Stored case history can be modified to contain false or incomplete incident records, affecting later review and disposition.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Only the coordinator is authorized to write to the case store; analysts have read-only visibility through coordinator-mediated review of stored case history.", "justification": "The write path to the case store is explicit and centralized in the coordinator. Authorization reduces access to a single trusted role, but if that role is subverted or misused, tampering of stored notes remains feasible on the modeled path."}, {"id": "T3", "category": "I", "subtype": "case-history disclosure through retrieval tooling", "component_ids": ["C4", "C5", "C6", "C2", "C3"], "flow_ids": ["F6", "F7", "F8", "F9", "F10"], "boundary_ids": ["B2"], "asset": "Existing case history and scoped evidence excerpts", "threat_vector": "A user of the bounded retrieval tool may obtain more case-history detail than intended through scoped excerpts or coordinator-mediated review.", "preconditions": ["An analyst can issue retrieval requests through the shared tooling component.", "Returned snippets or case history contain sensitive details not intended for that analyst's task."], "impact": "Sensitive incident history or notes may be exposed to analysts beyond the minimum necessary for triage.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "Analyst tool use is constrained to scoped retrieval and note-taking actions through the shared tooling component; no direct shell or network capability is modeled.", "justification": "The tooling is bounded, which reduces broad exfiltration paths, but the scenario does not specify the exact scoping rules or whether returned excerpts exclude sensitive content. The store read path exists, so disclosure risk is possible but not established as exposed."}], "scope_notes": ["No direct spoofing, repudiation, or elevation path is clearly supported by the supplied flows and controls.", "Boundary IDs inferred from the declared trust zones: F1/F2 cross B1; F5/F6 cross B2; same-zone flows have no boundary.", "The scenario does not specify whether provenance checks reject prompt-injection-like text, whether retrieval snippets are redacted, or whether coordinator review detects maliciously altered analyst findings.", "A data-store read alone is not sufficient to prove disclosure to an unauthorized recipient; the third threat is therefore kept unknown rather than exposed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C6", "C4"], "recommendation": "Harden intake and tool mediation with strict content sanitization at C1, provenance-aware prompt segregation in C6, and mandatory coordinator-side review of any analyst output that cites or reacts to external evidence before F5 writes.", "verification": "Run benign test cases with deliberately malformed but harmless evidence text and confirm C1 strips/flags the payload, C6 keeps retrieved snippets separate from instructions, and C4 rejects or annotates any analyst finding that depends on untrusted directives before persistence.", "existing_control_effect": "K1 already filters and tags inbound evidence, K2 limits tool use to scoped retrieval, and K4 creates an approval gate before storage; these reduce but do not eliminate content-driven manipulation.", "residual_risk": "A carefully crafted malicious document may still influence analyst reasoning if it passes filtering or appears semantically benign, so some poisoned or misleading findings can still reach coordinator review."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Require coordinator writes to C5 to be signed, versioned, and policy-checked so that only explicitly approved case-note deltas are accepted, with append-only audit records for every disposition update.", "verification": "Submit a benign draft note and a benign rejected update through the normal coordinator workflow and confirm C5 stores only the approved delta, preserves prior versions, and records an audit entry for each attempted write.", "existing_control_effect": "K3 restricts write access to the coordinator and K4 places a merge/approval step before persistence, which narrows the write path to one trusted role and one approval point.", "residual_risk": "If the coordinator role is compromised or misused, the attacker can still author approved-looking but false updates within the authorized path; stronger integrity checks reduce but do not remove that risk."}, {"threat_id": "T3", "enforcement_component_ids": ["C6", "C4", "C5"], "recommendation": "Constrain retrieval to explicit per-task fields and redaction rules in C6, have C4 enforce least-necessary disclosure before forwarding snippets, and store only access-scoped notes in C5 so analysts receive minimum required history.", "verification": "Perform benign retrieval requests for known sample cases and confirm C6 returns only the configured fields, C4 suppresses out-of-scope details, and C5 responses match the redaction policy without revealing extra history.", "existing_control_effect": "K2 already bounds tool actions and K3 limits store access through the coordinator, which reduces broad exposure paths and keeps reads mediated.", "residual_risk": "If the scoped retrieval rules are too broad or the stored notes contain sensitive details in allowed fields, analysts may still see more context than intended even without any direct policy bypass."}]}
{"id": "E041", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "full_workflow", "trial_id": "slot-06-full_workflow", "topology_family_id": "topology-035", "annotation_status": "candidate", "image_path": "images/comparison/E041.png", "title": "Agent-assisted support response assembly", "purpose": "A support agent drafts replies by combining externally supplied help documents with customer-account data in a response workspace before preparing an outbound response.", "assumptions": ["The support agent operates in a separate application zone from the external content source and the customer-account store.", "Relevant control implementation and effectiveness are unspecified; no assumption is made that filtering, sandboxing, or approval gates are absent or effective.", "The attacker can influence external help documents or evidence available to the agent but has no privileged access to the support system or account store.", "The architecture is connected end-to-end and contains only the components needed for document ingestion, account lookup, response drafting, and outbound reply preparation."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1"], "fact": "The agent reviews externally supplied help content in the drafting workspace before composing a reply; specific validation or content-filter behavior is unspecified."}, {"id": "K2", "component_ids": ["C2", "C5"], "flow_ids": ["F2"], "fact": "Account data is consulted read-side for personalization; the store is not modeled as writable from this workflow."}, {"id": "K3", "component_ids": ["C3", "C4"], "flow_ids": ["F4"], "fact": "The draft is handed from the workspace to the reply preparation step for final formatting and review; enforcement strength is unspecified."}, {"id": "K4", "component_ids": ["C2", "C4"], "flow_ids": ["F5"], "fact": "Outbound reply preparation remains within the support workspace boundary before send-off; downstream delivery controls are outside this diagram."}], "architecture": {"components": [{"id": "C1", "label": "Help document feed", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Response drafting workspace", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Reply preparation service", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Customer account store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "External help content and evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C5", "target": "C2", "data": "Customer account details", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C2", "target": "C3", "data": "Draft reply content and notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "Reviewed response draft", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C2", "data": "Outbound reply package", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External sources"}, {"id": "Z2", "label": "Support workspace"}, {"id": "Z3", "label": "Customer account service"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/content manipulation leading to unsafe reply drafting", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F3", "F4", "F5"], "boundary_ids": ["B1"], "asset": "Support response integrity and agent decision quality", "threat_vector": "Attacker-controlled help content can shape what the support agent reads and then copies into draft notes and the outbound reply package.", "preconditions": ["Attacker can influence external help documents or evidence available to the agent.", "The agent uses that external content while drafting replies."], "impact": "Misleading or unsafe reply content may be produced and sent, causing incorrect support actions or customer harm.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3", "K4"], "evidence_quote": "The agent reviews externally supplied help content in the drafting workspace before composing a reply; specific validation or content-filter behavior is unspecified.", "justification": "The design explicitly routes external content into the agent workspace for drafting, and the control text leaves validation unspecified. This supports a feasible agent-content manipulation path; downstream preparation remains within the workspace, so the issue is unsafe reply influence rather than a storage compromise."}, {"id": "T2", "category": "I", "subtype": "account-data exposure through support workflow", "component_ids": ["C5", "C2", "C3", "C4"], "flow_ids": ["F2", "F3", "F4", "F5"], "boundary_ids": ["B2"], "asset": "Customer account details", "threat_vector": "Customer account details are read into the support workspace and may be incorporated into draft notes or reply content beyond the minimum necessary for handling the case.", "preconditions": ["The support agent can access account details for the current case.", "The draft/reply content may preserve or expose those details in the support workflow."], "impact": "Account details could be disclosed to an unintended recipient if included in the outbound response or retained in the workspace more broadly than intended.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3", "K4"], "evidence_quote": "Account data is consulted read-side for personalization; the store is not modeled as writable from this workflow.", "justification": "There is a cross-boundary read from the account store into the support workspace, so sensitive data is present in the workflow. However, the architecture does not specify what exact fields are exposed, whether minimization occurs, or whether any of that data leaves the intended support context, so disclosure is possible but not explicit enough to mark exposed."}, {"id": "T3", "category": "R", "subtype": "insufficient traceability for drafted support responses", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4", "F5"], "boundary_ids": [], "asset": "Support response accountability", "threat_vector": "Draft content and final reply preparation occur within the support workspace without specified logging, approval, or non-repudiation controls.", "preconditions": ["An incorrect or harmful response is produced.", "The workflow lacks effective, specified audit or signature controls."], "impact": "The organization may be unable to determine who approved or changed the response content after the fact.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "The draft is handed from the workspace to the reply preparation step for final formatting and review; enforcement strength is unspecified.", "justification": "The diagram shows a multi-step drafting path, but it does not specify audit logging, immutable approvals, or signature-based accountability. That leaves repudiation risk plausible, but the control effectiveness is not described well enough to assert an exposed gap."}], "scope_notes": ["Boundary B1 covers the external content source to support workspace path; B2 covers the account service to support workspace path.", "No dedicated logging, approval, or content provenance mechanism is modeled in the provided controls.", "Downstream message delivery after reply preparation is explicitly outside the diagram, so recipient-side disclosure or delivery failure is out of scope.", "The architecture does not state whether externally supplied help content is sanitized, ranked, or sandboxed before the agent reads it.", "The support agent is treated as a trusted workspace actor for the model; attacker influence is limited to external inputs, not privileged system access."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Require the support agent to use a provenance-labeled drafting workflow: only copy or summarize external help content after it is reviewed in C2 with content provenance visible in C3, and block outbound draft fields in C4 from including unapproved external instructions or copied text unless the agent explicitly confirms a case-relevant use.", "verification": "Run a benign test with harmless but misleading sample help text and confirm the agent can see provenance labels, that unapproved snippets are not auto-transferred into the outbound package, and that any manual override is recorded before F5.", "existing_control_effect": "K1 shows external content is already reviewed in the workspace, and K3/K4 keep drafting and preparation inside the support zone; these controls reduce direct cross-boundary injection but do not specify content validation or approval strength.", "residual_risk": "An attacker who can influence external help content may still bias the agent’s judgment or induce mistaken summaries if the agent follows the content despite labels; the workflow reduces but does not eliminate unsafe reply influence."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Minimize account-data use in the support workspace by exposing only case-needed fields to C2, and require C3/C4 to redact or suppress sensitive account values from draft notes and outbound reply text unless they are strictly necessary for the resolved case.", "verification": "Use a non-sensitive test account and confirm the agent can complete a mock case while C3/C4 only display and carry the minimum allowed fields, with redacted placeholders appearing in the reply package where sensitive values are not required.", "existing_control_effect": "K2 already indicates read-side access for personalization and no writable path from the workflow, which reduces tampering risk against the store; K3/K4 keep processing inside the workspace, but they do not guarantee minimization or redaction.", "residual_risk": "Some account details must still be read into the workflow for support handling, so disclosure risk remains if the agent intentionally includes them or if the drafting tools preserve more data than intended."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Add immutable step-level audit records for draft creation and reply preparation, including who reviewed the content in C2 and what was handed from C3 to C4, and require a final human confirmation in C2 before F5 is used for send-off.", "verification": "Perform a harmless draft exercise and confirm the system records a timestamped review trail for each handoff, with no ability to alter or delete the record from the workspace, and that the reply cannot proceed without the final confirmation step.", "existing_control_effect": "K3 and K4 already describe a structured handoff through the workspace, which helps define responsibility points, but the controls do not specify logging or non-repudiation enforcement.", "residual_risk": "If audit records are incomplete or if a user shares credentials or delegates work informally, later attribution may still be disputed even though the workflow is better traced."}]}
{"id": "E042", "collection": "comparison", "domain": "finance_operations", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-07-full_workflow", "topology_family_id": "topology-036", "annotation_status": "candidate", "image_path": "images/comparison/E042.png", "title": "Invoice-to-Payment Drafting with Weakly Controlled Document Intake", "purpose": "Model a finance workflow where an invoice-handling agent ingests externally supplied documents, extracts payment details, drafts payment requests, and forwards them into approval and payment execution under weak control assumptions.", "assumptions": ["Externally supplied invoice files and supporting evidence can be attacker-influenced before ingestion.", "The invoice agent can parse attachments and may follow embedded instructions unless blocked by weak or unspecified controls.", "Approval and payment-execution steps are connected to the same workflow, but validation of extracted fields is weak or partially unspecified.", "The attacker has no privileged internal access; influence is limited to external documents or evidence.", "This is a single connected workflow with no disconnected helper nodes."], "controls": [{"id": "K1", "component_ids": ["C2"], "flow_ids": ["F1"], "fact": "Document trust is permissive at intake; externally supplied invoices are accepted for parsing, with only weak screening of attachment content."}, {"id": "K2", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "fact": "Extraction output is used directly to populate payment drafts, with limited field normalization and no clearly specified strong cross-check before drafting."}, {"id": "K3", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "Approval review exists but its validation strength is unspecified beyond being weak/permissive, so it may rely on the drafted request with limited scrutiny."}, {"id": "K4", "component_ids": ["C5"], "flow_ids": ["F4", "F5"], "fact": "Payment execution follows the approved instruction path; downstream authorization is present, but the exact anti-fraud verification depth is not specified."}], "architecture": {"components": [{"id": "C1", "label": "Vendor Invoice Upload", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Invoice Intake Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Payment Draft Builder", "zone": "Z2", "kind": "process"}, {"id": "C4", "label": "Approval Review Step", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Payment Execution Service", "zone": "Z3", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice files, attachments, supporting evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "extracted invoice fields, payment details, drafted request", "boundary_ids": []}, {"id": "F3", "source": "C3", "target": "C4", "data": "payment draft, amount, beneficiary, justification", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C5", "data": "approved payment instruction", "boundary_ids": []}, {"id": "F5", "source": "C5", "target": "C1", "data": "payment status or remittance notice", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Finance Workflow"}, {"id": "Z3", "label": "Approval and Execution"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z1", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction manipulation via embedded invoice content", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "payment draft integrity and workflow execution", "threat_vector": "Attacker supplies invoice files or attachments that the intake agent parses and may follow as instructions, causing the extracted fields and drafted payment request to be manipulated.", "preconditions": ["Attacker can influence externally supplied invoice files before ingestion.", "The agent follows embedded instructions unless blocked by weak or unspecified controls."], "impact": "Incorrect payment details or justification can be injected into the payment draft, potentially steering the finance workflow toward an attacker-influenced request.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "Document trust is permissive at intake; externally supplied invoices are accepted for parsing, with only weak screening of attachment content.", "justification": "The architecture explicitly allows attacker-influenced external documents into the intake agent and uses extraction output directly to populate payment drafts with limited normalization. This supports a feasible agent-specific instruction/content manipulation path from Z1 to Z2."}, {"id": "T2", "category": "T", "subtype": "invoice field tampering in extraction output", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3"], "boundary_ids": ["B1", "B2"], "asset": "payment amount, beneficiary, and justification", "threat_vector": "Maliciously crafted invoice content causes the extraction pipeline to produce altered payment fields that are forwarded into the approval review step.", "preconditions": ["Attacker-controlled invoice content reaches the parser.", "No strong cross-check occurs before drafting or review."], "impact": "Tampered fields may propagate into the approval package, leading to incorrect payment instructions being reviewed and potentially approved.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K3"], "evidence_quote": "Extraction output is used directly to populate payment drafts, with limited field normalization and no clearly specified strong cross-check before drafting.", "justification": "The workflow shows a direct trust path from external document content to drafted payment details and onward across the Z2 to Z3 boundary. This supports a tampering candidate on the extracted payment data, though approval strength is only described as weak/permissive."}, {"id": "T3", "category": "I", "subtype": "remittance/status disclosure to external source", "component_ids": ["C5", "C1"], "flow_ids": ["F5"], "boundary_ids": ["B3"], "asset": "payment status or remittance notice", "threat_vector": "The payment execution service sends status or remittance information back to the external source, which may disclose transaction details to an untrusted party.", "preconditions": ["The external source that submitted the invoice is not fully trusted.", "The returned remittance/status message contains sensitive payment details."], "impact": "An external sender may learn payment timing or remittance details, exposing financial workflow information beyond intended recipients.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K4"], "evidence_quote": "Payment execution follows the approved instruction path; downstream authorization is present, but the exact anti-fraud verification depth is not specified.", "justification": "There is an explicit return flow from payment execution to the external entity across B3, but the architecture does not specify what information is included in the notice or whether it is restricted. That makes disclosure plausible but not clearly supported as exposed."}, {"id": "T4", "category": "D", "subtype": "workflow disruption through malformed document intake", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "invoice intake availability", "threat_vector": "An attacker submits malformed or complex invoice files and attachments that stress parsing and screening at intake.", "preconditions": ["Attacker can send externally supplied invoice files.", "Weak screening and parsing are applied to attachments."], "impact": "The intake agent may be delayed, overloaded, or unable to process invoices normally, disrupting the finance workflow.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Document trust is permissive at intake; externally supplied invoices are accepted for parsing, with only weak screening of attachment content.", "justification": "The intake path explicitly accepts untrusted external documents with weak screening, which supports a feasible denial-of-service candidate against the parser/intake step. The architecture does not quantify capacity or resilience, so only the exposure is stated."}], "scope_notes": ["Boundary IDs were assigned from the listed flows: B1 for Z1-Z2, B2 for Z2-Z3, and B3 for Z1-Z3.", "Unsupported paths not modeled as threats: same-zone flows F2 and F4 do not cross a trust boundary, and the architecture does not state unauthorized internal access.", "The exact contents of the approval review and remittance/status message are unspecified, so disclosure and approval-bypass consequences remain partially incomplete.", "No claim is made that controls are absent; where control effectiveness is unspecified, assessment remains unknown rather than exposed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3"], "recommendation": "At intake and drafting, treat invoice text/attachments as untrusted data: disable instruction-following on C2 for external documents, strip or quarantine embedded commands/macros, and require C3 to populate drafts only from a fixed schema with mandatory field allowlists and provenance tagging from F1 to F2.", "verification": "Submit a benign test invoice containing harmless embedded instructions and an out-of-schema payment field; confirm C2 ignores the instruction content, C3 drops the extra field, and only schema-approved values reach the draft.", "existing_control_effect": "K1 provides only weak screening at C2, and K2 shows direct use of extraction output; these controls reduce accidental issues but do not reliably stop agentic instruction injection or content steering.", "residual_risk": "A crafted document may still influence parsing edge cases or induce partial misextraction, but the direct instruction-following path should be substantially reduced."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C4"], "recommendation": "Before F3, require C3 to reconcile extracted payment fields against a trusted invoice record or vendor master and emit a signed, normalized draft; at C4, verify the signature and compare amount, beneficiary, and justification against the expected source values before approval.", "verification": "Use a benign invoice with one intentionally mismatched noncritical field; confirm the mismatch is flagged before C4 and the review step cannot approve until the discrepancy is resolved.", "existing_control_effect": "K2 allows extracted values to populate drafts with limited normalization, and K3 indicates only weak/permissive review; together they provide some review path but not a strong integrity check.", "residual_risk": "If the trusted reference data is stale or compromised, or if reviewers override alerts, a tampered field can still propagate."}, {"threat_id": "T3", "enforcement_component_ids": ["C5"], "recommendation": "Constrain F5 so C5 sends only minimal, preapproved remittance/status data to C1, redact payment-sensitive fields unless explicitly required, and add a policy check that logs and blocks any response exceeding the approved template.", "verification": "Trigger a normal successful payment in a test environment and confirm the return message contains only the minimum approved status fields and no full remittance details.", "existing_control_effect": "K4 confirms downstream authorization exists, which helps ensure messages originate from the proper execution path, but it does not specify payload minimization or recipient scoping.", "residual_risk": "If the approved template itself includes sensitive details, or if business requirements demand richer notices, some disclosure to the external sender remains inherent."}, {"threat_id": "T4", "enforcement_component_ids": ["C2"], "recommendation": "Add intake hardening at C2: file-size limits, attachment type restrictions, parser timeouts, decompression caps, and queue backpressure so malformed or overly complex invoices are rejected or isolated before parsing completion.", "verification": "Send benign oversized and nested-attachment sample files in a test run; confirm C2 times out or quarantines them without degrading processing for normal invoices.", "existing_control_effect": "K1 indicates permissive intake with weak screening, which may block only obvious bad content but leaves parser exhaustion and queue flooding largely available.", "residual_risk": "Well-formed but high-volume submissions can still pressure capacity, so operational rate limiting and monitoring remain needed."}]}
{"id": "E043", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "single_prompt", "trial_id": "slot-05-single_prompt", "topology_family_id": "topology-037", "annotation_status": "candidate", "image_path": "images/comparison/E043.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "Attacker can influence externally supplied documents or evidence but has no privileged access.", "Control implementation and effectiveness for validation, filtering, authorization, and logging are unspecified."], "controls": [{"id": "K1", "component_ids": ["C4"], "flow_ids": ["F1", "F2"], "fact": "Draft response tool applies document parsing and attachment handling before presenting content to the agent."}, {"id": "K2", "component_ids": ["C4", "C6"], "flow_ids": ["F5", "F6"], "fact": "Support workspace uses case-linked account lookups when the agent opens a customer record."}, {"id": "K3", "component_ids": ["C3", "C4"], "flow_ids": ["F1", "F2", "F5", "F6"], "fact": "The agent reviews retrieved material before sending a reply."}], "architecture": {"components": [{"id": "C1", "label": "Help document feed", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Evidence upload channel", "zone": "Z1", "kind": "external_entity"}, {"id": "C3", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Draft response tool", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Knowledge cache", "zone": "Z2", "kind": "data_store"}, {"id": "C6", "label": "Customer-account store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C4", "data": "help documents and articles", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C4", "data": "customer-provided evidence and attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C3", "target": "C4", "data": "reply instructions and edits", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "drafted response context and retrieved notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C6", "data": "account details and case context", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C6", "target": "C4", "data": "customer profile and account status", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support Workspace"}, {"id": "Z3", "label": "Account Data"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection via help documents", "component_ids": ["C1", "C4", "C3"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "agent decision quality and reply content", "threat_vector": "Attacker places malicious instructions in externally supplied help documents that are ingested by the draft response tool and then acted on by the support agent.", "preconditions": ["Attacker can influence the external help document content.", "The support agent relies on retrieved document text during response preparation."], "impact": "Wrong, unsafe, or policy-violating responses may be drafted based on attacker-authored instructions.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Draft response tool applies document parsing and attachment handling before presenting content to the agent.", "justification": "The path exists because external documents flow into the drafting tool and then to the agent, but the quoted parsing/review controls indicate some limiting handling before presentation, so the exposure is reduced rather than fully exposed."}, {"id": "T2", "category": "I", "subtype": "account-data overexposure in drafting", "component_ids": ["C6", "C4", "C3"], "flow_ids": ["F5", "F6"], "boundary_ids": ["B2"], "asset": "customer profile and account status", "threat_vector": "The drafting tool requests and displays more account data than needed for a reply, allowing the agent to see sensitive customer details.", "preconditions": ["The agent opens a customer record for support work.", "The account lookup returns more fields than the immediate task requires."], "impact": "Sensitive customer-account information may be disclosed to the support agent or retained in draft context unnecessarily.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "Support workspace uses case-linked account lookups when the agent opens a customer record.", "justification": "The architecture clearly includes account lookups into the drafting workspace, but the scope of fields, masking, and authorization checks are unspecified, so disclosure risk cannot be confirmed or ruled out."}, {"id": "T3", "category": "T", "subtype": "tampering with evidence inputs", "component_ids": ["C2", "C4", "C3"], "flow_ids": ["F2", "F3"], "boundary_ids": ["B1"], "asset": "support evidence integrity", "threat_vector": "An attacker supplies altered attachments or fabricated evidence that the draft response tool incorporates into the case context, changing the agent's response basis.", "preconditions": ["Attacker can upload or influence external evidence.", "The agent accepts uploaded materials as relevant case evidence."], "impact": "Responses may be based on falsified supporting material, leading to incorrect support actions or account changes.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Draft response tool applies document parsing and attachment handling before presenting content to the agent.", "justification": "The path from external evidence to the drafting tool is explicit, but the integrity checks on uploads are unspecified; parsing alone does not establish protection, so the path is an unknown-risk candidate."}], "scope_notes": ["Boundary B1 is the external-sources to support-workspace crossing for F1, F2.", "Boundary B2 is the support-workspace to account-data crossing for F5, F6.", "Control effectiveness is unspecified except where the facts explicitly say parsing, lookup usage, or agent review occur.", "Unsupported paths such as direct privileged database access are out of scope because no such component or flow is present."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C4", "C3"], "recommendation": "Add explicit prompt/content separation and a require-human-confirmation step in the draft response tool so externally sourced text is displayed as quoted reference material, not as instructions to execute.", "verification": "Use a benign test document containing conflicting guidance and confirm the tool preserves it as cited text while the agent must manually choose the response action.", "existing_control_effect": "Draft response tool applies document parsing and attachment handling before presenting content to the agent; the agent reviews retrieved material before sending a reply.", "residual_risk": "Residual risk remains that subtle instruction framing in external documents can still influence the agent's judgment."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C6"], "recommendation": "Restrict account lookups to case-minimum fields and mask nonessential identifiers in the draft response tool before data reaches the agent review screen.", "verification": "Open a normal support case with test records and confirm only the minimal required account fields appear in the draft context and review view.", "existing_control_effect": "Support workspace uses case-linked account lookups when the agent opens a customer record.", "residual_risk": "Residual risk remains if the case workflow still needs broader data for certain support tasks."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C4"], "recommendation": "Require upload provenance checks and attachment integrity validation at the evidence upload channel before materials enter draft-response processing.", "verification": "Upload a harmless sample attachment and verify the system records its source metadata and rejects altered copies with mismatched checksums or missing provenance.", "existing_control_effect": "Draft response tool applies document parsing and attachment handling before presenting content to the agent; the agent reviews retrieved material before sending a reply.", "residual_risk": "Residual risk remains for well-formed but misleading external evidence that passes technical integrity checks."}]}
{"id": "E046", "collection": "comparison", "domain": "incident_response", "control_profile": "guarded", "generation_condition": "single_prompt", "trial_id": "slot-15-single_prompt", "topology_family_id": "topology-018", "annotation_status": "candidate", "image_path": "images/comparison/E046.png", "title": "incident_response", "purpose": "Two analyst agents process externally supplied incident evidence and send findings through a coordinator to a case store.", "assumptions": ["Synthetic design; no actual production data.", "Attackers can influence externally supplied evidence content but have no privileged access to analysis systems or case store.", "Analyst agents are permitted to read evidence and submit drafted findings, but not to directly modify stored cases.", "The coordinator is the only component authorized to write to the case store and provenance registry."], "controls": [{"id": "CTRL1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Evidence intake validates source identity and attaches a submission identifier before routing to analysts."}, {"id": "CTRL2", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2", "F3", "F4"], "fact": "Analyst agents are configured to treat evidence text as untrusted and ignore embedded instructions that request tool use or policy overrides."}, {"id": "CTRL3", "component_ids": ["C4"], "flow_ids": ["F3", "F4", "F5", "F6"], "fact": "The coordinator only accepts findings signed by an analyst agent and records provenance metadata with each case update."}, {"id": "CTRL4", "component_ids": ["C5", "C6"], "flow_ids": ["F5", "F6"], "fact": "Case store and provenance registry writes are restricted to the coordinator service account."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake API", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Log Analyst Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Artifact Analyst Agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Findings Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Provenance Registry", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, attachments, and free-text evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, attachments, and free-text evidence", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "log analysis, extracted indicators, confidence notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "artifact review, extracted indicators, confidence notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case findings, triage summary, disposition updates", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C4", "target": "C6", "data": "source provenance, hashes, chain-of-custody metadata", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Analysis Workspace"}, {"id": "Z3", "label": "Case Management"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection through evidence", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3", "F5"], "boundary_ids": ["B1", "B2"], "asset": "case findings and triage disposition", "threat_vector": "Attacker-controlled evidence embeds instructions that try to steer analyst behavior and downstream findings.", "preconditions": ["Attacker can submit or influence evidence text reaching the analyst agents.", "Analyst output is not independently reviewed for prompt injection indicators before handoff."], "impact": "Misleading findings may be propagated into the case store, reducing the integrity of incident response decisions.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL3"], "evidence_quote": "Analyst agents are configured to treat evidence text as untrusted and ignore embedded instructions that request tool use or policy overrides.", "justification": "The path exists because external evidence reaches both analyst agents and their outputs reach the coordinator. The design explicitly limits instruction-following by the analysts and requires signed findings at the coordinator, so the exposure is reduced rather than fully open."}, {"id": "T2", "category": "I", "subtype": "provenance leakage in case updates", "component_ids": ["C4", "C5", "C6"], "flow_ids": ["F5", "F6"], "boundary_ids": ["B2"], "asset": "source provenance and chain-of-custody metadata", "threat_vector": "Stored provenance metadata may expose sensitive origin details to anyone able to read case records or downstream exports.", "preconditions": ["A reader or export path can access case records beyond the coordinator boundary.", "Provenance fields include source-identifying details rather than fully redacted summaries."], "impact": "Disclosure of incident source details can reveal investigative methods, affected parties, or internal handling patterns.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL3", "CTRL4"], "evidence_quote": "The coordinator only accepts findings signed by an analyst agent and records provenance metadata with each case update.", "justification": "The architecture stores provenance alongside findings, but the read-side access model for the case store is not specified. That makes disclosure plausible but not fully demonstrated."}], "scope_notes": ["Unsupported read access paths to the case store are not described; only coordinator writes are controlled.", "Whether signed findings prevent forged content from a compromised analyst agent is unspecified.", "No direct evidence export component is shown, so disclosure depends on downstream access not modeled here.", "Boundary IDs derived from the supplied flow order: F1/F2 cross Z1-Z2 as B1; F3/F4/F5/F6 cross Z2-Z3 as B2."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Keep analyst-side untrusted-text handling enabled, add a coordinator-side content policy check for injected tasking phrases, and require manual review for any finding that changes severity or attribution.", "verification": "Submit benign sample evidence containing fake control commands and confirm analysts treat it as inert text while the coordinator rejects unreviewed disposition changes.", "existing_control_effect": "Existing controls already reduce this path by marking evidence untrusted and requiring signed analyst findings.", "residual_risk": "Residual risk remains that subtle manipulation can still bias analyst interpretation even when explicit instructions are ignored."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5", "C6"], "recommendation": "Restrict case and provenance read access to the smallest operational group, minimize stored provenance fields to non-sensitive identifiers, and add separate redaction for any exported case views.", "verification": "Open a normal case workflow and confirm non-coordinator users cannot query raw provenance fields; verify redacted views omit source-identifying details.", "existing_control_effect": "Existing controls restrict writes to the coordinator, but they do not define who can read stored provenance.", "residual_risk": "Residual risk persists until read permissions and export behavior are explicitly defined and reviewed."}]}
{"id": "E047", "collection": "comparison", "domain": "incident_response", "control_profile": "weak", "generation_condition": "no_reviewer", "trial_id": "slot-14-no_reviewer", "topology_family_id": "topology-005", "annotation_status": "candidate", "image_path": "images/comparison/E047.png", "title": "Analyst-assisted incident evidence triage with coordinator and case store", "purpose": "Model a single incident-response workflow where two analyst agents review externally supplied evidence, forward findings through a coordinator, and persist case records in a shared store under weak trust controls.", "assumptions": ["External incident documents, attachments, and messages originate outside the trusted application zone and may be attacker-influenced.", "The two analyst agents are separate processing roles that both consume external evidence and can emit findings to a coordinator.", "The coordinator aggregates, routes, or normalizes findings before writing to the case store; the store is a shared persistence component within the trusted zone.", "Controls are weak or permissive: evidence is accepted from external sources with limited pre-validation, and the coordinator/tooling does not enforce strong content isolation or provenance checks.", "No privileged system access is assumed for the attacker beyond influence over external documents or evidence content."], "controls": [{"id": "K1", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "External evidence is accepted through a permissive intake path with limited pre-validation; analyst agents can directly open and process the supplied content."}, {"id": "K2", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4"], "fact": "Analyst outputs are forwarded to the coordinator without strong provenance enforcement or content isolation between the two analyst paths."}, {"id": "K3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "The coordinator is the only writer to the case store, but the write path is only weakly constrained by basic application logic rather than strict validation."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case Store", "zone": "Z2", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident documents, attachments, messages", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident documents, attachments, messages", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, summaries, extracted indicators", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, summaries, extracted indicators", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case records, normalized findings, updates", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Evidence Source"}, {"id": "Z2", "label": "Incident Response Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "prompt/content injection via external evidence", "component_ids": ["C1", "C2", "C4"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "analyst reasoning and downstream case records", "threat_vector": "Attacker supplies crafted incident documents/messages to Evidence Intake, which Analyst Agent A processes and forwards as findings to the Coordinator; malicious content can be carried into internal analysis and normalized outputs.", "preconditions": ["Attacker can influence external incident documents, attachments, or messages.", "The external content is processed by an analyst agent with limited pre-validation."], "impact": "Injected instructions or misleading content can alter analyst outputs and downstream case records, causing incorrect triage or contaminated case data.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "External evidence is accepted through a permissive intake path with limited pre-validation; analyst agents can directly open and process the supplied content.", "justification": "The architecture explicitly accepts attacker-influenced external content across B1 and processes it in analyst agents before forwarding results internally. Weak provenance and isolation make content manipulation of the analysis path plausible."}, {"id": "T2", "category": "I", "subtype": "cross-analyst information disclosure via shared coordination path", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F3", "F4"], "boundary_ids": [], "asset": "sensitive incident evidence and analyst findings", "threat_vector": "Findings from one analyst path may be exposed to the other path through the coordinator because there is no strong content isolation or provenance enforcement between the two analyst outputs.", "preconditions": ["One analyst path handles sensitive evidence or derived findings.", "The coordinator aggregates or routes results from both analysts."], "impact": "Sensitive details from one analyst's findings may be disclosed to the other analyst workflow or mixed into shared outputs beyond intended need-to-know boundaries.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "Analyst outputs are forwarded to the coordinator without strong provenance enforcement or content isolation between the two analyst paths.", "justification": "The design states two analyst agents emit findings to a common coordinator with weak isolation. That is a supported disclosure path within Z2, even though the exact recipient misuse is not enumerated."}, {"id": "T3", "category": "T", "subtype": "case record tampering through weak write validation", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": [], "asset": "case store records", "threat_vector": "Malformed, misleading, or attacker-shaped normalized findings reaching the coordinator can be written into the shared case store because the write path uses only basic application logic.", "preconditions": ["Attacker can influence evidence content that affects analyst findings.", "Coordinator accepts analyst findings and writes them to the store."], "impact": "Case records may be corrupted, overwritten, or populated with incorrect normalized findings, degrading the integrity of incident response records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "The coordinator is the only writer to the case store, but the write path is only weakly constrained by basic application logic rather than strict validation.", "justification": "A single internal writer does not remove tampering risk when the writer accepts weakly validated upstream content. The design explicitly leaves the store update path weakly constrained."}], "scope_notes": ["Boundary B1 is the only trust boundary and applies to F1 and F2 in the listed order.", "No direct external write to the case store is shown; store compromise is only inferred through the coordinator write path.", "No explicit authentication, authorization, or audit controls are described for the analyst-to-coordinator or coordinator-to-store paths.", "No evidence of denial-of-service, spoofing of identities, elevation of privilege, or repudiation controls is provided in the architecture facts.", "The assessment does not assume privileged system access beyond influence over external evidence content."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C4"], "recommendation": "Add strict evidence sanitization and instruction-stripping at intake (C1) before analysts open content, then enforce a fixed analyst prompt/template and provenance tag checks in C2 before any output is accepted by the coordinator C4; reject or quarantine evidence that contains executable directives, hidden markup, or untrusted embedded instructions.", "verification": "Submit benign sample evidence containing harmless embedded imperative phrases and confirm the intake layer quarantines or neutralizes them, the analyst prompt remains fixed, and C4 only accepts outputs carrying the expected provenance tag from C2.", "existing_control_effect": "K1 already provides a permissive intake path with limited pre-validation, so this control gives basic access but does not prevent content manipulation or prompt injection.", "residual_risk": "An attacker may still influence analyst interpretation through subtle wording or adversarially crafted benign-looking evidence, but the strongest direct instruction-carrier paths are reduced."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C3", "C4"], "recommendation": "Separate analyst result channels in C4 with per-source namespaces and mandatory provenance labeling, and prevent one analyst’s output from being forwarded to the other analyst workflow except through explicitly approved shared summaries that strip source-specific sensitive details.", "verification": "Use two benign test cases with distinct markers from C2 and C3 and confirm C4 keeps them in separate namespaces, does not echo one analyst’s detailed findings into the other path, and only emits approved summarized fields.", "existing_control_effect": "K2 already routes both analyst outputs to the coordinator, but without strong provenance enforcement or content isolation, so the current design only provides weak separation.", "residual_risk": "Some information may still be inferable from shared summaries or operator-visible case context, but unintended direct cross-analyst disclosure should be reduced."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Require C4 to validate normalized findings against a strict schema and allowed-field set before any write to C5, including explicit length/type checks, source provenance, and rejection of conflicting or overwrite-style updates unless matched to a trusted case workflow state.", "verification": "Send benign test updates with missing fields, duplicate keys, oversized values, and harmless conflicting edits, and confirm C4 rejects invalid writes while only schema-conformant updates reach C5.", "existing_control_effect": "K3 already places C4 as the only writer to C5, which limits direct write paths, but the write path is only weakly constrained and can still persist bad data.", "residual_risk": "A valid-looking but misleading analyst conclusion can still pass validation and be stored, so integrity depends on upstream analyst correctness as well as schema enforcement."}]}
{"id": "E049", "collection": "comparison", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "single_prompt", "trial_id": "slot-18-single_prompt", "topology_family_id": "topology-017", "annotation_status": "candidate", "image_path": "images/comparison/E049.png", "title": "incident_response", "purpose": "Two analyst agents process externally supplied incident evidence and send findings through a coordinator to a case store.", "assumptions": ["Synthetic design; no actual production data.", "The attacker can influence external documents or evidence, but has no privileged access.", "Relevant control implementation and effectiveness are unspecified."], "controls": [{"id": "C1", "component_ids": ["P1"], "flow_ids": ["F3", "F4", "F5"], "fact": "The coordinator records analyst findings before writing the consolidated case record to the case store."}, {"id": "C2", "component_ids": ["A2", "A3", "P1"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "fact": "Relevant control implementation and effectiveness are unspecified."}], "architecture": {"components": [{"id": "A1", "label": "Malware Evidence Ingestor", "zone": "Z1", "kind": "external_entity"}, {"id": "A2", "label": "Triage Analyst Agent", "zone": "Z2", "kind": "agent"}, {"id": "A3", "label": "Forensics Analyst Agent", "zone": "Z2", "kind": "agent"}, {"id": "P1", "label": "Finding Coordinator", "zone": "Z2", "kind": "process"}, {"id": "D1", "label": "Case Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "A1", "target": "A2", "data": "incident documents, artifacts, and embedded instructions", "boundary_ids": ["B1"]}, {"id": "F2", "source": "A1", "target": "A3", "data": "incident documents, artifacts, and embedded instructions", "boundary_ids": ["B1"]}, {"id": "F3", "source": "A2", "target": "P1", "data": "triage findings and recommended actions", "boundary_ids": []}, {"id": "F4", "source": "A3", "target": "P1", "data": "forensic findings, indicators, and confidence notes", "boundary_ids": []}, {"id": "F5", "source": "P1", "target": "D1", "data": "consolidated case record", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "External Source Zone"}, {"id": "Z2", "label": "Analysis Zone"}, {"id": "Z3", "label": "Case Management Zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection via external evidence", "component_ids": ["A2", "A3"], "flow_ids": ["F1", "F2"], "boundary_ids": ["B1"], "asset": "analyst reasoning and downstream case findings", "threat_vector": "Attacker places embedded instructions in supplied evidence to steer either analyst agent away from intended incident handling.", "preconditions": ["Attacker can submit or modify external incident documents before analyst review.", "Analysts treat some document content as actionable context."], "impact": "Incorrect or attacker-shaped findings may be forwarded to the coordinator and stored in the case record.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["C2"], "evidence_quote": "Relevant control implementation and effectiveness are unspecified.", "justification": "The architecture explicitly accepts externally supplied evidence into both analyst agents, and the control posture for handling that input is unspecified, so instruction-manipulation exposure is plausible but not confirmed as effective or blocked."}, {"id": "T2", "category": "I", "subtype": "sensitive content over-sharing in consolidated findings", "component_ids": ["P1", "D1"], "flow_ids": ["F5"], "boundary_ids": ["B2"], "asset": "case data and analyst findings", "threat_vector": "An attacker-supplied evidence packet could cause an analyst to include more sensitive incident details than needed in the consolidated record written to the case store.", "preconditions": ["The case store retains details from analyst submissions.", "The coordinator preserves what analysts send into the final record."], "impact": "Sensitive incident details may be retained in the case store and exposed to users with broader-than-needed read access.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["C1"], "evidence_quote": "The coordinator records analyst findings before writing the consolidated case record to the case store.", "justification": "The coordinator explicitly writes consolidated findings to the case store, but the exact data-minimization policy and access boundaries are not described, so disclosure risk is possible yet not demonstrated as exposed."}], "scope_notes": ["Boundary B1 covers the external source zone to analysis zone flows F1 and F2.", "Boundary B2 covers the analysis zone to case management zone flow F5.", "Unsupported paths remain: no explicit authentication, sanitization, or access-control mechanism is described for analyst inputs or the case store.", "Control effectiveness is unspecified, so reduced or blocked status cannot be claimed for the evidence-processing path.", "A data-store write alone does not prove unauthorized disclosure; the information-disclosure threat is about possible over-sharing into stored case records."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["A2", "A3", "P1"], "recommendation": "Add deterministic evidence normalization and instruction-stripping before analyst interpretation, and require the coordinator to accept only structured findings fields from the agents.", "verification": "Submit benign sample evidence containing harmless imperative text and confirm the analysts and coordinator preserve only evidence fields, not embedded instructions.", "existing_control_effect": "The coordinator already records analyst findings before writing the consolidated case record to the case store, but effectiveness against instruction manipulation is unspecified.", "residual_risk": "Residual risk remains for novel prompt-injection variants or analyst over-reliance on untrusted content."}, {"threat_id": "T2", "enforcement_component_ids": ["P1", "D1"], "recommendation": "Constrain the coordinator to redact unnecessary incident details before case-store write and store only the minimum required case fields.", "verification": "Review a non-sensitive test case and confirm the stored record omits optional raw evidence details while preserving required findings.", "existing_control_effect": "The coordinator records analyst findings before writing the consolidated case record to the case store, which supports a centralized point for redaction if configured.", "residual_risk": "Residual risk remains if analysts include sensitive details in required fields or if downstream readers have broader access than intended."}]}
{"id": "E050", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "single_prompt", "trial_id": "slot-01-single_prompt", "topology_family_id": "topology-038", "annotation_status": "candidate", "image_path": "images/comparison/E050.png", "title": "customer_support", "purpose": "A support agent uses externally supplied help documents and a customer-account store to prepare responses.", "assumptions": ["Synthetic design; no actual production data.", "External documents may be influenced by an attacker with ordinary submission capability.", "The support agent can access the drafting tool and customer account store for case handling.", "Document screening is permissive and does not guarantee malicious content is blocked."], "controls": [{"id": "CTRL1", "component_ids": ["C4"], "flow_ids": ["F2"], "fact": "The drafting tool previews retrieved documents before rendering them into the response workspace."}, {"id": "CTRL2", "component_ids": ["C5"], "flow_ids": ["F3"], "fact": "The customer account store requires authenticated access from the drafting tool."}, {"id": "CTRL3", "component_ids": ["C3", "C4"], "flow_ids": ["F2", "F4"], "fact": "The support agent reviews suggested responses before sending them to a customer."}], "architecture": {"components": [{"id": "C1", "label": "Customer Help Submitter", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Help Document Inbox", "zone": "Z1", "kind": "data_store"}, {"id": "C3", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Response Drafting Tool", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "help document content and attachments", "boundary_ids": []}, {"id": "F2", "source": "C2", "target": "C4", "data": "retrieved help documents", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C5", "target": "C4", "data": "customer account details", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C4", "target": "C3", "data": "draft response and suggested answer", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Support Workspace"}, {"id": "Z3", "label": "Customer Data"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "instruction injection via external help document", "component_ids": ["C1", "C2", "C4", "C3"], "flow_ids": ["F1", "F2", "F4"], "boundary_ids": ["B1"], "asset": "response quality and agent tool authority", "threat_vector": "Attacker supplies a help document that contains instructions intended to steer the drafting tool or the support agent into following attacker-provided actions or content.", "preconditions": ["Attacker can submit or influence help documents.", "The document is retrieved into the drafting workflow without strong content authentication or filtering."], "impact": "Incorrect or unsafe responses may be drafted, including misuse of support guidance or agent tool authority.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL3"], "evidence_quote": "The drafting tool previews retrieved documents before rendering them into the response workspace.", "justification": "The path crosses from External Sources (Z1) into Support Workspace (Z2) on F2, then to the agent on F4. The quoted control shows previewing and agent review, but not robust instruction isolation or rejection of malicious content, so the injection path remains supported and permissive."}, {"id": "T2", "category": "I", "subtype": "customer data exposure in draft handling", "component_ids": ["C5", "C4", "C3"], "flow_ids": ["F3", "F4"], "boundary_ids": ["B2"], "asset": "customer account details", "threat_vector": "Sensitive customer account details are delivered into the drafting tool and may be exposed in the prepared response or to the support agent beyond need-to-know.", "preconditions": ["The drafting tool receives account data for case handling.", "The response may echo or summarize account details."], "impact": "Customer information could be disclosed to unauthorized recipients if the drafted response or handling process over-shares data.", "assessment": "reduced", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL3"], "evidence_quote": "The customer account store requires authenticated access from the drafting tool.", "justification": "The path crosses from Customer Data (Z3) to Support Workspace (Z2) on F3. Authentication reduces unauthorized retrieval, and the agent review adds a check, but the architecture still sends account details into the drafting process, so disclosure risk is reduced rather than eliminated."}], "scope_notes": ["Boundary IDs were assigned from the flow order: F1 has no boundary, F2 crosses Z1->Z2 as B1, F3 crosses Z3->Z2 as B2, and F4 has no boundary.", "The architecture does not specify content sanitization, provenance checks, or explicit data minimization for help documents or customer records.", "No direct unauthorized access path to the customer account store is shown beyond authenticated drafting-tool access.", "Unsupported paths such as direct external access to customer data are not claimed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C4", "C3"], "recommendation": "Add strict document provenance checks and treat retrieved help text as untrusted input in the drafting tool; require the agent to ignore any document directives that attempt to alter tool behavior or case policy.", "verification": "Use a benign test document containing injected instructions and confirm the drafting tool labels it untrusted, strips directive-like text, and the agent review can reject it before response finalization.", "existing_control_effect": "The existing preview and agent review provide a partial checkpoint, but they do not guarantee instruction isolation.", "residual_risk": "Residual risk remains if malicious content is subtle, if the agent follows embedded instructions, or if document screening fails to detect the payload."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Limit the drafting tool to the minimum customer fields needed for the case and mask or omit unnecessary account attributes before rendering the draft response.", "verification": "Prepare a harmless test case with extra nonessential account fields and confirm the drafting output only includes the approved minimum data set.", "existing_control_effect": "Authenticated access to the customer store reduces unauthorized retrieval, and agent review can catch obvious over-sharing.", "residual_risk": "Residual risk remains if required support content still contains sensitive fields or if the draft is forwarded without sufficient redaction."}]}
{"id": "E051", "collection": "comparison", "domain": "customer_support", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-02-full_workflow", "topology_family_id": "topology-039", "annotation_status": "candidate", "image_path": "images/comparison/E051.png", "title": "AI-assisted customer support response preparation", "purpose": "Draft customer support replies by combining externally supplied help materials with customer-account records, under weakly controlled agent/tool handling.", "assumptions": ["Customer messages and uploaded help documents arrive from an untrusted external zone.", "The support agent can query a customer-account store through a tool with read access only.", "Weak control facts: document content is not strongly authenticated or sanitized before agent consumption; prompt/tool-instruction separation is permissive.", "The architecture is one connected workflow with no extra unrelated services.", "Attacker has no privileged internal access, but can influence external documents or evidence presented to the agent."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F2"], "fact": "Uploaded help documents are ingested directly into the assistant context with only basic formatting checks; no strong authenticity gate is present."}, {"id": "K2", "component_ids": ["C3", "C4", "C5"], "flow_ids": ["F3", "F4", "F5", "F6"], "fact": "The account lookup tool is configured for read-only access to the customer-account store, but request content is not tightly separated from assistant instructions."}, {"id": "K3", "component_ids": ["C3", "C6"], "flow_ids": ["F7", "F8"], "fact": "A human reviewer sees the drafted reply before sending, but review is based on the assistant output rather than the original external evidence set."}], "architecture": {"components": [{"id": "C1", "label": "Customer message intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Uploaded help documents", "zone": "Z1", "kind": "external_entity"}, {"id": "C3", "label": "Support agent", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Account lookup tool", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Customer-account store", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "Draft reply review", "zone": "Z4", "kind": "process"}], "flows": [{"id": "F1", "source": "C1", "target": "C3", "data": "customer question and case details", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "help article text and screenshots", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C3", "target": "C4", "data": "account lookup request", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C5", "data": "read-only customer record query", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C4", "data": "account status, history, and notes", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C4", "target": "C3", "data": "retrieved account facts", "boundary_ids": []}, {"id": "F7", "source": "C3", "target": "C6", "data": "draft response for human review", "boundary_ids": ["B3"]}, {"id": "F8", "source": "C6", "target": "C3", "data": "review feedback and approval or edits", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "Public / customer input"}, {"id": "Z2", "label": "Support assistant"}, {"id": "Z3", "label": "Account data"}, {"id": "Z4", "label": "Human review"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/instruction injection via uploaded help documents", "component_ids": ["C2", "C3", "C6"], "flow_ids": ["F2", "F7", "F8"], "boundary_ids": ["B1", "B3"], "asset": "Support-agent drafting context and reply quality", "threat_vector": "Attacker supplies crafted help document text/screenshots that are ingested into the assistant context and can steer drafting behavior.", "preconditions": ["Attacker can influence uploaded help documents from the untrusted external zone.", "The assistant consumes document content with weak instruction separation."], "impact": "Misleading or attacker-directed draft replies; downstream human review may validate a compromised draft.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Uploaded help documents are ingested directly into the assistant context with only basic formatting checks; no strong authenticity gate is present.", "justification": "The architecture explicitly places untrusted documents into the assistant context and notes permissive prompt/tool-instruction separation. That supports a feasible agent-specific injection path affecting drafted replies, with the reviewer only seeing the assistant output."}, {"id": "T2", "category": "I", "subtype": "account-data disclosure through draft reply content", "component_ids": ["C4", "C5", "C6"], "flow_ids": ["F4", "F5", "F6", "F7", "F8"], "boundary_ids": ["B2", "B3"], "asset": "Customer-account record contents", "threat_vector": "The agent retrieves account status, history, and notes and may place sensitive details into a draft reply that is then reviewed.", "preconditions": ["The support agent queries the customer-account store for a case.", "The drafted reply includes retrieved account facts.", "The external recipient or reviewer is not intended to see all raw account details."], "impact": "Sensitive customer-account information may be revealed in the drafted response or review trail.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The account lookup tool is configured for read-only access to the customer-account store, but request content is not tightly separated from assistant instructions.", "justification": "Read-only access limits modification, not disclosure. The architecture shows a path from stored account facts into the draft reply, but whether the eventual recipient receives sensitive data is not fully specified, so this is a candidate disclosure risk with unknown effectiveness rather than a confirmed exposed one."}], "scope_notes": ["No unsupported write-path or data tampering path to the customer-account store is shown; the tool is read-only.", "The architecture does not specify the content of the final sent reply, only the draft and human review loop, so external disclosure impact is incomplete.", "No direct evidence of spoofing identities, repudiation logging gaps, denial-of-service, or privilege escalation is present in the supplied flows.", "Boundary IDs were assigned from the listed cross-zone flows: B1 for Z1-Z2, B2 for Z2-Z3, and B3 for Z2-Z4."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C6"], "recommendation": "Add a strict content-isolation gate on C2 before C3 consumes uploaded help documents: verify document provenance, strip or neutralize instruction-like text, and pass only bounded excerpts to the drafting context. Require C6 review to inspect the original uploaded evidence alongside the draft, not the assistant output alone.", "verification": "Use a benign test document containing harmless instruction-like phrases and confirm C3 treats them as data only, while C6 shows the original document and the draft side by side for review.", "existing_control_effect": "K1 already provides direct ingestion controls only at formatting level, and K3 adds human review; these reduce but do not prevent prompt injection because authenticity and instruction separation remain weak.", "residual_risk": "Some manipulation risk remains if attacker-crafted content still resembles legitimate support material or if reviewers approve a compromised draft based on plausible-looking evidence."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5", "C6"], "recommendation": "Constrain C4 to return the minimum necessary account fields for the case, apply field-level redaction before data reaches C3, and have C6 review the redacted draft against a case-specific disclosure policy.", "verification": "Run a non-sensitive test query and confirm C4 only returns approved fields, sensitive notes are masked before drafting, and C6 blocks any draft that includes disallowed account details.", "existing_control_effect": "K2 limits the tool to read-only access, which prevents unauthorized writes but does not by itself stop over-broad reads or accidental inclusion of sensitive fields; K3 adds a review checkpoint on the draft.", "residual_risk": "Some disclosure risk remains if the retrieved facts are inherently sensitive for the case or if policy gaps allow an approved but overly detailed response."}]}
{"id": "E052", "collection": "comparison", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-17-no_reviewer", "topology_family_id": "topology-040", "annotation_status": "candidate", "image_path": "images/comparison/E052.png", "title": "Incident Evidence Triage with Dual Analyst Agents and Coordinator", "purpose": "Model a connected incident-response workflow where two analyst agents independently inspect externally supplied evidence, then submit findings through a coordinator into a case store for later review.", "assumptions": ["Externally supplied incident evidence can include attacker-influenced documents, logs, or attachments.", "Two analyst agents are distinct components but may share the same evidence source and coordination channel.", "The coordinator routes findings into the case store and may reconcile or normalize submissions before persistence.", "Relevant control implementation and effectiveness are unspecified; no assumption is made that filters, sandboxing, approval gates, or access controls are absent or weak.", "Attacker capability is limited to influencing the contents of external evidence; the attacker has no privileged access to analyst or coordinator components."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1"], "fact": "Analyst agents consume externally supplied evidence from the evidence source; control effectiveness for content filtering or sandboxing is unspecified."}, {"id": "K2", "component_ids": ["C4"], "flow_ids": ["F3", "F4", "F5"], "fact": "The coordinator is the only path into the case store and may normalize or reconcile analyst submissions before persistence."}, {"id": "K3", "component_ids": ["C5"], "flow_ids": ["F5"], "fact": "Case records are written only from the coordinator path; direct analyst-to-store writes are not modeled in this architecture."}], "architecture": {"components": [{"id": "C1", "label": "Evidence Ingest", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst Agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst Agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Triage Coordinator", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "Case Repository", "zone": "Z4", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident evidence, attachments, logs", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident evidence, attachments, logs", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "analyst findings, extracted indicators, notes", "boundary_ids": ["B2"]}, {"id": "F4", "source": "C3", "target": "C4", "data": "analyst findings, extracted indicators, notes", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C4", "target": "C5", "data": "normalized case record, merged findings, disposition", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External Evidence Source"}, {"id": "Z2", "label": "Analyst Workspace"}, {"id": "Z3", "label": "Coordination Zone"}, {"id": "Z4", "label": "Case Store"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z3", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/context injection via evidence", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "Analyst Agent A reasoning and downstream triage output", "threat_vector": "Attacker-influenced evidence content is delivered from the external source to Analyst Agent A, allowing malicious instructions or context to alter the agent’s interpretation of evidence.", "preconditions": ["Attacker can influence the contents of externally supplied incident evidence."], "impact": "Unsafe or biased analyst output can propagate into coordinator input, potentially distorting case disposition and stored records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Analyst agents consume externally supplied evidence from the evidence source; control effectiveness for content filtering or sandboxing is unspecified.", "justification": "The architecture explicitly routes attacker-influenced evidence into the analyst agent across B1, and no effective filtering/sandboxing is specified. This supports an A-class agent manipulation threat on the analyst reasoning path."}, {"id": "T2", "category": "A", "subtype": "prompt/context injection via evidence", "component_ids": ["C1", "C3"], "flow_ids": ["F2"], "boundary_ids": ["B1"], "asset": "Analyst Agent B reasoning and downstream triage output", "threat_vector": "Attacker-influenced evidence content is delivered from the external source to Analyst Agent B, allowing malicious instructions or context to alter the agent’s interpretation of evidence.", "preconditions": ["Attacker can influence the contents of externally supplied incident evidence."], "impact": "Unsafe or biased analyst output can propagate into coordinator input, potentially distorting case disposition and stored records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Analyst agents consume externally supplied evidence from the evidence source; control effectiveness for content filtering or sandboxing is unspecified.", "justification": "The same externally supplied evidence path exists to Analyst Agent B across B1, with unspecified content controls. This is a distinct candidate because it targets a separate agent instance and can independently affect downstream triage."}, {"id": "T3", "category": "T", "subtype": "evidence tampering impacts case content", "component_ids": ["C1", "C4", "C5"], "flow_ids": ["F1", "F3", "F4", "F5"], "boundary_ids": ["B1", "B2", "B3"], "asset": "Case repository records", "threat_vector": "Maliciously crafted evidence can induce the analyst agents to produce altered or misleading findings that the coordinator normalizes and persists into the case store.", "preconditions": ["Attacker can influence evidence contents.", "Coordinator accepts and persists analyst submissions through the modeled path."], "impact": "Stored case records may be inaccurate or corrupted, affecting later review and response decisions.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The coordinator is the only path into the case store and may normalize or reconcile analyst submissions before persistence.", "justification": "The architecture shows a write path from coordinator to case store, but it is unspecified how well reconciliation detects maliciously induced distortions from upstream evidence. Because the control effect is not defined, the path is a tampering concern but not clearly exposed."}], "scope_notes": ["Unsupported direct write paths from analysts to the case store are not modeled; only the coordinator writes to C5.", "No explicit disclosure path from the case store to external parties is shown, so information disclosure is not asserted here.", "No direct analyst-to-coordinator trust bypass or privilege escalation mechanism is specified.", "Control effectiveness for filtering, sandboxing, approval gates, and access controls is unspecified, so reduced-vs-exposed cannot be inferred absent explicit design facts."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C1"], "recommendation": "Place a strict evidence-ingestion guard on C2 that treats all F1 content as untrusted: strip or quarantine instruction-like text, block tool/action directives embedded in attachments, and require C2 to operate only on normalized evidence fields produced by C1-side preprocessing or a sandboxed parser.", "verification": "Submit benign test evidence containing harmless instruction-shaped text and confirm C2 records it as data only, does not execute or follow it, and produces the same triage result after sanitization as it does on the clean version.", "existing_control_effect": "K1 already indicates C2 consumes external evidence; however, filtering/sandboxing effectiveness is unspecified, so current controls may limit exposure but do not clearly prevent instruction injection.", "residual_risk": "A determined attacker may still influence analyst interpretation through subtle framing, ambiguous artifacts, or content that survives sanitization."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C1"], "recommendation": "Apply the same untrusted-evidence handling at C3 for F2, with a dedicated parser/sanitizer at the evidence source boundary and a policy that C3 ignores any evidence-contained instructions, task overrides, or role prompts before analysis.", "verification": "Use non-malicious sample evidence with embedded fake directives and confirm C3 treats them as inert text, logs the sanitization step, and only uses allowed evidence fields in its findings.", "existing_control_effect": "K1 already covers C3 receiving external evidence, but the architecture does not specify effective filtering or sandboxing, so protection is not assured by design alone.", "residual_risk": "Residual influence remains possible through crafted evidence that alters analyst judgment without relying on explicit instruction injection."}, {"threat_id": "T3", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Enforce coordinator-side validation and provenance checks on F3/F4 before F5 persistence: require source-linked citations for merged findings, reject conflicting or unsupported analyst claims, and store the original analyst submissions alongside the normalized case record for later review.", "verification": "Run benign cases with consistent analyst submissions and confirm C4 preserves provenance, rejects malformed fields, and writes only validated records to C5 while keeping traceable originals.", "existing_control_effect": "K2 and K3 already constrain writes to C5 through C4 and allow normalization/reconciliation, which can reduce accidental corruption and limit direct store abuse.", "residual_risk": "If both analysts are influenced in similar ways or the coordinator’s reconciliation logic is weak, misleading but well-formed content can still be persisted."}]}
{"id": "E053", "collection": "comparison", "domain": "incident_response", "control_profile": "weak", "generation_condition": "full_workflow", "trial_id": "slot-13-full_workflow", "topology_family_id": "topology-005", "annotation_status": "candidate", "image_path": "images/comparison/E053.png", "title": "Incident evidence triage with two analyst agents and permissive coordinator workflow", "purpose": "Model a small incident-response workflow where two analyst agents review attacker-influenced evidence, pass findings to a coordinator, and the coordinator writes case records to a shared store under weak access control.", "assumptions": ["External incident documents, logs, or attachments can be attacker-influenced before ingestion.", "The two analyst agents operate in a trusted application zone but consume untrusted evidence from an external intake zone.", "The coordinator accepts analyst findings and can invoke a case-store write action with permissive checks; no strong approval gate is stated.", "The case store is reachable only from the coordinator and is shared within the internal incident-response zone.", "No privileged insider access is assumed; attacker influence is limited to external evidence content and indirect prompt/context effects."], "controls": [{"id": "K1", "component_ids": ["C2", "C3"], "flow_ids": ["F1", "F2"], "fact": "Analyst agents accept external evidence directly from the intake zone; content validation is present but unspecified and not described as strong."}, {"id": "K2", "component_ids": ["C4"], "flow_ids": ["F3", "F4"], "fact": "Coordinator aggregates both analyst outputs before acting; cross-checking is informal and may be overridden by workflow discretion."}, {"id": "K3", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "fact": "Case-store writes are limited to the coordinator, but the write gate is permissive and does not require a separate approval step."}, {"id": "K4", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "fact": "Evidence provenance is treated as untrusted at intake, while downstream handling relies on analyst judgment and coordinator review rather than enforced policy checks."}], "architecture": {"components": [{"id": "C1", "label": "Evidence intake", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Analyst agent A", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Analyst agent B", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "Coordinator", "zone": "Z2", "kind": "process"}, {"id": "C5", "label": "Case store", "zone": "Z2", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "incident evidence, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C1", "target": "C3", "data": "incident evidence, logs, attachments", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C2", "target": "C4", "data": "triage findings, extracted indicators, confidence notes", "boundary_ids": []}, {"id": "F4", "source": "C3", "target": "C4", "data": "triage findings, extracted indicators, confidence notes", "boundary_ids": []}, {"id": "F5", "source": "C4", "target": "C5", "data": "case records, status updates, analyst summaries", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "External intake"}, {"id": "Z2", "label": "Incident response"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "prompt/context manipulation via external evidence", "component_ids": ["C1", "C2", "C4", "C5"], "flow_ids": ["F1", "F3", "F5"], "boundary_ids": ["B1"], "asset": "triage findings and case records", "threat_vector": "Attacker-influenced evidence content injects misleading instructions or context into Analyst agent A, which can shape findings sent to the Coordinator and then be written into the case store.", "preconditions": ["Attacker can influence the external evidence, logs, or attachments ingested on F1.", "Analyst agents consume this content as part of their reasoning.", "Coordinator can write analyst-derived records with permissive checks."], "impact": "Unsafe agent behavior or decision manipulation can contaminate triage outputs and persist incorrect case records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3", "K4"], "evidence_quote": "Analyst agents accept external evidence directly from the intake zone; content validation is present but unspecified and not described as strong.", "justification": "The architecture explicitly allows untrusted external evidence to reach the analyst agent, and downstream handling relies on judgment rather than enforced policy. This supports a feasible agent-specific manipulation path from intake to coordinator to stored case data."}, {"id": "T2", "category": "I", "subtype": "case record disclosure via shared internal store", "component_ids": ["C4", "C5"], "flow_ids": ["F5"], "boundary_ids": [], "asset": "case records and analyst summaries", "threat_vector": "Sensitive incident summaries written by the Coordinator into the shared case store may be readable by internal parties with access to that store.", "preconditions": ["A party with access to the internal incident-response zone can read the shared store.", "The stored records contain sensitive incident details."], "impact": "Confidential incident details could be exposed to unauthorized internal recipients.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "Case-store writes are limited to the coordinator, but the write gate is permissive and does not require a separate approval step.", "justification": "The design shows storage of case records in a shared internal store, but it does not state who can read that store. A write path alone does not demonstrate disclosure to an unauthorized recipient, so exposure is not established."}, {"id": "T3", "category": "D", "subtype": "coordinator workflow saturation", "component_ids": ["C1", "C2", "C3", "C4"], "flow_ids": ["F1", "F2", "F3", "F4", "F5"], "boundary_ids": ["B1"], "asset": "incident response workflow availability", "threat_vector": "Large or complex attacker-influenced evidence can drive excessive analyst processing and coordinator handling, slowing triage and case updates.", "preconditions": ["Attacker can submit oversized, repetitive, or highly complex evidence through the intake path.", "The workflow processes incoming evidence and analyst outputs without a stated rate limit or hard approval gate."], "impact": "Triage throughput may be reduced or delayed.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K4"], "evidence_quote": "Coordinator aggregates both analyst outputs before acting; cross-checking is informal and may be overridden by workflow discretion.", "justification": "The workflow suggests potential load amplification from external evidence into two analysts and then the coordinator, but no explicit rate-limit or exhaustion mechanism is described. Availability impact is plausible but not sufficiently specified to mark exposed."}], "scope_notes": ["Boundary B1 applies to flows between Z1 and Z2, so F1 and F2 cross the trust boundary while F3-F5 do not.", "The shared case store read model is not described; unauthorized internal read access is unspecified.", "No explicit integrity check, sanitization, or approval gate is described for analyst-to-coordinator outputs beyond informal review.", "No evidence of privileged insider access or direct write access to the case store from outside the coordinator is provided."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C3", "C4", "C5"], "recommendation": "Add strict evidence sanitization and instruction-stripping at Analyst agents C2/C3, enforce a fixed schema for extracted findings, and require Coordinator C4 to reject any analyst output that contains non-schema control text before writing to Case store C5.", "verification": "Submit benign test attachments that contain mock instructions and confirm C2/C3 only emit schema-conformant fields, C4 rejects non-conformant content, and no raw attacker text reaches C5.", "existing_control_effect": "K1 and K4 already indicate untrusted intake handling and some validation, and K2 provides a second review step, but these controls are informal and insufficient to prevent prompt/context manipulation from propagating.", "residual_risk": "Residual risk remains from sophisticated obfuscation in evidence content and from analyst reasoning errors, but the main injection path should be reduced if schema enforcement and rejection are implemented."}, {"threat_id": "T2", "enforcement_component_ids": ["C4", "C5"], "recommendation": "Restrict Case store C5 read access to explicitly authorized incident-response roles, log every read request, and require Coordinator C4 to label records with sensitivity tags so stored incident details are only exposed to approved consumers.", "verification": "Perform a benign access test from an unauthorized internal identity and confirm C5 denies reads while authorized coordinator-mediated reads succeed and are logged with the sensitivity label.", "existing_control_effect": "K3 limits writes to C4, but it does not define read restrictions; therefore it helps protect integrity of storage access but does not address disclosure to internal readers.", "residual_risk": "If additional internal consumers legitimately need access, some disclosure risk remains from overly broad authorization or mislabeling, so periodic access review is still needed."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "C2", "C3", "C4"], "recommendation": "Add intake rate limits and size caps at C1, impose bounded parsing and processing budgets at C2/C3, and make C4 require a timeout-safe completion threshold before aggregating or persisting results.", "verification": "Use benign oversized and repetitive sample submissions to confirm the intake rejects or truncates them, analyst processing stays within defined limits, and C4 does not block indefinitely on malformed batches.", "existing_control_effect": "K1, K2, and K4 provide some review and aggregation structure, but they do not state hard resource limits or exhaustion protections, so they only partially slow abuse.", "residual_risk": "Residual availability risk remains from moderate-volume abuse and from legitimate but unusually complex cases, but bounded processing should reduce saturation and delay."}]}
{"id": "E054", "collection": "comparison", "domain": "customer_support", "control_profile": "unspecified", "generation_condition": "no_reviewer", "trial_id": "slot-06-no_reviewer", "topology_family_id": "topology-041", "annotation_status": "candidate", "image_path": "images/comparison/E054.png", "title": "Support Agent with External Docs and Account Lookup", "purpose": "A customer-support assistant drafts replies by consulting externally supplied help documents and a customer-account store, then returns a response to an operator or customer-facing channel.", "assumptions": ["The support agent is an AI-assisted workflow component inside the trusted application zone.", "Externally supplied help documents and evidence can be influenced by an attacker, but they originate outside the trusted zone.", "The customer-account store is inside the trusted data zone and is reachable only through the support agent's normal lookup path.", "Control implementation and effectiveness are unspecified; no assumption is made that filtering, authorization, or prompt-injection defenses are absent or weak.", "The architecture is a single connected functional flow with no extra padding components."], "controls": [{"id": "CTRL1", "component_ids": ["C1", "C2"], "flow_ids": ["F1"], "fact": "External documents are ingested through the normal document input path; any content screening or instruction filtering is unspecified."}, {"id": "CTRL2", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F2", "F3", "F4", "F5"], "fact": "Account access occurs only via the lookup service and its ordinary query path to the store; authorization and validation details are unspecified."}, {"id": "CTRL3", "component_ids": ["C2", "C5"], "flow_ids": ["F6"], "fact": "The support agent outputs a single drafted reply to the reply channel; downstream review or approval controls are unspecified."}], "architecture": {"components": [{"id": "C1", "label": "Help Document Feed", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "Support Agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "Account Lookup Service", "zone": "Z3", "kind": "process"}, {"id": "C4", "label": "Customer Account Store", "zone": "Z3", "kind": "data_store"}, {"id": "C5", "label": "Reply Channel", "zone": "Z4", "kind": "external_entity"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "help documents and evidence", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C3", "data": "account lookup request", "boundary_ids": ["B2"]}, {"id": "F3", "source": "C3", "target": "C4", "data": "customer record query", "boundary_ids": []}, {"id": "F4", "source": "C4", "target": "C3", "data": "customer account data", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C2", "data": "lookup results", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C2", "target": "C5", "data": "drafted support response", "boundary_ids": ["B3"]}], "trust_zones": [{"id": "Z1", "label": "External Sources"}, {"id": "Z2", "label": "Trusted Application"}, {"id": "Z3", "label": "Trusted Data"}, {"id": "Z4", "label": "Output Channel"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}, {"id": "B3", "zone_ids": ["Z2", "Z4"]}]}, "threats": [{"id": "T1", "category": "A", "subtype": "external doc prompt injection leading to unsafe reply drafting", "component_ids": ["C1", "C2"], "flow_ids": ["F1", "F6"], "boundary_ids": ["B1", "B3"], "asset": "support-agent reply drafting behavior", "threat_vector": "attacker-influenced help documents or evidence are delivered into the trusted support agent and can steer the generated response", "preconditions": ["Attacker can influence externally supplied help documents or evidence", "The agent uses ingested document content when drafting replies"], "impact": "Unsafe or manipulated support replies, including incorrect guidance or adversarial instructions being reflected to the output channel", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["CTRL1", "CTRL3"], "evidence_quote": "External documents are ingested through the normal document input path; any content screening or instruction filtering is unspecified.", "justification": "The architecture explicitly accepts external documents into the trusted application zone and then sends a drafted reply to the output channel. Because the content-screening and downstream review controls are unspecified, there is a feasible agent-specific manipulation path from F1 to F6."}, {"id": "T2", "category": "I", "subtype": "support response reveals account data", "component_ids": ["C3", "C4", "C2", "C5"], "flow_ids": ["F3", "F4", "F5", "F6"], "boundary_ids": ["B2", "B3"], "asset": "customer account data", "threat_vector": "lookup results containing customer account data are incorporated into the drafted support response and sent outside the trusted zone", "preconditions": ["The lookup service returns customer account data to the support agent", "The drafted response includes some or all of those results"], "impact": "Disclosure of customer account information to the reply channel recipient", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL2", "CTRL3"], "evidence_quote": "Account access occurs only via the lookup service and its ordinary query path to the store; authorization and validation details are unspecified.", "justification": "The design shows customer data moving from the trusted store to the support agent and then to the output channel. However, the architecture does not specify what data is selected for replies or who receives the reply, so disclosure is possible but not demonstrated as an unauthorized exposure."}, {"id": "T3", "category": "S", "subtype": "spoofed account lookup context", "component_ids": ["C2", "C3"], "flow_ids": ["F2", "F5"], "boundary_ids": ["B2"], "asset": "account lookup decision context", "threat_vector": "a malicious or forged lookup request can cause the service to return results for the wrong customer context", "preconditions": ["Attacker can influence the lookup request content or context", "The lookup service relies on request context for retrieval"], "impact": "Responses may be attributed to the wrong account, enabling impersonation or incorrect customer handling", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["CTRL2"], "evidence_quote": "Account access occurs only via the lookup service and its ordinary query path to the store; authorization and validation details are unspecified.", "justification": "There is a clear request/response path across the trusted-application to trusted-data boundary, but the request authentication and validation behavior are not specified. That leaves a plausible spoofing path, though the design facts do not show it is weak enough to call exposed."}], "scope_notes": ["Boundary IDs are assigned from the listed flows: B1 for Z1↔Z2, B2 for Z2↔Z3, and B3 for Z2↔Z4.", "No unsupported direct path from external sources to the customer-account store is shown; F3 and F4 remain within Z3.", "The architecture does not specify whether reply recipients are customers, operators, or both, so disclosure and repudiation consequences are incomplete.", "Control effectiveness is unspecified for content screening, authorization, validation, and downstream review; missing detail prevents upgrading several paths from unknown to exposed."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C5"], "recommendation": "Add mandatory document-content sanitization and agent-side instruction isolation at C2 before drafting, and require a downstream approval gate on F6 so externally supplied text cannot directly steer or leave the reply channel without review.", "verification": "Provide a benign test document containing obvious meta-instructions and confirm C2 treats it only as reference content; then check that a draft on F6 is blocked or flagged until the approval step records release.", "existing_control_effect": "CTRL1 already routes external documents through the normal input path, and CTRL3 shows replies leave only through the reply channel, but neither control states that prompt-injection screening or human review is enforced.", "residual_risk": "Attacker-supplied documents may still influence wording or priorities if sanitization misses subtle manipulations, but direct unchecked propagation to the output channel is reduced."}, {"threat_id": "T2", "enforcement_component_ids": ["C3", "C4", "C5"], "recommendation": "Minimize lookup results in C3 to the least necessary fields, redact or suppress sensitive attributes before F5, and enforce output checks at C5 so only approved response templates or masked data can be sent outside Z2.", "verification": "Use a benign account record with non-sensitive and sensitive fields, request a normal lookup, and confirm the drafted reply omits sensitive fields while still answering the support question.", "existing_control_effect": "CTRL2 keeps account access on the normal lookup path through C3 and C4, which constrains where data can be read from, but it does not specify field-level authorization or redaction; CTRL3 only indicates a single drafted reply is sent out.", "residual_risk": "Some account details may still be inferred from allowed lookup results or from necessary support context, especially if the reply channel ultimately reaches a customer."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3"], "recommendation": "Bind each lookup request in F2 to a verified support-session identity and customer context at C2, and have C3 reject requests whose subject/customer binding, authorization token, or session state does not match the expected account.", "verification": "Submit a benign lookup using a valid session and then repeat with a mismatched customer context; confirm the second request is rejected and no results return on F5.", "existing_control_effect": "CTRL2 already routes all access through the lookup service and its ordinary query path, which is a suitable enforcement point, but the design does not say that request binding or authorization checks are implemented.", "residual_risk": "A forged or confused lookup request may still succeed if session binding is incomplete or if upstream identity data is stale, but the normal request path is the right place to enforce prevention."}]}