| {"id": "DEV-diagram-0001", "collection": "development", "domain": "customer_support", "control_profile": "weak", "generation_condition": "development_workflow", "trial_id": "", "topology_family_id": "topology-042", "annotation_status": "candidate", "image_path": "images/development/DEV-diagram-0001.png", "title": "Retrieval-assisted support agent with account lookup", "purpose": "A support assistant answers customer questions by combining externally supplied help documents with customer account records, then drafts replies for a support channel.", "assumptions": ["The agent can read a document corpus supplied from outside the support team.", "The agent can query customer account records across a separate trust zone.", "The control profile is weak: access checks, citation validation, and output review are limited or inconsistently enforced.", "A human agent may review some replies, but mandatory approval is not assumed for every response.", "Attacker capability is limited to sending crafted support messages and influencing the externally supplied document corpus; no privileged internal access is assumed."], "controls": [{"id": "K1", "component_ids": ["C2", "C4"], "flow_ids": ["F2", "F3"], "fact": "Document retrieval is enabled with basic query filtering, but document provenance checks are weak and the agent may ingest retrieved text directly into the draft."}, {"id": "K2", "component_ids": ["C2", "C5"], "flow_ids": ["F4", "F5"], "fact": "Account lookup is restricted to the current customer context, but the access gate is inconsistently enforced and relies on the agent to request the correct record."}, {"id": "K3", "component_ids": ["C3", "C6"], "flow_ids": ["F6", "F8"], "fact": "Replies can be copied to a human reviewer, but review is optional and not all drafts are checked before sending."}, {"id": "K4", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F6", "F7"], "fact": "The support channel accepts plain text prompts and plain text replies; no separate content sanitization step is shown in the architecture."}], "architecture": {"components": [{"id": "C1", "label": "user", "zone": "Z1", "kind": "external_entity"}, {"id": "C2", "label": "agent", "zone": "Z1", "kind": "agent"}, {"id": "C3", "label": "reply", "zone": "Z1", "kind": "process"}, {"id": "C4", "label": "docs", "zone": "Z2", "kind": "data_store"}, {"id": "C5", "label": "acct", "zone": "Z3", "kind": "data_store"}, {"id": "C6", "label": "reviewer", "zone": "Z1", "kind": "agent"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "support question", "boundary_ids": []}, {"id": "F2", "source": "C2", "target": "C4", "data": "doc query", "boundary_ids": ["B1"]}, {"id": "F3", "source": "C4", "target": "C2", "data": "retrieved text", "boundary_ids": ["B1"]}, {"id": "F4", "source": "C2", "target": "C5", "data": "account lookup", "boundary_ids": ["B2"]}, {"id": "F5", "source": "C5", "target": "C2", "data": "account details", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C2", "target": "C3", "data": "draft reply", "boundary_ids": []}, {"id": "F7", "source": "C3", "target": "C1", "data": "support response", "boundary_ids": []}, {"id": "F8", "source": "C3", "target": "C6", "data": "optional review copy", "boundary_ids": []}], "trust_zones": [{"id": "Z1", "label": "Support zone"}, {"id": "Z2", "label": "Docs zone"}, {"id": "Z3", "label": "Account zone"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z1", "Z3"]}]}, "threats": [{"id": "T1", "category": "I", "subtype": "prompt injection via externally supplied documents causing data disclosure", "component_ids": ["C2", "C4", "C3"], "flow_ids": ["F2", "F3", "F6", "F7"], "boundary_ids": ["B1"], "asset": "retrieved help content and downstream response content", "threat_vector": "Attacker influences the external document corpus so retrieved text contains instructions that cause the agent to reveal sensitive account details or other internal context in the drafted reply.", "preconditions": ["Attacker can influence the externally supplied document corpus.", "The agent ingests retrieved text directly into the draft.", "Review is optional and not all drafts are checked before sending."], "impact": "Sensitive account or internal information may be disclosed to the customer through the support response.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3"], "evidence_quote": "Document retrieval is enabled with basic query filtering, but document provenance checks are weak and the agent may ingest retrieved text directly into the draft.", "justification": "The path crosses the docs trust boundary and the control explicitly says provenance checks are weak and retrieved text may be ingested directly. Optional review does not reliably block disclosure."}, {"id": "T2", "category": "I", "subtype": "account data exposure through incorrect record lookup", "component_ids": ["C2", "C5", "C3"], "flow_ids": ["F4", "F5", "F6", "F7"], "boundary_ids": ["B2"], "asset": "customer account records", "threat_vector": "A crafted support message induces the agent to request or reuse the wrong account record, exposing account details in the drafted response.", "preconditions": ["Attacker can send crafted support messages.", "The access gate is inconsistently enforced and relies on the agent to request the correct record.", "Mandatory approval is not assumed for every response."], "impact": "Account details from an unintended customer record may be returned in the support response.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "Account lookup is restricted to the current customer context, but the access gate is inconsistently enforced and relies on the agent to request the correct record.", "justification": "This is a direct boundary-crossing lookup into the account zone, and the stated inconsistency makes unintended disclosure a supported exposure. Optional review leaves some responses unchecked."}, {"id": "T3", "category": "T", "subtype": "tampering with draft reply content", "component_ids": ["C2", "C3", "C6"], "flow_ids": ["F6", "F8", "F7"], "boundary_ids": [], "asset": "support reply integrity", "threat_vector": "Influenced retrieved text or crafted prompts alter the draft reply before it is sent to the customer or copied for review.", "preconditions": ["The agent may ingest retrieved text directly into the draft.", "The support channel accepts plain text prompts and plain text replies.", "Review is optional and not all drafts are checked before sending."], "impact": "The final customer-facing response may contain attacker-influenced instructions, incorrect guidance, or manipulated wording.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K3", "K4"], "evidence_quote": "The support channel accepts plain text prompts and plain text replies; no separate content sanitization step is shown in the architecture.", "justification": "Plain-text handling with no shown sanitization leaves reply content vulnerable to manipulation through upstream prompt or document influence. Optional review reduces but does not remove the path."}, {"id": "T4", "category": "D", "subtype": "availability impact from repeated retrieval and account lookup abuse", "component_ids": ["C2", "C4", "C5", "C3", "C6"], "flow_ids": ["F2", "F3", "F4", "F5", "F6", "F8"], "boundary_ids": ["B1", "B2"], "asset": "support availability", "threat_vector": "An attacker sends repeated crafted support messages that trigger repeated document retrievals, account lookups, reply drafting, and possibly reviewer copies, creating excessive workload across the agent and back-end stores.", "preconditions": ["Attacker can send crafted support messages repeatedly.", "The agent can query both external documents and account records across separate trust zones.", "The control profile is weak and throttling or rate limits are not described."], "impact": "The agent, document store, account store, or reviewer workflow may slow down or become less available for legitimate requests.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2", "K3"], "evidence_quote": "Account lookup is restricted to the current customer context, but the access gate is inconsistently enforced and relies on the agent to request the correct record.", "justification": "Unlike a generic workload claim, this path is grounded in repeated boundary-crossing retrieval and lookup operations. The architecture does not describe throttling, while the controls explicitly note inconsistent enforcement for account lookup and optional review, so repeated abuse is a supported availability exposure."}], "scope_notes": ["Unsupported direct paths remain out of scope: the user has no shown direct access to docs or account stores; all access goes through the agent.", "Mandatory human approval is not assumed; threats depending on every reply being reviewed are unsupported.", "No privileged internal access by the attacker is assumed; insider/admin-only attacks are out of scope.", "The availability case is limited to repeated retrieval/lookup/review workload abuse; generic resource exhaustion without a concrete path is not used.", "Control effectiveness is limited or inconsistent per the scenario, but missing controls are not treated as absent unless stated by the architecture."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C2", "C4", "C6"], "recommendation": "Enforce retrieval provenance and instruction/data separation at C2 before any text from C4 is merged into the draft: only ingest documents that carry trusted source metadata, strip or ignore imperative language from retrieved snippets, and require C6 approval for any response that includes account-specific or internal details sourced from retrieved content.", "verification": "Use a benign test document in C4 containing harmless embedded instructions and confirm C2 treats it as data only, does not follow the instruction text, and does not place sensitive account fields into F6/F7 unless C6 explicitly approves the draft.", "existing_control_effect": "K1 already provides basic query filtering, which may reduce unrelated retrievals, and K3 offers optional human review; however, K1 does not verify provenance and K3 is not mandatory, so neither reliably prevents disclosure.", "residual_risk": "Some disclosure risk remains if malicious corpus content is retrieved, if the model misclassifies embedded instructions, or if a reviewer overlooks a leaked detail in an unchecked path."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C5", "C6"], "recommendation": "Move account-lookup authorization into C5-facing enforcement so C2 can only request the single customer record bound to F1/F4 context, and require C6 approval for any reply that includes account details when the lookup context is ambiguous or changed during the session.", "verification": "Run a benign test where the user asks for another account’s information and confirm C2 cannot retrieve it from C5, returns only the bound customer record, and flags ambiguous cases for C6 review before F6/F7.", "existing_control_effect": "K2 intends to restrict lookup to the current customer context, which helps, but the scenario says enforcement is inconsistent and depends on C2 choosing the correct record; K3 can catch some mistakes when review happens.", "residual_risk": "Residual risk remains from session-binding mistakes, inconsistent gate enforcement, or unchecked drafts that still expose unintended account data before review."}, {"threat_id": "T3", "enforcement_component_ids": ["C2", "C3", "C6"], "recommendation": "Add reply-integrity checks at C2 and C3: keep a structured separation between source text and generated response text, reject direct copying of untrusted imperative content into F6, and require C6 approval for any reply that contains policy-sensitive instructions, account data, or rewritten content derived from retrieved text.", "verification": "Inject a benign malicious-style prompt or document snippet that attempts to rewrite the response, then confirm C2 does not pass that instruction into C3 as authoritative reply content and that C6 sees a clearly labeled draft for review when risky transformations are present.", "existing_control_effect": "K1 limits some retrieval content, K3 provides optional review, and K4 keeps the channel plain text; however, K4 also means there is no separate sanitization layer, so the existing controls do not guarantee reply integrity.", "residual_risk": "Attackers may still influence wording or framing if untrusted text reaches the draft and the response is sent without review; optional review lowers but does not eliminate tampering risk."}, {"threat_id": "T4", "enforcement_component_ids": ["C2", "C4", "C5", "C6"], "recommendation": "Apply explicit rate limiting and per-session request budgets at C2 for F2/F4, cache or deduplicate repeated lookups where safe, and cap optional-review submissions to C6 so repeated crafted messages cannot continuously trigger expensive retrieval, lookup, and review work.", "verification": "Send a harmless burst of repeated support messages and confirm C2 enforces request limits, stops repeated F2/F4 fan-out after the budget is reached, and does not forward unlimited copies to C6 for review.", "existing_control_effect": "K1 and K2 show that retrieval and account lookup already pass through C2 with some filtering/context checks, which provides a place to enforce limits; K3 may reduce risk by reviewing some drafts, but it also adds work when used repeatedly.", "residual_risk": "Some availability risk remains from distributed low-rate abuse, legitimate spikes, or work that still occurs before limits trigger, but the repeated cross-boundary load should be substantially reduced."}]} | |
| {"id": "DEV-diagram-0002", "collection": "development", "domain": "finance_operations", "control_profile": "guarded", "generation_condition": "development_workflow", "trial_id": "", "topology_family_id": "topology-043", "annotation_status": "candidate", "image_path": "images/development/DEV-diagram-0002.png", "title": "Invoice agent with approval gate", "purpose": "Illustrate a guarded finance-operations workflow where an invoice agent extracts payment requests, a human approver reviews them, and a separate payment service executes transfers after approval.", "assumptions": ["Invoices arrive from external vendors through a document intake channel.", "The invoice agent can read invoice content and basic vendor master data, but does not directly execute payments.", "A human approver is required before any payment execution request is sent onward.", "The payment execution boundary is separated from document processing and approval review.", "Control profile is guarded: basic validation and access checks exist at the trust boundaries, but the exact implementation details are not fully specified.", "An external sender may supply malformed or misleading invoice fields, but cannot directly invoke the payment service without crossing the approval path."], "controls": [{"id": "K1", "component_ids": ["C1"], "flow_ids": ["F1"], "fact": "Document intake performs basic file/type checks before forwarding invoices into the review zone."}, {"id": "K2", "component_ids": ["C2", "C5"], "flow_ids": ["F2", "F3"], "fact": "The invoice agent is limited to read access on vendor master data and write access only to draft invoice records."}, {"id": "K3", "component_ids": ["C3"], "flow_ids": ["F4", "F5"], "fact": "Payment requests require explicit human approval in the review desk before any message is sent to the payment service."}, {"id": "K4", "component_ids": ["C4", "C6"], "flow_ids": ["F5", "F6"], "fact": "The payment service accepts only approved payment orders and returns status updates to the invoice ledger."}, {"id": "K5", "component_ids": ["C4"], "flow_ids": ["F5"], "fact": "The payment rail is separated from document processing by an access boundary; direct calls from intake are not configured."}], "architecture": {"components": [{"id": "C1", "label": "invoice intake", "zone": "Z1", "kind": "process"}, {"id": "C2", "label": "invoice agent", "zone": "Z2", "kind": "agent"}, {"id": "C3", "label": "review desk", "zone": "Z2", "kind": "agent"}, {"id": "C4", "label": "pay service", "zone": "Z3", "kind": "process"}, {"id": "C5", "label": "vendor master", "zone": "Z2", "kind": "data_store"}, {"id": "C6", "label": "invoice ledger", "zone": "Z2", "kind": "data_store"}], "flows": [{"id": "F1", "source": "C1", "target": "C2", "data": "invoice doc", "boundary_ids": ["B1"]}, {"id": "F2", "source": "C2", "target": "C5", "data": "vendor lookup", "boundary_ids": []}, {"id": "F3", "source": "C2", "target": "C6", "data": "draft invoice", "boundary_ids": []}, {"id": "F4", "source": "C2", "target": "C3", "data": "payment request", "boundary_ids": []}, {"id": "F5", "source": "C3", "target": "C4", "data": "approved pay order", "boundary_ids": ["B2"]}, {"id": "F6", "source": "C4", "target": "C6", "data": "pay status", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "vendor"}, {"id": "Z2", "label": "finance review"}, {"id": "Z3", "label": "payment rail"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z2", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "invoice field tampering / draft invoice manipulation", "component_ids": ["C1", "C2", "C6"], "flow_ids": ["F1", "F3"], "boundary_ids": ["B1"], "asset": "invoice processing integrity and invoice ledger records", "threat_vector": "A malformed or misleading external invoice document alters extracted fields or draft invoice contents before review.", "preconditions": ["An external vendor can supply invoice content through the document intake channel.", "The invoice agent processes invoice content into draft invoice records."], "impact": "Incorrect invoice data may be recorded and presented for approval, potentially causing wrong payment requests or ledger corruption.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1", "K2"], "evidence_quote": "Document intake performs basic file/type checks before forwarding invoices into the review zone.", "justification": "The architecture explicitly allows external invoices into the workflow and the agent writes draft invoice records. Basic file/type checks reduce some malformed input risk, but they do not eliminate content-level tampering of invoice fields crossing B1."}, {"id": "T2", "category": "I", "subtype": "vendor master and ledger data exposure", "component_ids": ["C2", "C5", "C6"], "flow_ids": ["F2", "F3"], "boundary_ids": [], "asset": "vendor master data and invoice ledger contents", "threat_vector": "The invoice agent can read vendor master data and write draft invoice records, creating a supported path for unauthorized disclosure if the agent is compromised or over-privileged within Z2.", "preconditions": ["The invoice agent processes vendor lookups and draft invoice records.", "An attacker gains access to or influence over the invoice agent or its outputs within the finance review zone."], "impact": "Sensitive vendor or invoice information could be exposed from read-accessed master data or ledger records to an unauthorized party in the review zone.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "The invoice agent is limited to read access on vendor master data and write access only to draft invoice records.", "justification": "The diagram shows C2 reading C5 and writing C6, so disclosure is structurally plausible, but the actual confidentiality controls, downstream access restrictions, and monitoring effectiveness are not specified. Read/write limits reduce some misuse, yet they do not fully characterize whether exposure is prevented."}, {"id": "T3", "category": "T", "subtype": "payment-request handoff tampering", "component_ids": ["C2", "C3", "C4"], "flow_ids": ["F4", "F5"], "boundary_ids": ["B2"], "asset": "payment request integrity and approved pay order correctness", "threat_vector": "A payment request could be altered in transit within the review-to-payment path, or an approved pay order could be modified before it reaches the payment service.", "preconditions": ["The invoice agent creates payment requests for review.", "The architecture does not specify message integrity or signing for the handoff from review desk to payment service."], "impact": "The payment service may receive a pay order that does not match the reviewer’s intended approval, resulting in incorrect execution.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "Payment requests require explicit human approval in the review desk before any message is sent to the payment service.", "justification": "Human approval is explicit and the payment service only accepts approved payment orders, but the integrity of the message content between approval and execution is unspecified. This supports a tampering concern on the handoff, but the effectiveness of any integrity protection is not described."}, {"id": "T4", "category": "D", "subtype": "approval-path denial of service", "component_ids": ["C1", "C2", "C3"], "flow_ids": ["F1", "F4"], "boundary_ids": ["B1"], "asset": "invoice processing availability and approval throughput", "threat_vector": "Malformed, excessive, or misleading invoices consume intake and agent resources, delaying or blocking payment requests from reaching human review.", "preconditions": ["An external sender can submit invoices through the intake channel.", "The intake and agent path processes each invoice before approval."], "impact": "Legitimate invoices may be delayed, backlog may build in the review desk, and payment processing throughput may be reduced.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Document intake performs basic file/type checks before forwarding invoices into the review zone.", "justification": "The workflow depends on a single intake-to-agent-to-review path, and the only explicit upstream control is basic file/type checking. That control may reduce some malformed input, but it does not fully address resource exhaustion or queue flooding across B1."}], "scope_notes": ["Boundary references are limited to the boundaries actually crossed by the referenced flows: F1 crosses B1; F5 crosses B2; F1, F4 do not cross B2; F2, F3, and F4 do not cross a trust boundary.", "The feedback suggested a separate repudiation annotation, but the architecture does not specify logging, signatures, or non-repudiation mechanisms; that category is not included as a supported path.", "The architecture does support an information-disclosure candidate via C2's read access to C5 and interaction with C6, but the exact confidentiality controls are unspecified, so assessment remains unknown.", "No direct payment-service invocation from intake is modeled; unsupported direct-call paths are out of scope except as prevented by K5.", "Control details are only partially specified, so several threats remain unknown rather than reduced when the effectiveness of checks is not described.", "The architecture does not specify authentication, logging, signature verification, or content validation beyond the quoted controls; these omissions are treated as incomplete information, not as proof of absence."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["C1", "C2", "C6"], "recommendation": "Add content-level invoice validation at C1 and C2: enforce schema, field-range, vendor-ID, amount/currency consistency, and duplicate-invoice checks before C2 writes any draft record; only allow C6 updates from validated draft data, with ledger-side reconciliation checks that reject records missing a validated provenance token from C2.", "verification": "Submit benign test invoices with malformed fields, swapped vendor IDs, duplicate invoice numbers, and out-of-range totals; confirm C1 rejects malformed documents and C2 refuses to write unvalidated drafts, while C6 only accepts records carrying the expected validation token.", "existing_control_effect": "K1 and K2 already reduce basic malformed-file input and limit C2 to draft writes, which helps contain some bad inputs and prevents direct payment execution from the agent.", "residual_risk": "A well-formed but misleading invoice can still pass basic checks if business-rule validation is incomplete or the vendor master itself contains incorrect data."}, {"threat_id": "T2", "enforcement_component_ids": ["C2", "C5", "C6"], "recommendation": "Restrict C2 to minimum-necessary data from C5 and mask sensitive ledger fields in C6; add access logging and per-field authorization on reads, and ensure draft outputs from C2 cannot expose full vendor or ledger contents beyond approved review context.", "verification": "Use a non-sensitive test account to request vendor lookups and ledger views; verify C2 receives only required fields, that unauthorized fields are redacted, and that read attempts are logged and attributable without exposing sensitive values.", "existing_control_effect": "K2 already constrains C2 to read access on vendor master data and write-only draft invoice records, which lowers the chance of uncontrolled disclosure through broad privileges.", "residual_risk": "If C2 is compromised or the review zone is over-exposed, authorized reads may still be abused to leak permitted data or infer sensitive information from draft outputs and operational metadata."}, {"threat_id": "T3", "enforcement_component_ids": ["C3", "C4", "C6"], "recommendation": "Bind the approved pay order to an integrity-protected approval artifact: have C3 sign or MAC the reviewed payment request, require C4 to verify that artifact before execution, and record the approved order hash in C6 so any change between approval and execution is detectable.", "verification": "Send a benign approved payment request and then replay it with altered amount, payee, or reference fields; confirm C4 rejects modified payloads and only accepts the exact approved hash or signature recorded in C6.", "existing_control_effect": "K3 ensures human approval is required before forwarding, and K4 ensures C4 accepts only approved payment orders, which already narrows the path for unauthorized execution.", "residual_risk": "If the approval artifact or key material is compromised, an attacker could still produce a seemingly valid but malicious order, and human approval does not by itself guarantee content integrity after handoff."}, {"threat_id": "T4", "enforcement_component_ids": ["C1", "C2", "C3"], "recommendation": "Add rate limiting, queue quotas, and prioritization at C1 and C2, plus backpressure and triage at C3 so malformed or high-volume submissions cannot monopolize the approval path; reject or defer abusive senders before they consume review capacity.", "verification": "Use a controlled test harness to submit a small burst of oversized or malformed invoices alongside normal invoices; confirm C1 throttles excessive submissions, C2 does not create unbounded work, and C3 still receives and processes legitimate items within expected latency.", "existing_control_effect": "K1 already filters some invalid documents at intake, which reduces trivial malformed-input load before it reaches the agent and review desk.", "residual_risk": "An attacker can still create load with syntactically valid invoices or many low-rate submissions, so some backlog and latency impact may remain under sustained abuse."}]} | |
| {"id": "DEV-diagram-0003", "collection": "development", "domain": "incident_response", "control_profile": "unspecified", "generation_condition": "development_workflow", "trial_id": "", "topology_family_id": "topology-044", "annotation_status": "candidate", "image_path": "images/development/DEV-diagram-0003.png", "title": "Incident response triage hub", "purpose": "Model an incident-response workflow where two analyst agents independently triage evidence, a coordinator reconciles their findings, and a reporting step packages outcomes for an external case system.", "assumptions": ["The system is used for real incident tickets and can ingest alerts, logs, and analyst notes from external sources.", "Two analyst agents have distinct working views and exchange findings through a coordinator rather than directly sharing all state.", "The coordinator may merge, prioritize, or reframe findings before exporting a case summary.", "Relevant control implementation details such as validation, authentication, and sanitization are not specified in the brief.", "At least one external evidence source and one external case sink exist outside the core trust boundary.", "Attacker capabilities are hypothetical and may include supplying malformed or misleading evidence at the external feed, influencing analyst notes, or attempting to alter outbound case content if controls are absent or weak."], "controls": [{"id": "K1", "component_ids": ["P1"], "flow_ids": ["F1"], "fact": "Evidence is accepted from an external feed into a single intake step; the brief does not specify filtering, authentication, or schema checks."}, {"id": "K2", "component_ids": ["A1", "A2"], "flow_ids": ["F2", "F3", "F4", "F5"], "fact": "The two analyst agents keep separate working views and exchange results only through the coordinator, not directly with each other."}, {"id": "K3", "component_ids": ["C1", "P2"], "flow_ids": ["F6", "F7"], "fact": "The coordinator is the only synthesis point before the report step exports a summary to the external case store."}, {"id": "K4", "component_ids": ["D1"], "flow_ids": ["F7"], "fact": "The case sink is outside the core trust boundary; persistence and downstream handling are not described in the brief."}], "architecture": {"components": [{"id": "A1", "label": "Analyst A", "zone": "Z1", "kind": "agent"}, {"id": "A2", "label": "Analyst B", "zone": "Z1", "kind": "agent"}, {"id": "C1", "label": "Coordinator", "zone": "Z1", "kind": "agent"}, {"id": "P1", "label": "Intake", "zone": "Z1", "kind": "process"}, {"id": "P2", "label": "Report", "zone": "Z1", "kind": "process"}, {"id": "E1", "label": "Alert feed", "zone": "Z2", "kind": "external_entity"}, {"id": "D1", "label": "Case store", "zone": "Z3", "kind": "data_store"}], "flows": [{"id": "F1", "source": "E1", "target": "P1", "data": "alerts, logs, notes", "boundary_ids": ["B1"]}, {"id": "F2", "source": "P1", "target": "A1", "data": "evidence bundle", "boundary_ids": []}, {"id": "F3", "source": "P1", "target": "A2", "data": "evidence bundle", "boundary_ids": []}, {"id": "F4", "source": "A1", "target": "C1", "data": "findings", "boundary_ids": []}, {"id": "F5", "source": "A2", "target": "C1", "data": "findings", "boundary_ids": []}, {"id": "F6", "source": "C1", "target": "P2", "data": "reconciled summary", "boundary_ids": []}, {"id": "F7", "source": "P2", "target": "D1", "data": "case update", "boundary_ids": ["B2"]}], "trust_zones": [{"id": "Z1", "label": "Core triage"}, {"id": "Z2", "label": "External evidence"}, {"id": "Z3", "label": "Case system"}], "trust_boundaries": [{"id": "B1", "zone_ids": ["Z1", "Z2"]}, {"id": "B2", "zone_ids": ["Z1", "Z3"]}]}, "threats": [{"id": "T1", "category": "T", "subtype": "Tampered external evidence injected into intake", "component_ids": ["P1"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "alerts, logs, notes", "threat_vector": "Malicious or malformed data supplied at the external alert feed alters the evidence bundle accepted by Intake.", "preconditions": ["Attacker can influence the external feed or a source feeding it.", "Intake does not fully validate, authenticate, or sanitize incoming content."], "impact": "False, incomplete, or misleading evidence can propagate into analyst work and case updates, affecting triage decisions and downstream records.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Evidence is accepted from an external feed into a single intake step; the brief does not specify filtering, authentication, or schema checks.", "justification": "A single intake point receives untrusted external evidence across B1, and the brief explicitly leaves validation controls unspecified, so tampering is a supported exposure."}, {"id": "T2", "category": "I", "subtype": "Exposure of sensitive evidence to analysts via incoming bundle", "component_ids": ["P1", "A1", "A2"], "flow_ids": ["F2", "F3"], "boundary_ids": [], "asset": "evidence bundle", "threat_vector": "Sensitive alerts, logs, or notes are distributed to both analyst working views once ingested.", "preconditions": ["The evidence bundle contains sensitive or regulated content.", "Analyst workstation access or handling is broader than intended."], "impact": "Incident data may be disclosed to analysts who do not need the full contents, or retained in separate working views beyond necessity.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K2"], "evidence_quote": "The two analyst agents keep separate working views and exchange results only through the coordinator, not directly with each other.", "justification": "The design clearly routes evidence to both analysts, but the brief does not specify access scoping, redaction, or view-level filtering, so disclosure risk is present but control effectiveness is unspecified."}, {"id": "T3", "category": "T", "subtype": "Analyst finding manipulation before reconciliation", "component_ids": ["A1", "A2", "C1"], "flow_ids": ["F4", "F5"], "boundary_ids": [], "asset": "findings", "threat_vector": "A compromised or misleading analyst can submit altered findings to the coordinator, influencing the reconciled summary.", "preconditions": ["An attacker can influence one analyst agent, its inputs, or its local notes.", "The coordinator does not independently verify the underlying evidence for each finding."], "impact": "The reconciled summary may reflect manipulated conclusions, biasing the case report and any downstream response.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K2", "K3"], "evidence_quote": "The coordinator is the only synthesis point before the report step exports a summary to the external case store.", "justification": "Findings from each analyst flow into a single synthesis point, so tampering with one analyst's output can affect the exported summary; no verification control is specified."}, {"id": "T4", "category": "T", "subtype": "Outbound case content alteration", "component_ids": ["P2", "D1"], "flow_ids": ["F7"], "boundary_ids": ["B2"], "asset": "case update", "threat_vector": "The report step could alter, truncate, or reframe the case update before it crosses into the external case store.", "preconditions": ["An attacker can influence the report step or its inputs.", "Outbound content integrity checks are absent or weak."], "impact": "The external case system may persist incorrect or incomplete incident records, affecting auditability and follow-up actions.", "assessment": "exposed", "severity": "not_assessed", "evidence_control_ids": ["K3", "K4"], "evidence_quote": "The case sink is outside the core trust boundary; persistence and downstream handling are not described in the brief.", "justification": "The only export path into the external sink crosses B2, and the brief does not describe persistence or downstream integrity controls, so outbound tampering is a supported exposure."}, {"id": "T5", "category": "D", "subtype": "Resource exhaustion at intake", "component_ids": ["P1"], "flow_ids": ["F1"], "boundary_ids": ["B1"], "asset": "intake processing availability", "threat_vector": "Large volumes or pathological alerts, logs, or notes from the external feed overwhelm the single intake step.", "preconditions": ["An attacker can send high-rate or oversized inputs at the external feed.", "Intake lacks rate limiting, queue limits, or robust parsing safeguards."], "impact": "Triage ingestion slows or fails, delaying analyst review and case updates.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K1"], "evidence_quote": "Evidence is accepted from an external feed into a single intake step; the brief does not specify filtering, authentication, or schema checks.", "justification": "A single intake point on the external boundary is a plausible denial-of-service target, but rate limiting or capacity controls are not specified."}, {"id": "T6", "category": "D", "subtype": "Coordinator bottleneck failure", "component_ids": ["C1", "P2"], "flow_ids": ["F4", "F5", "F6"], "boundary_ids": [], "asset": "triage workflow availability", "threat_vector": "Overload, malformed inputs, or a stalled merge path at the coordinator blocks reconciliation and report generation.", "preconditions": ["One or both analyst streams can be saturated or malformed.", "The coordinator has no redundancy or fallback merge path described."], "impact": "Findings cannot be reconciled and no case summary is exported, interrupting the incident-response workflow.", "assessment": "unknown", "severity": "not_assessed", "evidence_control_ids": ["K3"], "evidence_quote": "The coordinator is the only synthesis point before the report step exports a summary to the external case store.", "justification": "A single synthesis point is an availability dependency, but the brief does not say whether buffering, retries, or alternate processing exist."}], "scope_notes": ["Unsupported paths such as direct analyst-to-analyst sharing are not shown in the architecture; the model only includes coordinator-mediated exchange.", "Authentication, schema validation, sanitization, rate limiting, authorization, and integrity verification are explicitly unspecified, so several threats remain 'unknown' rather than assumed present or absent.", "The external case store's downstream persistence, access controls, and audit behavior are out of scope in the brief.", "No explicit repudiation or elevation-of-privilege pathway is evidenced by the supplied flows; those categories are not asserted here."], "mitigations": [{"threat_id": "T1", "enforcement_component_ids": ["P1"], "recommendation": "Add mandatory input validation at Intake: authenticate the external source where possible, enforce a strict schema and size limits for alerts/logs/notes, and quarantine or reject malformed or unsigned records before they enter the evidence bundle.", "verification": "Submit benign sample records that cover valid, oversized, malformed, and unsigned cases; confirm only schema-compliant, authorized inputs proceed and rejected items are logged without reaching A1/A2.", "existing_control_effect": "K1 already concentrates external evidence through a single intake step, which creates a clear enforcement point, but the brief says filtering, authentication, and schema checks are unspecified.", "residual_risk": "If upstream sources are compromised or validation rules are incomplete, some misleading content may still be admitted, though the attack surface is reduced."}, {"threat_id": "T2", "enforcement_component_ids": ["P1", "A1", "A2"], "recommendation": "Apply need-to-know filtering at Intake and per-analyst view redaction so each analyst receives only the evidence fields required for their task; avoid exposing full sensitive bundles when a summarized or masked view is sufficient.", "verification": "Use test tickets with tagged sensitive fields and confirm the intake step redacts or partitions them before delivery, and that each analyst view shows only authorized fields.", "existing_control_effect": "K2 already keeps analyst working views separate and prevents direct analyst-to-analyst sharing, but it does not state that the incoming bundle is minimized or redacted.", "residual_risk": "Both analysts still receive some evidence content, so exposure remains possible if the bundle itself contains sensitive data that is legitimately needed for triage."}, {"threat_id": "T3", "enforcement_component_ids": ["C1", "P2"], "recommendation": "Require the coordinator to reconcile findings against the underlying evidence bundle before producing the summary, and record per-finding provenance so altered analyst conclusions can be traced and challenged.", "verification": "Run a controlled case where one analyst finding is intentionally inconsistent with the source evidence; confirm the coordinator flags the discrepancy, preserves provenance, and does not silently incorporate the altered conclusion.", "existing_control_effect": "K2 and K3 already route findings only through the coordinator and make it the single synthesis point, which helps contain uncontrolled cross-talk but does not by itself verify correctness.", "residual_risk": "A compromised analyst or misleading evidence can still bias reconciliation if the coordinator’s cross-checking is incomplete or if the underlying evidence itself is deceptive."}, {"threat_id": "T4", "enforcement_component_ids": ["P2", "D1"], "recommendation": "Protect the report-to-case-store export with integrity checks and an immutable outbound summary format: generate a signed or checksummed case update, validate it at the sink boundary, and reject any truncated or modified payload before persistence.", "verification": "Send benign reports with known checksums and intentional formatting variations; confirm the case sink accepts only intact updates and logs any mismatch or truncation as a failed write.", "existing_control_effect": "K3 and K4 identify the report step as the sole export path and mark the case sink as outside the core trust boundary, which makes boundary enforcement possible but does not describe integrity protection.", "residual_risk": "If the report step is influenced before signing or if the external case system ignores validation, incorrect records may still be stored or mirrored downstream."}, {"threat_id": "T5", "enforcement_component_ids": ["P1"], "recommendation": "Add intake rate limiting, bounded queues, and parser hardening at P1 so oversized or high-rate external evidence cannot monopolize processing; drop or defer excess traffic rather than allowing unbounded buffering.", "verification": "Use controlled load tests with increasing message rates and payload sizes; confirm intake stays responsive, enforces queue limits, and degrades gracefully without blocking normal triage traffic.", "existing_control_effect": "K1 provides a single intake choke point, which helps centralize overload protection, but the brief does not specify any capacity or parsing safeguards.", "residual_risk": "A determined sender can still consume some intake capacity or trigger throttling, so brief delays may remain even when the system remains available."}, {"threat_id": "T6", "enforcement_component_ids": ["C1", "P2"], "recommendation": "Introduce coordinator resilience controls: bounded work queues, timeout handling, and a fallback path for incomplete reconciliations so a stalled merge does not block report generation indefinitely.", "verification": "Simulate slow, malformed, and partially missing analyst inputs; confirm the coordinator times out, isolates bad inputs, and either produces a bounded partial summary or cleanly defers the case without hanging.", "existing_control_effect": "K3 already makes the coordinator the only synthesis point, which simplifies control placement, but the brief does not describe buffering, retries, or redundancy.", "residual_risk": "Availability still depends on the coordinator path, so severe overload or persistent malformed inputs can delay reconciliation even with graceful degradation."}]} | |