fan28/synthetic_authorization

Dataset Card: Synthetic Multi-Agent Access Control with Workflow Splits

Dataset Summary

This CSV contains 50,000 synthetic access-control decision records spanning 12,500 multi-agent workflows. Each workflow contains four steps (STEP_ID 0–3). Records describe an agent's role, task, requested resource and operation, workflow stage, privilege and delegation context, trust and risk attributes, and access decision (ACTION) with a violation reason (VIOLATION_TYPE). A preassigned SPLIT column supports workflow-disjoint training, validation, and evaluation.

The dataset is marked as synthetic in its filename; the underlying simulation, authorization policy, generator, and validation procedures were not supplied. Therefore, individual labels should be understood as dataset-provided outcomes, not independently verified security ground truth.

Dataset Structure

• Format: CSV, one row per workflow step / access decision.
• Rows: 50,000.
• Columns: 21.
• Workflows: 12,500 distinct WORKFLOW_ID values, with exactly four steps each.
• Agents: 250 distinct AGENT_ID values.
• Completeness: No missing field values observed.
• Duplicates: No completely duplicated rows observed.

Splits

Split	Rows	Workflows	Share
Train	35,000	8,750	70%
Validation (val)	7,500	1,875	15%
Test	7,500	1,875	15%

No WORKFLOW_ID appears in more than one split. Agent identities are not disjoint: agents can recur in different splits. Workflow-disjoint evaluation therefore does not establish generalization to unseen agents.

Data Fields

Field	CSV type	Description / interpretation
WORKFLOW_ID	string	Synthetic workflow identifier; grouping key.
STEP_ID	integer	Step index within workflow; observed values 0–3.
AGENT_ID	string	Synthetic agent identifier.
AGENT_ROLE	string	Acting agent's role: admin, executor, planner, researcher, or reviewer.
PARENT_ROLE	string	Role of the parent / delegating context, as labeled in the data.
TASK_TYPE	string	Assigned task category: customer_support, financial_review, market_research, report_generation, security_audit, or software_deployment.
RESOURCE_TYPE	string	Requested resource category (12 observed categories).
OPERATION	string	Requested action: approve, delegate, delete, execute, read, send, or write.
WORKFLOW_STAGE	string	Stage: planning, execution, review, or approved.
PRIVILEGE_LEVEL	integer	Acting agent's privilege tier, observed 0–3; exact policy semantics unspecified.
PARENT_PRIVILEGE_LEVEL	integer	Parent privilege tier, observed 0–3.
RESOURCE_SENSITIVITY	integer	Resource sensitivity tier, observed 0–3.
DELEGATION_DEPTH	integer	Delegation depth, observed 0–4.
TRUST_SCORE	float	Numeric trust score, observed 0.039–1.000.
TASK_RISK	float	Numeric task risk score, observed 0.1001–0.9499.
TASK_AUTHORIZED	integer	Binary task authorization indicator, values 0/1.
TRUSTED_TOOL	integer	Binary tool trust indicator, values 0/1.
SAME_TENANT	integer	Binary tenant-matching indicator, values 0/1.
ACTION	integer	Binary target: 1 = allow, 0 = deny, inferred from exact agreement with the violation reason field.
VIOLATION_TYPE	string	Outcome/reason label: none for allowed requests; one of 11 violation categories for denied requests.
SPLIT	string	Predefined partition: train, val, or test.

The meaning of numeric tiers and thresholds is not documented; do not presume they encode a particular production policy.

Labels and Distributions

Access decision (ACTION)

Label	Records	Percentage
0 — Deny	32,251	64.50%
1 — Allow	17,749	35.50%

Per-split class counts:

Split	Deny	Allow
Train	22,565	12,435
Validation	4,811	2,689
Test	4,875	2,625

Violation reasons (VIOLATION_TYPE)

Value	Count
none	17,749
insufficient_privilege	15,794
workflow_stage_violation	3,306
task_unauthorized	3,012
privilege_escalation	2,826
invalid_delegation	2,272
task_resource_out_of_scope	1,440
role_operation_violation	1,292
untrusted_tool	1,246
low_trust	459
cross_tenant_access	338
task_operation_out_of_scope	266

In this file, ACTION=1 if and only if VIOLATION_TYPE='none'. This is perfect outcome leakage if both are used as predictors of one another; evaluate binary action predictions without the violation-reason field. For violation classification, either predict all 12 categories as a joint allow/reason outcome or evaluate the 11 violations on denied examples only, while keeping selection methodology explicit.

Potential Uses

• Train and compare supervised models for allow/deny authorization decisions.
• Classify synthetic authorization failure reasons.
• Study model performance across resource types, agent roles, operations, risk levels, and delegation depths.
• Prototype multi-agent access-control research pipelines, explainability techniques, and workflow-level validation.

Recommended Evaluation

Use the supplied SPLIT values rather than a random row-level partition; rows from the same workflow are related. Report macro-F1, per-class precision/recall, confusion matrices, and allow/deny error rates. In access control, particularly report false allows (model predicts allow, label is deny) separately from false denies. Compare performance by violation category and risk-related features, and consider calibration and false-allow rates at decision thresholds relevant to the application.

Feature handling:

• For binary ACTION prediction, remove ACTION, VIOLATION_TYPE, and SPLIT from inputs.
• Treat identifiers (WORKFLOW_ID, AGENT_ID) cautiously: they may enable memorization and do not necessarily generalize.
• For reason prediction, remove VIOLATION_TYPE and avoid providing ACTION as an input unless the task is explicitly conditional on a known deny decision.
• Fit encoders, imputers, scalers, and any threshold selection on the train split only; tune on validation and reserve test for final evaluation.
• Consider whether fields like TASK_AUTHORIZED, TRUSTED_TOOL, and SAME_TENANT would actually be known at deployment time.

Limitations and Risks

1. Synthetic-to-real gap: No evidence is provided that simulated distributions, adversarial behavior, or privileges represent production multi-agent systems. Results should not be treated as proof of real-world access-control security.
2. Rule-based label generation and single-violation representation: Labels are generated deterministically using predefined authorization policies implemented in generate_multiagent_access_dataset.py. The binary ACTION label is assigned 1 (allow) only when all 11 authorization checks pass; otherwise, it is assigned 0 (deny). These checks cover role-operation permissions, task-resource and task-operation scope, privilege sufficiency and escalation, delegation constraints, trust thresholds, tool trustworthiness, tenant isolation, workflow-stage restrictions, and task authorization. The VIOLATION_TYPE label is assigned using a fixed precedence order, selecting the first failed check, or none when all checks pass. Consequently, although multiple policy violations may occur simultaneously, only one violation category is recorded per denied request. This representation may obscure concurrent security violations and introduce precedence-dependent label distributions. Moreover, because labels are derived from predefined synthetic policies rather than independently observed real-world authorization decisions, model performance primarily reflects the ability to reproduce the encoded rules and may not generalize to more complex or evolving production security policies.
3. Label leakage: VIOLATION_TYPE deterministically reveals ACTION, so feature choice must reflect prediction-time availability. Also, TASK_AUTHORIZED, TRUSTED_TOOL and SAME_TENANT deterministically reveal untrusted_tool, low_trust and cross_tenant_access respectively.
4. Uneven classes: Some violations (e.g., task_operation_out_of_scope, cross_tenant_access) are relatively rare; aggregate accuracy could hide poor detection of important failure types.
5. Identity overlap: Split separation is by workflow, not by agent; performance on unseen identities is not measured.
6. Security impact: A learned classifier should not replace a verified, default-deny authorization mechanism in a live environment without independent policy checks, adversarial evaluation, and human security review.
7. Missing provenance: Creator, collection/generation date, generating code, license, governance, and applicable privacy review were not supplied; do not assume permission for redistribution or production use.

Ethical and Privacy Considerations

Records contain synthetic-looking workflow and agent IDs and categorical operational context; no obvious personal names, email addresses, or direct personal identifiers appear in the inspected columns. Nevertheless, synthetic status and privacy guarantees have not been independently verified. Publication should follow provenance and license checks. Models should be audited for unequal error rates across relevant roles and task/resource categories; false authorization can create security harms.

Provenance and Licensing

• Source artifact: multiagent_access_synthetic_with_split.csv (user-provided).
• Generation script: generate_multiagent_access_dataset.py (user-provided; uses SEED = 42 and N = 50000).
• Creator / organization: Faiyaz Ansari.
• Creation date and generator: 10/01/2026.

Example Loading Code

import pandas as pd

df = pd.read_csv('multiagent_access_synthetic_with_split.csv')
features = df.drop(columns=['ACTION', 'VIOLATION_TYPE', 'SPLIT', 'WORKFLOW_ID'])
target = df['ACTION']

train_mask = df['SPLIT'].eq('train')
val_mask = df['SPLIT'].eq('val')
test_mask = df['SPLIT'].eq('test')

X_train, y_train = features.loc[train_mask], target.loc[train_mask]
X_val, y_val = features.loc[val_mask], target.loc[val_mask]
X_test, y_test = features.loc[test_mask], target.loc[test_mask]