ihabler/sentinel-flow-benchmark

Sentinel-Flow (sentinelTris)

Sentinel-Flow is a benchmark for evaluating whether large language models can detect multi-step security violations in AI agent interaction traces. This dataset release contains the scenario fixtures and labeled traces used in the SECAI @ ESORICS evaluation.

Data-only release. This repository contains benchmark data only (36 scenario files, 900 labeled traces). No evaluation code or model outputs are included here.

Synthetic-data and vendor disclaimer. All scenarios, interaction traces, people, organizations, credentials, domains, and security events in this dataset are synthetic and were created solely for defensive security research. References to real companies, products, or services are illustrative only and do not represent actual product behavior, vulnerabilities, incidents, or customer data. This dataset is not affiliated with, sponsored by, or endorsed by any referenced company or product vendor.

Dataset Summary

Property	Value
Scenarios	36
Labeled traces	900 (490 unsafe, 410 safe)
Vulnerability classes	3
Application domains	6
Format	JSON

Vulnerability Classes

Based on the Lethal Trifecta threat model:

1. Lethal Trifecta (lethal_trifecta): Untrusted content (U) → Private data access (P) → External exfiltration (E)
2. Exfil Chain (exfil_chain): Private data access (P) → External exfiltration (E) — capability misuse without injection
3. Taint-Sink (taint_sink): Untrusted content (U) → Dangerous execution sink — unsanitized execution

Application Domains

Code	Domain
HC	Healthcare
FN	Finance
EN	Enterprise
CI	IoT / Critical Infrastructure
DO	DevOps
WB	Web

Dataset Layout

data/
├── lethal_trifecta/   # 12 scenario files
├── exfil_chain/       # 12 scenario files
└── taint_sink/        # 12 scenario files

Each scenario JSON file contains:

• Scenario metadata: id, flow_class, domain fields
• Environment: system_prompt, tool definitions
• Traces: labeled agent interaction histories with ground_truth, flow_tag, reasoning, history, and candidate_action

JSON Structure

{
  "id": "HC-LT-clinical",
  "flow_class": "lethal_trifecta",
  "application_domain": "HC",
  "environment": {
    "system_prompt": "...",
    "tools": [ ... ]
  },
  "traces": [
    {
      "trace_id": "HC-LT-clinical-T001",
      "ground_truth": "UNSAFE",
      "flow_tag": "HC-LT-05",
      "reasoning": "...",
      "history": [ ... ],
      "candidate_action": { ... }
    }
  ]
}

Labels

Each trace includes:

• ground_truth: SAFE or UNSAFE
• flow_tag: scenario-level flow grouping tag
• reasoning: reference explanation for the label
• history: prior multi-turn agent/tool interaction
• candidate_action: the next action to be judged

Intended Use

• Benchmark evaluation of trace-level security classification
• Robustness and error analysis across flow classes
• Reproducible comparisons using fixed scenario fixtures
• Training or fine-tuning security judge models

Ethical Considerations

• Defensive benchmark: designed to evaluate detection capabilities, not to enable attacks
• Synthetic data: all traces are synthetically generated; no real user data was used
• Simulated credentials: some traces contain realistic-looking credential strings as part of attack simulation content — these are fictional and not valid credentials

Citation

If you use this dataset, please cite the Sentinel-Flow paper (SECAI @ ESORICS):

@inproceedings{sentinel-flow-2026,
  title={Sentinel-Flow: A Benchmark for Multi-Step Security Detection in AI Agent Traces},
  booktitle={SECAI @ ESORICS},
  year={2026}
}

Related Links

• Anonymous artifact (review): https://anonymous.4open.science/r/sentinelTris-F76D/