airblackbox/eu-ai-act-compliance-benchmark

EU AI Act Compliance Benchmark

55 Python AI agent files with ground-truth compliance labels across 6 EU AI Act articles.

The first public benchmark dataset for evaluating EU AI Act compliance scanners. Each sample is a realistic Python AI agent that either passes or fails specific articles — verified against the AIR Blackbox scanner with 100% label accuracy.

Why This Exists

The EU AI Act deadline is August 2, 2026. Fines reach €35M or 7% of global annual turnover. Multiple tools claim to scan for compliance, but there's no standard way to measure their accuracy. This dataset fixes that.

What's Inside

samples/          # 55 Python AI agent files (.py)
metadata/
  labels.jsonl    # Ground-truth labels (one JSON object per line)

Each sample is a realistic 20-80 line Python AI agent using one of 5 frameworks: OpenAI, LangChain, CrewAI, AutoGen, or RAG pipelines.

Each label contains:

{
  "id": "bare_openai_agent",
  "filename": "bare_openai_agent.py",
  "description": "Bare OpenAI agent with zero compliance patterns",
  "framework": "openai",
  "difficulty": "easy",
  "labels": {
    "art9": "FAIL",
    "art10": "FAIL",
    "art11": "PASS",
    "art12": "FAIL",
    "art14": "FAIL",
    "art15": "FAIL"
  },
  "score": "1/6",
  "scanner_checks": { ... }
}

Articles Checked

Article	Requirement	What the Scanner Looks For
9	Risk Management	Risk classification, access control, RBAC
10	Data Governance	Input validation (Pydantic), PII handling, data schemas
11	Technical Documentation	Logging imports, docstrings, type hints
12	Record-Keeping	Structured logging (structlog), audit trails, timestamps, HMAC
14	Human Oversight	Human-in-the-loop, kill switch, notifications
15	Robustness & Security	Injection detection, error handling, testing, rate limiting

Distribution

By compliance score:

Score	Count	Description
1/6	10	Minimal compliance (usually just docstrings)
2/6	9	Two articles passing
3/6	13	Half compliant
4/6	10	Most articles passing
5/6	4	Nearly complete
6/6	9	Full compliance

By framework:

Framework	Count
OpenAI	26
LangChain	10
CrewAI	7
AutoGen	2
RAG	6
Other	4

Usage

Evaluate a compliance scanner

import json

# Load ground truth
labels = {}
with open("metadata/labels.jsonl") as f:
    for line in f:
        entry = json.loads(line)
        labels[entry["filename"]] = entry

# Run your scanner on each sample
for filename, truth in labels.items():
    with open(f"samples/{filename}") as f:
        code = f.read()

    # your_scanner_result = your_scanner(code)
    # Compare your_scanner_result to truth["labels"]

Evaluate AIR Blackbox scanner

pip install air-blackbox-mcp

from air_blackbox_mcp.scanner import scan_code

with open("samples/bare_openai_agent.py") as f:
    result = scan_code(f.read())

print(result["compliance_score"])  # "1/6"

Sample Categories

The dataset includes these categories of samples:

Zero/minimal compliance — Bare agents from tutorials with no governance patterns. Common in production despite being high-risk.

Single-article compliance — Agents that pass exactly one article. Tests scanner precision.

Partial compliance (2-4 articles) — The most realistic category. Most production agents have some compliance but significant gaps.

Near-full compliance (5/6) — Agents missing exactly one article. Tests whether scanners can identify the specific gap.

Full compliance (6/6) — Reference implementations across multiple frameworks showing what "passing" looks like.

Edge cases — Dead code with compliance patterns, imported-but-unused libraries, compliance terms in comments/strings only, multi-framework files.

High-risk domains — Medical diagnosis, hiring/screening, autonomous trading, code execution, email sending. These domains face extra scrutiny under the EU AI Act.

How Labels Were Generated

Labels are generated by running the AIR Blackbox scanner (regex-based, deterministic) on each sample file. The scanner's output IS the ground truth. This means:

1. Labels are 100% reproducible — run the scanner yourself and get identical results
2. Labels reflect regex-based detection (not semantic understanding)
3. Some edge cases exist: comments mentioning compliance terms may trigger detection

This is intentional. The benchmark tests what scanners actually detect, including their false positives and limitations.

Limitations

This benchmark tests pattern-based compliance detection, not semantic compliance. A file with risk_classification in a comment will pass Article 9 even though no risk management is actually implemented. This reflects the current state of compliance scanning tools, which are predominantly regex-based.

Future versions may include:

• Semantic labels (does the code actually implement the compliance measure?)
• More samples (targeting 200+)
• Additional frameworks (Anthropic Agent SDK, Pydantic AI, DSPy)
• Multilingual samples

Citation

@dataset{air_blackbox_eu_ai_act_benchmark_2026,
  title={EU AI Act Compliance Benchmark},
  author={AIR Blackbox},
  year={2026},
  url={https://huggingface.co/datasets/air-blackbox/eu-ai-act-compliance-benchmark},
  note={55 Python AI agent files with ground-truth compliance labels across 6 EU AI Act articles}
}

License

Apache-2.0

Links

• Scanner: air-blackbox-mcp
• Website: airblackbox.ai
• GitHub: github.com/airblackbox