Dataset

All scan results on this site are available as data.

Download

File	Contents
/data/summary.json	Current run: totals, severity and category counts, per-skill index, top risks, analyzers used
/data/scans/<slug>-scan.json	Full findings for one skill, current run
/data/categories.json	Category taxonomy with OWASP AST10 mapping and definitions
Repository results/	Every run, JSON and Markdown, with scanner version

A CSV export of the per-skill index and a single-file archive of all scan JSON are planned; until then, clone the repository.

Schema

Per-skill record (abridged):

{
  "skill_name": "...",
  "is_safe": true,
  "max_severity": "MEDIUM",
  "findings_count": 2,
  "scanner_version": "2.3.1",
  "analyzers_used": ["static_analyzer", "llm_analyzer", "meta_analyzer"],
  "timestamp": "2026-05-09T01:44:52",
  "findings": [{
    "id": "...", "rule_id": "LLM_DATA_EXFILTRATION",
    "category": "data_exfiltration", "severity": "MEDIUM",
    "title": "...", "description": "...", "file_path": "SKILL.md",
    "line_number": null, "snippet": "...", "remediation": "...",
    "analyzer": "llm",
    "metadata": { "model": "...", "overall_assessment": "...",
                  "validated": true, "confidence": "HIGH",
                  "exploitability": "Low", "impact": "..." }
  }]
}

Category slugs are the thirteen in categories.json. Severity is one of CRITICAL, HIGH, MEDIUM, LOW, INFO.

License

The dataset (scan results, summaries, taxonomy) is licensed CC BY 4.0. The pipeline and site code are MIT. Skill source code is not redistributed and remains under its own license at the linked repository. Finding text includes snippets quoted from scanned skills for identification.

Citation

AppSecHQ. skillscan.io: Security scan results for AI agent skills. Dataset, scan run YYYY-MM-DD. https://skillscan.io/data/

A machine-readable CITATION.cff is in the repository root. Please include the scan run date, because the corpus changes.

Caveats

Labels are scanner output, not ground truth. Selection is not random. Nothing was executed. Read the methodology before drawing conclusions, and the false-positive register before counting CRITICALs.