Story

Hi HN,

I posted SkillFortify here a few days ago as a formal verification tool for 3 agent skill formats. Based on feedback, v0.3 now supports 22 agent frameworks and can scan your entire system with zero configuration.

The problem: In January 2026, the ClawHavoc campaign planted 1,200 malicious skills into agent marketplaces. CVE-2026-25253 was the first RCE in agent software. Researchers catalogued 6,000+ malicious agent tools. The industry responded with heuristic scanners — pattern matching, YARA rules, LLM-as-judge. One popular scanner states in its docs: "No findings does not mean no risk."

SkillFortify eliminates that caveat with formal verification.

What it does:

  pip install skillfortify
  skillfortify scan

  [*] Auto-discovering AI tools on system...
  [+] Found: Claude Code (12 skills)
  [+] Found: Cursor (8 skills)
  [+] Found: VS Code MCP (5 servers)
  [+] Found: Windsurf (3 skills)
  [*] Scanning 28 skills across 4 tools...

  RESULTS
    Critical:  2 skills with capability violations
    High:      3 skills with excessive permissions
    Clean:     23 skills passed all checks

Why formal verification, not heuristics: Heuristic scanners check for known bad patterns. Novel attacks pass through. SkillFortify verifies what a skill CAN do against what it CLAIMS to do. Five mathematical theorems guarantee soundness — if it says safe, it provably cannot exceed declared capabilities.

Results on 540-skill benchmark (270 malicious, 270 benign): - F1 = 96.95% - Precision = 100% (zero false positives) - Recall = 94.07% - Speed: ~2.5ms per skill

9 CLI commands: - scan — auto-discover + analyze all AI tools on your system - verify — formally verify a single skill - lock — generate skill-lock.json (like package-lock.json for agent skills) - trust — compute graduated trust score (L0-L3, inspired by SLSA) - sbom — generate CycloneDX 1.6 Agent Software Bill of Materials - frameworks — list all 22 supported frameworks + detection status - dashboard — generate standalone HTML security report (zero dependencies) - registry-scan — scan MCP/PyPI/npm registries before installing - verify --recursive — batch verify entire directory trees

1,818 tests. 22 parsers. 97 source modules. MIT licensed. Peer-reviewed paper on Zenodo.

GitHub: https://github.com/varun369/skillfortify PyPI: https://pypi.org/project/skillfortify/ Paper: https://zenodo.org/records/18787663 Wiki: https://github.com/varun369/skillfortify/wiki Landing page: https://www.superlocalmemory.com/skillfortify

Built this as part of my research on making AI agents reliable enough for production. The companion project AgentAssert (arXiv:2602.22302) handles behavioral contracts — SkillFortify handles the supply chain.

Happy to answer questions about the formal model, framework support, or auto-discovery.