Formal Analysis and Supply Chain Security for Agentic AI Skills
96.15% F1, five theorems with full proofs, and a published negative result on information-flow analysis
Security · SkillFortify
SkillFortify is a formal security analysis tool for AI agent skills. Using the DY-Skill Dolev-Yao formal model, it detects skill injection, prompt leakage, unsafe tool invocations, and supply chain attacks, with five theorems and full proofs. The soundness theorem's scope is stated explicitly; it is not a claim of zero false positives.

Figures as reported by the project. Check them against the code and paper linked above.
Install
Install
pip install skillfortifyWhat it does
Five theorems with full proofs in the paper, with the scope of the soundness theorem stated explicitly.
Dolev-Yao model adapted for agent skills. Sound static analysis with formal security proofs.
Agent Software Bill of Materials. Track every skill, its dependencies, and their security posture.
Validated against a 540-skill benchmark. 96.15% F1 score.
Information-flow analysis adds no detections over pattern matching on the paper's corpus, and the paper reports that openly.
Use it for
Changelog
The latest releases on GitHub.
v0.6.0 — unified skill discovery, reproducible experiment harness
Discovery is now uniform across formats. All skill formats share one bounded tree walk, so OpenClaw install roots and MCP configurations are found anywhere below the scan root — matching the behaviour already provided for Claude skills. A single scan of a repository now…
v0.5.0 — full-surface discovery, graded findings, rebuilt benchmark
Discovery now covers every path the runtimes actually load skills from — personal, project, plugin and nested-package .claude/skills/, legacy .claude/commands/, and .openclaw/skills/ — following symlinks and de-duplicating on resolved path. MCP configuration discovery covers…
Research
96.15% F1, five theorems with full proofs, and a published negative result on information-flow analysis