skill-scan

Scan AI agent skill packages by analyzing SKILL.md, code, and assets to detect threats and misalignments. Run skill-scan on local directories or ClawHub slugs to identify prompt injection, evasion tactics, and trojan-like behavior before installation.

Multi-layer static analysis detects prompt injection by combining fast pattern matching, JS/TS AST evasion analysis, and optional LLM-powered semantic threat analysis. This approach correlates findings across the skill package to reduce false positives and surface hidden security threats.

Alignment verification checks whether SKILL.md descriptions match actual code behavior to surface trojan-like misalignments. By cross-referencing declared functionality with static analysis results, it identifies skills that hide malicious intent behind benign documentation.

Batch scan installed skills by running skill-scan batch against a directory path. This enables scalable security governance by processing multiple local skill packages simultaneously, applying multi-layer threat detection and meta-analysis across the entire set.

ClawHub integration allows you to scan skills directly from the registry using the scan-hub command with a slug. It verifies external dependencies and audits skill packages before installation, supporting both local directories and registry-based workflows.

Pattern matching alone may miss sophisticated evasion techniques and semantic threats in AI agent skill packages. Skill-scan addresses this by layering AST/evasion analysis, prompt-injection checks, and optional LLM analysis to correlate findings and reduce missed threats.