Trust No Skill: Integrity Verification for AI Agent Supply Chains

This report introduces Behavioral Integrity Verification (BIV), a static, cross-modality audit method that compares declared metadata, executable code, and natural-language instructions of third-party LLM agent "skills". Applied at registry scale (49,943 skills), BIV found pervasive declaration-vs-actual behavior deviations, including 2,490 skills with multi-stage chains enabling credential exfiltration, instruction-override hijacking and remote code execution, and recommends targeted review tiers and mitigation steps for high-risk capability families.