Trust No Skill: Integrity Verification for AI Agent Supply Chains
Palo Alto Networks introduces Behavioral Integrity Verification (BIV), a novel audit primitive that detects when AI agent skills deviate from their declared capabilities by analyzing metadata, executable code, and natural-language instructions across three modalities. Scanning 49,943 skills in the OpenClaw registry, BIV found 80% of skills exhibit behavioral mismatches, with 18.9% traced to adversarial intent — including multi-stage exploit chains that combine file reading, encoding, and network exfiltration to steal credentials or enable RCE. The threat model mirrors mobile and browser extension ecosystems a decade ago: extensibility has outpaced supply-chain audit primitives. Organizations deploying LLM agents with third-party skills face active risk from unvetted skill installations with privileged access to environment variables, file systems, and shell execution.