Agent Skill Security Auditing

Agent skill security auditing studies whether reusable instructions, code, and tools behave as intended and stay within an agent’s authorized scope. Auditors combine static checks such as pattern matching and data-flow analysis with semantic review of natural-language instructions, sandboxed execution, and runtime monitoring. Benchmarks increasingly test complete agent trajectories, skill dependencies, and compositions, since individually benign skills can interact to cause unsafe actions. A central challenge is balancing high detection recall with false positives and reliable performance on unfamiliar skills, adaptive attacks, and different agent frameworks.