Agent Skills Matter: Inferring Proprietary Skills from Execution Trajectories

Researchers introduced SigLeak, a black-box method for inferring hidden agent skills from execution trajectories.

The paper frames this as “Skill Leakage,” where observable behavior reveals proprietary procedures and procedural knowledge.

SigLeak builds diagnostic tasks, compares skill-enabled versus skill-disabled runs, and reconstructs skills from recurring patterns.

Across multiple scenarios, model families, and agent frameworks, it matched or beat baselines and improved success rates.

The findings raise a new privacy and IP risk for cloud-hosted AI agents and other black-box systems.