AntiSkillBench: Evaluating Privacy Risks in Persona-Based Agent Skills
August 3, 2026
AntiSkillBench evaluates the risks of the persona-skill pipeline, where personal interaction histories are distilled into portable skills. The benchmark measures skill-level privacy leakage and the risk of agent-level attribute disclosure and behavioral impersonation.
HOW THIS AFFECTS YOU
●
builderYou must evaluate the risk of personal data leakage when distilling user history into executable agent skills.
●
policyYou should monitor the privacy implications of allowing agents to carry portable, distilled persona skills across platforms.