When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills

TL;DR - AntiSkillBench evaluates privacy leakage and impersonation risks when agents distill personal interaction histories into reusable persona skills. Results show persistent risks across agent backbones and weak generalization from existing defenses.

• Includes 7,500 persona-grounded dialogue traces from 50 behaviorally rich profiles.
• Measures attribute disclosure and impersonation of communication styles and personality traits across three skill-distillation strategies.
• Tests four online and post-hoc defense configurations, including risk suppression and provenance protection.
• Defense effectiveness varies by distillation method and does not generalize reliably across risks.