Benchmark Reveals Persistent Privacy and Impersonation Risks in AI Persona Skills
A new arXiv preprint introduces AntiSkillBench, a benchmark designed to systematically evaluate privacy leakage and impersonation risks in AI persona skills. The benchmark includes 7,500 persona-grounded dialogue traces from 50 diverse profiles and tests multiple skill-distillation strategies and defenses. Results show that privacy and impersonation risks persist across different AI agent architectures and distillation protocols, while current defenses are only partially effective and do not generalize well.
Why it matters: This work exposes a significant and under-addressed safety risk in AI personalization, with implications for privacy and trust in AI systems.
Full story at: arXiv Cryptography and Security ↗