Research
AntiSkillBench: Persona Skills Leak Identity Across All Three Frontier Agents Tested, and Every Existing Defense Fails to Generalize
Persona skills compress a user's interaction history into a portable executable artifact, which concentrates fragmented personal signals and defeats defenses built for individual records or retrieval-based memory. The benchmark uses 7,500 persona-grounded dialogue traces from 50 behaviorally rich profiles and measures skill-level privacy leakage plus agent-level attribute disclosure and behavioral impersonation across three distillation strategies. Risk persisted across all three frontier agent backbones and every distillation protocol, extending past explicit attributes into communication style and personality traits, while the four tested defenses were distillation-dependent and did not transfer.
↳ Follow the thread