Researchers introduce AntiSkillBench for persona-skill risks
The benchmark tests privacy leakage, attribute disclosure and impersonation in personalized AI agents.
Why it matters
The work highlights a safety gap in agent personalization: packaging user histories into reusable skills can concentrate personal signals and make leakage or impersonation harder to control. It gives researchers and builders a concrete evaluation target for defenses beyond record-level or retrieval-memory protections.
The key points
- 1.AntiSkillBench targets privacy and impersonation risks in persona skills.
- 2.The dataset contains 7,500 dialogue traces from 50 profiles.
- 3.Risks persisted across three frontier agents in experiments.
A new arXiv paper introduces AntiSkillBench, an end-to-end benchmark for evaluating risks and defenses in persona skills, which convert personal interaction histories into portable artifacts for downstream agents. The benchmark includes 7,500 persona-grounded dialogue traces from 50 behaviorally rich profiles, tests privacy leakage, attribute disclosure and behavioral impersonation across three skill-distillation strategies, and evaluates four defense configurations. Experiments across three frontier agents found persona-skill risks persist across agent backbones and distillation protocols.
⚡ Try this today
Read the paper before deploying persona-skill personalization, and test for leakage, attribute disclosure and impersonation risks.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Research
Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency
arXiv:2608.
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
arXiv:2608.
[Paper] Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iter