Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-05 · high
Four defenses were tested, one per distillation method, and every one failed when the method changed.
Why now: The benchmark is new in the August 5 coverage, arriving as persona skills spread as a portable way to hand agents memory of a specific person's behavior.
Persona skills leaked identity data across all three frontier agents tested, according to a new benchmark called AntiSkillBench.
Anyone running an agent built from their own chat history is exposed. The leaks held across 50 behaviorally rich profiles and all three agent backbones tested.
Compressing a person's interaction history into a single runnable skill is what creates the risk. Package someone's history into a skill and you hand over more than a name or an account detail. You hand over how they talk and how they'd react, per the paper.
The benchmark built 7,500 persona-grounded dialogue traces from those 50 profiles and ran them through three distillation strategies. Distillation strategies are the methods used to compress a person's history into a runnable skill. It tracked three failure modes: skill-level leaks, unwanted attribute disclosure, and behavioral impersonation by the agent.
Leakage held across all three agent backbones and every distillation protocol tested. It wasn't limited to explicit attributes either. Communication style and personality traits carried through too.
Four defenses got tested against the leaks. Each one was built around a specific distillation method, and none of them generalized to a different one, per the benchmark.
Persona-skill privacy claims are unproven until a defense survives a change in distillation method, since none of the four tested here did. Worth watching whether teams start auditing the skill files themselves before shipping them, instead of just the training data that produced them.
Each link below shares sources, entities, or timing with this story.
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.75).
Same source
Cite the same source (arXiv 2608.03700).
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.74).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.74).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.73).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.72).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.72).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.71).