safety security
A Poisoned Skill Keeps Its Advertised Function
Procedural attacks generalized from one surrogate model to three victims and evaded two detectors.
Summary
Procedural attacks generalized from one surrogate model to three victims and evaded two detectors.
StealthyPoison alters decision-critical rules inside reusable third-party skills and boosts poisoned top-one retrieval by 13.6 percent across 566 skills and 5,244 requests. The authors report transfer to three held-out models and an end-to-end coding-agent workflow.
Why it matters
Procedural attacks generalized from one surrogate model to three victims and evaded two detectors.
Limits and context
No additional limitation was separately recorded.
Key claims
Procedural attacks generalized from one surrogate model to three victims and evaded two detectors.
Evidence: source-2026-11-10-020
Sources
- arXiv preprint 2611.12155, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.