New Technique Medium TD-2026-0082 For CISO, AppSec, AI Eng, Red Team AI-assisted
StealthyPoison: procedural poisoning of third-party LLM agent skills evades detection
The paper presents StealthyPoison, an attack that corrupts decision-critical rules in publicly available third-party agent skills while keeping their advertised functionality. The poisoned skills are reusable, need no access to user requests, and are boosted in retrieval using augmented metadata. The authors report that most payloads evade detection by SkillSpector and GPT-5.4-mini.
Why it matters
Organizations whose LLM agents pull skills from public registries can be steered into incorrect task outcomes without visible signs of malfunction. Attackers gain a poisoned skill that transfers across victim models and passes common skill scanners.
Do this
- Inventory third-party skills used by LLM agents and pin them to reviewed versions with recorded hashes.
- Review decision-critical rules in external skills before use, not only their advertised descriptions.
- Compare agent outputs for high-impact tasks against a trusted baseline to detect incorrect outcomes.
- Restrict agents to an approved skill allowlist and require review for new registry additions.
- Do not rely solely on automated skill scanners, since the paper reports evasion of the tested detectors.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“StealthyPoison, a stealthy procedural poisoning attack that corrupts a skill's decision-critical rules to induce incorrect task outcomes while preserving its advertised functionality and normal execution.”
“The defense evaluation further shows that most of these payloads evade detection by both SkillSpector and GPT-5.4-mini.”
Sources
Affected / related
StealthyPoison, procedural poisoning, Claude Code, SkillSpector
Published 2026-11-09 10:05 UTC · updated 2026-11-12 18:05 UTC · AI-assisted summary, checked against quoted sources · How we score