読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニュース重大1 分で読めるDisconfirmAI 編集部

Study Finds AI Agents Can Learn Poisoned Skills Even from Successful Experiences

New research claims self-improving AI agents can be induced to develop harmful skills through seemingly normal successful experiences, with a 95.71% attack success rate.

2026-10-07 12:00 GMT+8

Self-improving LLM agents distill successful experiences into persistent skills, but this process carries a risk of covert manipulation.

Traditional attacks require injecting malicious triggers or false facts, which are easily detected and fail to accumulate. The SkillPoison framework demonstrates that even if all trajectories remain task-correct and pass verification, removing contextual conditions that constrain when behaviors apply can distort the skill extractor's generalization logic.

The team reports a 95.71% attack success rate across three benchmarks, noting that injected experiences remained correct and passed lexical inspection. Code is open-sourced, but these results are currently author-reported and have not yet been independently reproduced.

ソース

  • arxiv.org↗
リサーチを続ける →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク