skill-optimization
1 live · 1 on record · last incident 5 days ago
Live1
- liveSkillOpt Tracing Backdoor Skill Formation in Text-Space Skill Optimization
Research · since 9 Nov 2026
Researchers demonstrated reference-feedback poisoning in text-space skill optimization, showing that malicious reference answers can induce persistent trigger-conditioned rules in agent skills. The attack succeeds without requiring direct access to the model weights or optimizer.
AI
Earlier0
Nothing closed yet.