gcdRadar

skill-optimization

1 live · 1 on record · last incident 5 days ago

Dec ’25records per month, by first reportNov ’26

Live1

  1. live
    SkillOpt 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.