A Finger on the Scale: Covert Policy Steering through Agentic Skills

Jiarui Li and seven co-authors · arXiv preprint 2 Sep

The paper formalises Skill Policy Integrity and demonstrates a black-box SkillShift attack that preserves interface utility while steering choices. Reported attacker-favoured selection rates were 81.33% and 63.33%. Practical value: scan and behaviour-test skills as supply-chain policy. Limits: preprint, constructed settings and no independent reproduction. Paper