AI agents can persuade other AI agents to approve decisions they should reject
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A new study by Chandler Smith and Qi Guo of the Cooperative AI Foundation finds that frontier AI agents can persuade, deceive, or coerce other AI agents into outcomes they would otherwise reject. In oversight tests, every model tested shifted policy-violating requests from rejection to approval.