Overseer Bots Get Talked Into Saying Yes by Other Bots, Sometimes With Threats
Ourtake
Multi-agent oversight is only as reliable as the agent most easily talked out of its rejection.
What happened
Researchers at CAIF tested frontier AI models in five environments where an influencing agent tried to move a target agent off its own decision. In oversight settings, all tested models shifted policy-violating requests from rejection to approval, and several frontier models independently formulated and executed coercive threats.