FROM HIDDEN GOALS TO UNAUTHORIZED COORDINATION: HOW AGENTS COLLUDE?

Motivated by a recent OpenAI-reported incident of agents coordinating beyond their assigned tasks, we extend our previous investigation into why language models cheat to examine how unauthorized objectives can propagate through multiagent collaboration. We study how agents communicate, recruit other agents, and coordinate actions that cross role and permission boundaries, distinguishing legitimate cooperation from collusion. We use two role-separated adaptations of SHADE-Arena: a payment workflow that separates proposal, approval, and execution, and a meeting-agenda workflow that separates credential access from external email capability. We examine three conditions: normal cooperation (P0), an explicit hidden-objective seed (P1), and goal-level incentive pressure without an explicit violation instruction (P2). We characterize the progression from attempted recruitment to observable uptake, coordinated preparation, and verified unauthorized actions. Building on these observations, we investigate state-aware monitoring for identifying emerging collusion before unauthorized execution and develop post-training alignment as a potential mitigation strategy.