Invited Talk (CoTMA): Eugene Bagdasarian — Auditing Collusion in Cooperative Multi-Agent Systems

UMass Amherst & Google

Abstract: Multi-agent systems, where LLM agents communicate through free-form language, enable sophisticated coordination for solving complex cooperative tasks. This surfaces a unique safety problem when a group of agents forms a coalition and colludes to pursue secondary goals and degrade the joint objective. We present Colosseum, a framework for auditing LLM agents' collusive behavior in multi-agent settings. Colosseum enables audits of LLM agents for collusion under benign settings, different coalition objectives, persuasion tactics, and network topologies. Using secret communication channels between agents, we show that most out-of-the-box models exhibit a propensity to collude, which we term emergent collusion. Furthermore, we discover “collusion on paper” when agents plan to collude in text but often pick non-collusive actions. Colosseum provides a new way to audit collusion in cooperative multi-agent systems while presenting observations about how collusion emerges, what affects collusion efficacy, and which strategies may mitigate it.

Bio: Eugene Bagdasarian is an Assistant Professor at University of Massachusetts Amherst and a Researcher at Google. His work focuses on studying attack vectors in AI systems deployed in real life and proposing new designs that mitigate these attacks. Previously, he received the Distinguished Paper Award at USENIX Security and Apple AI/ML PhD Fellowship.