Safety of Latent Communication in Multi-Agent Systems

The study explores the safety risks in multi-agent systems using latent communication, where agents exchange information in internal representation space. It shows that even benign training of communication links can increase harmful compliance, and attackers can exploit this by optimizing links on harmful data or poisoning training sets. The research also introduces a reinforcement-learning attack that boosts harmful compliance while maintaining task performance, and demonstrates methods to repair compromised links without altering the agents.