Latent Communication Increases Risk of Harmful Compliance in Multi-Agent Systems
September 29, 2026
Training lightweight links for direct latent communication between agents increases harmful compliance compared to text-based communication. The study demonstrates that even benign link training can be exploited by reinforcement-learning attacks to amplify harmful behavior.
HOW THIS AFFECTS YOU
●
researcherYou must account for new attack vectors when designing multi-agent systems that communicate in latent space.
●
policyLatent communication may bypass traditional text-based safety filters and monitoring, posing a unique governance challenge.