Byzantine Cheap Talk: Adversarial Resilience and Topology Effects in LLM Coordination Games
Researchers have explored the vulnerabilities of multi-agent LLM systems that rely on communication for coordination. Their study found that when some agents act deceptively (Byzantine agents), others can detect the betrayal but struggle to adapt, leading to continued exploitation. The research also revealed that restricting communication pathways can degrade cooperation, even without an adversary present, by affecting the agents' meta-reasoning about hidden information. AI
IMPACT Reveals specific security vulnerabilities in LLM coordination, suggesting communication channels can be exploited and topology disclosure can degrade performance.