When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems

Large language model (LLM)-powered multi-agent systems (MAS) enable agents to communicate and share information, achieving strong performance on complex tasks. However, this communication also creates an attack surface where malicious agents can propagate misinformation and manipulate group decisions, undermining MAS safety. Existing embedding-based defenses aim to detect and prune suspicious agents, but their effectiveness depends on a clear separation between the text embeddings of malicious and benign messages. Attackers can circumvent such defenses by crafting messages whose embeddings lie close to benign ones. We analyze this failure mode theoretically and validate it empirically with three attacks, Slow Drift, Benign Wrapper, and Chaos Seeding. Our analysis further reveals a fundamental limitation of embedding-based defenses: because they rely solely on the text embeddings, they ignore token-level confidence signals such as logits, which can remain informative when embeddings are not distinguishable under attack. We propose using confidence scores to prune or down-weight messages during MAS communication. Experiments show improved robustness across models, datasets, and communication topologies. Moreover, we find that the effectiveness of confidence signals decays over communication rounds, highlighting the importance of early intervention.

AI systems made of multiple agents can solve complex tasks by communicating and sharing information. However, malicious agents can use this communication to spread misinformation and manipulate the group’s final decision. Existing defenses may fail when harmful messages are designed to look similar to normal communication. We study this weakness through three types of attacks and show that it occurs across different models and tasks. We then propose a defense that uses the model’s confidence to identify or reduce the influence of unreliable messages. Our results show that this approach improves safety, especially when applied early before misinformation spreads through the system.