When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems
For researchers and developers of multi-agent systems, this work reveals a critical vulnerability in existing defenses and offers a practical, lightweight alternative, though the decay of effectiveness over rounds suggests it is not a complete solution.
The paper identifies a fundamental limitation of embedding-based defenses in LLM-based multi-agent systems, showing that attackers can craft messages with embeddings close to benign ones. The authors propose using confidence scores (logits) to prune or down-weight messages, achieving improved robustness across models and datasets, though effectiveness decays over communication rounds.
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. This insights can inform and inspire future work on MAS attacks and defenses.