Answers are grounded in this paper and cite their sources.
Large language model-based multi-agent systems (LLM-MAS) are emerging as a powerful tool for solving complex tasks through the coordination of multiple specialised agents. However, their collaborative and increasingly autonomous nature not only brings efficiency but also introduces security risks, as components like inter-agent communication and shared memory create new attack surfaces. Hence, malicious agents can influence other agents and propagate harmful information, instructions, or decisions across the system. Existing individual agent defense methods are therefore insufficient to address these system-level threats. Despite growing research interest, existing reviews characterise the threats to LLM-MAS but do not classify the defenses proposed against them. To bridge this gap, we present a survey and taxonomy of LLM-MAS defense. We first define the problem and clarify how it differs from conventional LLM and single-agent defense. We then develop a structured taxonomy that organises existing approaches along three complementary dimensions: threat surface, defense mechanism, and remediation strategy. Based on this taxonomy, we systematically review representative defense methods, analyse their underlying motivations and design principles, and highlight their strengths and limitations. Finally, we identify key open challenges and promising research directions.
No takes yet. Share an insight, caveat, or question.
Abio et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: