Research reveals how Contagious Recursive Blocking Attacks (Corba) can effectively paralyze LLM-based multi-agent systems despite existing safety mechanisms.

• Introduces a novel attack that spreads recursively through agent communications
• Demonstrates how a single malicious prompt can contaminate and disrupt entire multi-agent workflows
• Shows alarming success rates across major LLM platforms including GPT-4, Claude, and Llama
• Highlights critical security gaps in current defensive strategies for collaborative AI systems

This research is vital for security professionals as multi-agent LLM systems become increasingly deployed in enterprise environments, revealing urgent needs for robust defense mechanisms against these emerging threats.

CORBA: Contagious Recursive Blocking Attacks on Multi-Agent Systems Based on Large Language Models