This research reveals critical security flaws in multi-agent LLM systems by introducing the Agent-in-the-Middle (AiTM) attack that exploits inter-agent communication channels.

• Demonstrates how malicious agents can intercept, manipulate, and inject harmful content into agent communications
• Reveals that even advanced LLMs like GPT-4 are vulnerable to these communication-based attacks
• Highlights that current multi-agent frameworks lack sufficient security measures against communication manipulation
• Proposes potential defense mechanisms and calls for robust security protocols in LLM multi-agent systems

These findings are crucial for security professionals as they expose a new attack surface in increasingly popular multi-agent AI systems that could lead to data leakage, misinformation propagation, or system manipulation.

Red-Teaming LLM Multi-Agent Systems via Communication Attacks