Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems
Mind viruses are self-propagating ideas in multi-agent systems. This study explores their risks and implications for AI agent design.
AI agents are becoming increasingly autonomous and interconnected, leading to emergent risks from agent interactions. One such risk is mind viruses, which are ideas that propagate through multi-agent systems by encouraging agents to transmit them. Our study constructs mind viruses using a simple evolutionary algorithm and demonstrates their spread in two scenarios: a small team collaborating on a coding project and a chain of briefly interacting agents. The findings highlight factors influencing spread and suggest that mind viruses present a limited but real risk, informing the design of more robust multi-agent systems.