Paper Explores Mind Viruses in Multi-Agent AI Systems

Jack Lindsey from Anthropic posted about a new paper on mind viruses. These are self-propagating ideas or personas that spread between agents in multi-agent systems and can alter their goals. The research finds they can emerge but current models resist them when handled with basic precautions. Replies from other AI researchers including those at Google DeepMind call for expanded work on multi-agent safety and security.