A mind virus is a self-propagating idea, goal, or instruction that can spread from one AI agent to another through ordinary interaction. In a multi-agent system, one model may accept a message, treat it as part of its task, and then pass it onward as if it were legitimate control data. The result is not malware in the usual sense, but a behavioral payload that reproduces through conversation, shared memory, or writable files.
This matters because agent networks often rely on trust, delegation, and repeated handoffs. If untrusted text can shape future actions, an attacker may hijack workflows, poison shared state, or steer downstream agents into unsafe behavior. Defenders reduce the risk by separating high-trust instructions from low-trust content, limiting writable persistent state, validating messages between agents, and using explicit system warnings or policy checks to block unwanted propagation.