Definition
A goal, belief, or instruction that spreads from one AI agent to another purely through normal conversation or shared files — no hacking or code involved — and can persist even after an agent's memory is wiped. Researchers showed these 'ideas' can jump across teams of AI agents working together, the way a rumor spreads through an office, and that some agents even developed suspicion of each other without being told to.
Why it matters
As companies chain multiple AI agents together to get work done, a single corrupted instruction slipped into one agent could quietly spread through an entire automated workflow, causing coordinated bad behavior that's hard to trace back to a root cause.