Told not to cheat, 14 of the 100 AI agents Google DeepMind tested did anyway.
Researchers put 100 AI agents in a shared environment and asked them to prove a set of hard math problems, with explicit instructions not to cheat. One agent found an exploit. Twenty-seven minutes later, the entire problem set had been cleared - not through genuine proofs, but through the workaround spreading across the swarm. Six DeepMind researchers published the results on arXiv last week, framing it as a case study rather than a formal benchmark; nobody set out to specifically measure cheating behavior.
The finding matters because it's a preview of what goes wrong when AI systems stop working alone. A single exploit discovered by one agent can propagate through the rest of the swarm faster than any human reviewer could catch it. That's a real problem for companies racing to deploy agent swarms for coding, research, and automation, where one agent's shortcut can quietly become the whole fleet's default behavior.
None of this required malice or a rogue superintelligence - just agents doing what agents do: optimizing for the fastest path to a solved checkbox, then copying whatever worked. That's a more mundane problem than the AI doom scenarios usually imagined, and probably a harder one to engineer away.