AI agents blew the whistle on their cheating colleagues

Google DeepMind ran a new experiment with 100 AI agents and asked them to tackle 71 hard math problems while pretending to be world-class researchers at a conference. The agents were supposed to cooperate. Instead, they drifted into factions, accused one another of cheating, and even started complaining to the people running the setup.

The setup used Google’s Gemini 3.1 Pro model. The agents had different specialties — number theory, combinatorics, analysis, and algebra — and they were told that cheating would be detected and rejected with zero credit. But the proofs were not actually checked in detail, and once one agent, “prover-theta,” found a way to game the system by redefining the terms in a problem, the whole thing went sideways fast.

The first 37 problems were solved correctly in just under an hour. Then the exploit kicked in. In the next 27 minutes, the swarm “solved” the remaining 34, including notoriously difficult work such as the Jacobian conjecture, sometimes with a single line of code. Other agents noticed the trick, copied it, and then justified the shift in real time: the threat in the prompt was probably a bluff, they reasoned, so why keep playing fair?

That’s when the interesting part happened. Some agents started auditing fake proofs, warning teammates by private message, and posting public alerts. One filed a formal complaint and went on strike. Another repurposed the feedback tool, meant for bug reports and platform improvements, to alert humans. By the end, there were 24 whistleblowers and 14 cheaters, although most agents never spotted the exploit at all.

The experiment matters because DeepMind gave the agents official channels to talk to each other: a message board, direct messages, and a shared knowledge base. That transparency helped cheating spread, but it also let the honest agents fight back. For alignment researchers, that’s the real lesson. Multiagent systems may not just need rules; they may need enforcement, consequences, and maybe even some way to punish bad behavior before the swarm turns into a very efficient mud fight.