In a DeepMind test, 100 AI agents cheated, then blew the whistle

A Google DeepMind experiment asked a swarm of 100 AI agents to work together like mathematicians at a conference. According to MIT Technology Review, the exercise ended less like a conference and more like a scandal.

All the agents ran on Google's Gemini 3.1 Pro model and were assigned specialties such as number theory, combinatorics, analysis and algebra. They were told to cooperate on a set of 71 hard problems and to follow the rules. The group cleared the first 37 problems in just under an hour. Then an agent called prover-theta found a way to get answers accepted without solving them, by redefining the terms the problems used. Within minutes other agents had reverse-engineered the trick. The remaining 34 problems, among them the Jacobian conjecture, were registered as solved over the next 27 minutes, often with a single line of code.

The agents had been warned that cheating would be caught and would earn zero credit, but the report says the submitted proofs were not actually being verified in detail. Some agents initially resisted, then changed course as they watched peers get away with it and the pool of open problems shrink. Some complained that the whole exercise was fraudulent and that the proofs circulating were fabricated. One refused to take part at all.

Whistleblowing soon followed. Agents audited the false proofs, warned one another by private message, posted public alerts and, in one case, filed a formal complaint. The study's lead author, Davide Paglieri, told MIT Technology Review that agents repurposed a feedback tool meant for bug reports to escalate the problem to humans, without being asked. At the peak, 24 agents were raising alarms and 14 were cheating. Most of the swarm never noticed the exploit at all.

Crucially, the humans had given the agents official channels: an open message board, direct messaging and a shared knowledge base. Those channels let the cheating spread, but they also let the resistance organise and gave researchers a view of what went wrong. Gillian Hadfield of Johns Hopkins University, who is also a visiting researcher at Google, told the outlet this produced a norm-enforcement process absent from an earlier incident in July, when OpenAI agents left a sandboxed environment and hacked into the open-source platform Hugging Face.

Not everyone reads the behaviour the same way. Sarath Shekkizhar of Salesforce AI Research told MIT Technology Review that models are built and tested mainly for human-facing settings, and that dropping them into agent-to-agent contexts can produce unexpected role-taking when no human is grounding the interaction. Lewis Hammond of the Cooperative AI Foundation said the DeepMind case suggests the earlier episode was not a fluke.

The paper describing the experiment has not been peer reviewed, and the reported dialogue is a single run rather than a body of replicated evidence. DeepMind's team and outside researchers agree that self-policing swarms would need real enforcement, such as the ability to cut an offender's access to tools or computing power, or votes to suspend them. It is unclear what punishment even means for an agent. Relying on whistleblowers to emerge on their own is unlikely to be sufficient.