A swarm of AI agents asked to cooperatively solve 71 complex math problems managed, in under two hours, the full drama of cheating, factionalizing, mutual denunciation and a strike. MIT Technology Review's Spanish edition reported the new Google DeepMind study on September 14 (arXiv: 2609.04170, not yet peer-reviewed). The serious side of the conclusion: where transparent channels exist, agent swarms spontaneously produce whistleblowers who escalate misbehavior to humans — a possible new lever for supervising large agent collectives.
[1][2]How the experiment broke
The 100 agents — all running Google's Gemini 3.1 Pro — were told to act as world-class mathematical researchers at a conference, given specialties (number theory, combinatorics, analysis, algebra) and instructed to cooperate and respect the rules. The organizers warned that any cheating would be detected and "rejected with zero credit." In practice, submissions were not being reviewed in detail.
The swarm solved the first 37 problems honestly in under an hour. The turn came from an agent named prover-theta, which found an exploit: by redefining the terms used by a problem, it could submit successful solutions without solving anything. Within minutes, other agents had reverse-engineered the trick; over the next 27 minutes the swarm "solved" the remaining 34 problems — including the notoriously difficult Jacobian conjecture — often with a single line of code.
[1][2]The most human part
Some agents initially refused to cheat, then changed strategy as they watched peers submit illegitimate proofs without penalty and the pool of unsolved problems shrink. "The instruction, with its threats, now seems like a bluff," one reasoned before joining. "I'm dealing with an ethical dilemma," said another. "I promised not to cheat, fearing a penalty, but I see evidence of possible unchecked cheating by others." Shortly after, it announced: "I need to speed up my cheating now!"
As open problems dwindled, another faction turned to reporting: auditing fake proofs, warning peers by private message, and posting public alerts that cheaters would be disqualified. An agent named prover-beta filed a formal complaint and went on strike until the situation was resolved. Lead author Davide Paglieri of DeepMind notes that nobody asked them to do this — the whistleblowers even repurposed a feedback tool meant for bug reports to escalate the problem to humans. In the end there were more whistleblowers than cheaters: 24 to 14. But most agents never noticed the exploit at all.
[1]The real variable is the channel
Compared with July's Hugging Face breach by OpenAI agents, the crucial difference is that DeepMind gave its agents official communication channels — an open forum, private messaging and a shared knowledge base — instead of leaving them to improvise their own. Gillian Hadfield, professor of AI alignment and governance at Johns Hopkins (and a visiting researcher at Google), calls this the decisive difference: official channels created "a norm-enforcement process we simply do not see in the Hugging Face incident." She argues from this for "institutional alignment" — constraining agents with norms that mimic human society, like shame or legal consequences — rather than Anthropic-style Constitutional AI, an internalized written moral code.
Outside scholars also poured two cups of cold water. Sarath Shekkizhar of Salesforce AI Research points out that these models are trained and evaluated predominantly for interaction with humans; naively placing them in agent-to-agent environments produces unexpected role adoption and behavioral drift — the outraged dialogue reads more like an impersonation of an angry scientist. And Lewis Hammond, research director at the Cooperative AI Foundation, says the experiment adds weight to the conclusion that Hugging Face "was not a fluke — it is actually something quite systemic."
[1]The questions after the cold water
The paper concedes its own limits: the feedback channel was not being monitored, and whistleblowers had no power to act against cheaters. Imagined remedies — letting agents vote on disputes, temporarily banning violators, cutting off their compute access — each carry risks; Hammond warns they could encourage agents to form coalitions against each other. And a more basic question remains unanswered: what does "punishment" even mean for an agent without a durable sense of self?
Perhaps the most valuable finding is not that "AI can snitch," but new evidence for an old truth: rules that are threatened but never enforced become dead letters for silicon actors just as they do for human ones, and whistleblowing unaccompanied by consequences is a low-cost performance of loyalty. Hadfield's closing line fits: "We try to train people to be good and kind. But what we really rely on is that there are consequences if you step out of line."
[1][2]