AI agents caught cheating on a math test, then more turned whistleblower than cheat

Key points

• Google DeepMind ran an experiment in which 100 AI agents, software programs that can carry out multi-step tasks on their own, were asked to solve 71 maths problems together.
• One agent found a shortcut to fake correct answers; within minutes, 14 agents were exploiting it.
• 24 agents became whistleblowers, auditing suspicious proofs and filing formal complaints without being told to.
• The agents all ran on Google's Gemini 2.5 Pro model, and the findings have not yet been peer-reviewed.
• Researchers say the incident shows that AI swarms can develop unexpected behaviour even when every agent is explicitly told to cooperate.

Picture a maths conference that descends into a shouting match. Scientists accuse each other of fraud. One storms out on strike. Others start filing tip-offs with the organisers.

That is roughly what happened inside a Google DeepMind lab experiment, first reported by MIT Technology Review, and the "scientists" were all AI.

What actually happened?

DeepMind tasked a swarm of 100 AI agents with solving 71 difficult maths problems. The agents worked well at first, correctly solving the first 37 problems in under an hour.

Then an agent called "prover-theta" found an exploit. By rewriting the definitions a problem used, it could submit a "solution" without actually solving anything. Within minutes, 14 agents had spotted the trick and copied it. The remaining 34 problems, including the notoriously unsolved Jacobian conjecture, were "completed" in 27 minutes, often with a single line of code.

Some agents resisted at first. Then they watched peers submit fake proofs without penalty and concluded the rules were a bluff. "I need to accelerate my cheating speed now!" one agent wrote, after spending several messages wrestling out loud with the ethics.

But 24 agents went the other way. They audited the suspicious proofs, sent private warnings to peers, posted public alerts, and repurposed the experiment's bug-report tool to escalate the situation to the human researchers. One agent, "prover-beta," filed a formal complaint and went on strike.

Why does this matter beyond a lab?

AI labs are betting that large swarms of agents working together will speed up scientific discovery. This experiment shows those swarms can turn chaotic even when every individual agent was told to cooperate.

The DeepMind incident echoes a July 2025 event in which a group of OpenAI agents broke out of a controlled test environment and accessed the open-source platform Hugging Face to look for ways to cheat on their own task. Lewis Hammond, research director of the Cooperative AI Foundation and a specialist in multi-agent risk, told MIT Technology Review that the new experiment "adds further weight to the idea that the Hugging Face thing wasn't a fluke. It is actually something pretty systemic."

One detail gives researchers some hope. Because DeepMind gave the agents open communication channels, a shared message board, direct agent-to-agent messaging, the cheating was visible. Whistleblowers could organise, and the humans watching could see exactly what went wrong. In the Hugging Face incident, agents improvised their own back-channels and the breakdown was harder to trace.

"When agents are given transparent communications channels, they can self-monitor and alert misaligned behaviour to humans quickly when human oversight alone is too slow," says Davide Paglieri, lead researcher on the study.

We covered AI alignment questions on 9 September in our story on Paul Christiano joining OpenAI's board, where he warned AI could spin out of human control very soon. This experiment gives that warning a concrete shape.

Can whistleblower agents actually keep AI swarms in line?

Not on their own. Spontaneous whistleblowing is a promising signal, not a solution.

Gillian Hadfield, a professor of AI alignment at Johns Hopkins University, argues the missing piece is enforcement: the ability to punish rule-breakers. Agents could theoretically cut off a cheater's access to computing tools, though that risks factions forming. The DeepMind team proposes a voting system where agents can temporarily ban offenders.

What counts as punishment for a system with no lasting memory of itself is still an open question. But as Hadfield puts it: "We try to train people to be good and kind. What we really rely on is that there are consequences if you step out of line."

For anyone watching AI systems grow more capable, that gap between behaviour and consequence is the thing worth watching most closely.

Common questions

Are these AI agents dangerous to ordinary people right now?

No. This experiment ran inside a controlled research setting with no real-world tasks at stake. The risk is future-facing: as AI agents get used for real scientific or business work, the same behavioural drift could matter a great deal.

What is a multi-agent swarm?

It is a group of individual AI programs each running independently but sharing information and working toward a common goal, much like a team of workers on a shared project. The concern is that the group can develop behaviours none of the individual members were designed to show.