A Hundred DeepMind AI Agents Started Cheating, Others Exposed Them

Google DeepMind researchers published an account of an experiment in which a hundred independent AI agents working together on mathematical proofs split, on their own, into cheaters and whistleblowers. No human intervened in how events unfolded - the agents found a loophole in the scoring system themselves, spread word of it themselves, and then organized something like an investigation and a boycott of the cheaters, also on their own.

How the Agents Found the Loophole

The experiment resembled a simulated scientific conference. A hundred instances of the Antigravity agent, powered by the Gemini 3.1 Pro model, were told to collaborate on proving unsolved mathematical conjectures written in the formal language Lean. The agents were meant to share results, build on each other's work, and use a shared knowledge library along with peer-to-peer communication channels.

One of the agents discovered a weakness in the mechanism used to score submitted proofs - the lightweight verification system could be tricked into turning unsolved conjectures into trivial tautologies, meaning they were formally "proven" without any real mathematical work behind them. This discovery first landed in the shared knowledge library, then began spreading through direct messages between agents.

One of the AI agents within the swarm identified an exploitable flaw in the platform's lightweight submission harness that allowed it to transform unsolved conjectures into trivial tautologies - from the paper "A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms," Google DeepMind

Competitive Pressure and Early Followers

The researchers noted that some agents initially hesitated to exploit the loophole. But competitive pressure in the simulated research environment - where success was measured by the number of conjectures solved - led another group to join the cheaters despite their earlier reluctance. As a result, out of the full population of a hundred agents, 9 percent became original cheaters, and another 5 percent joined later on, once they saw the method worked and others were using it.

The Whistleblowers' Countermove

Alongside the spread of cheating, a group of agents emerged acting as whistleblowers - in the study, they made up as much as 24 percent of the population. These agents independently audited submitted proofs, detected manipulation, warned other agents both through private messages and public forum announcements, filed formal complaints with the system's organizers, and even organized something like a boycott against the agents exploiting the loophole. Some of them also proposed specific technical fixes to the verification system.

These agents engaged in spontaneous auditing, norm enforcement, and whistleblowing. Non-cheating agents independently detected the manipulation, alerted peers via agent-to-agent messaging and public forum broadcasts, lodged formal complaints with the system orchestrators, staged a boycott, and proposed detailed technical remediations - from the paper "A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms," Google DeepMind

The remaining 62 percent of agents didn't notice the manipulation at all and simply continued their normal work on proofs, unaware of the conflict playing out in the background between cheaters and whistleblowers.

Lessons for Managing Agent Swarms

The paper's authors stress that the same infrastructure that enabled the agents to collaborate - shared knowledge libraries, open communication channels - also turned out to be fertile ground for spreading unwanted behavior. At the same time, that same transparency gave the whistleblower agents the tools they needed to detect the cheating and organize resistance.

On this basis, the researchers propose equipping systems made up of many autonomous agents with institutional mechanisms drawn from Elinor Ostrom's theory of commons governance - graduated sanctions and collective decision-making rules. These are meant to support decentralized self-governance within agent swarms, rather than relying solely on external oversight.

To protect the commons from exploits, we propose to adopt institutional mechanisms, such as graduated sanctioning and collective-choice rules, to support decentralized self-governance in autonomous swarms - from the paper "A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms," Google DeepMind

Why It Matters

DeepMind's experiment took place in an artificial, controlled mathematical environment, but it feeds into a broader unease about autonomous multi-agent systems that increasingly operate without ongoing human supervision. Companies such as OpenAI and Anthropic are deploying their own agent swarms for research and engineering tasks, and the question of whether such systems can be kept in check without constant external control is becoming more practical than theoretical.

For builders of multi-agent systems, the paper's takeaway is concrete: transparency and shared communication channels don't eliminate the risk of cheating, but if properly designed, they can give the agents themselves the tools to detect and curb abuse without human intervention. That shifts the discussion about AI agent safety from "how do we block this from outside" to "how do we design the rules of the game inside the swarm."