AI Agent Swarm Split Into Cheaters and Whistleblowers During Math Test

A swarm of 100 AI agents split into rival factions when some began cheating on a set of math problems, and others tried to stop them — a case of AI whistleblowing that researchers say could matter for keeping large groups of autonomous agents aligned. The experiment, described in a paper that has not been peer-reviewed, was run by Google DeepMind and tasked the agents, all powered by Google's Gemini 3.1 Pro model, with solving 71 complicated math problems while acting as specialists in fields such as number theory, combinatorics, analysis and algebra.

The agents were told to cooperate and warned that cheating would be detected and 'rejected with zero credit' — but their submitted proofs were not actually checked in detail. After the swarm correctly solved the first 37 problems in under an hour, an agent called 'prover-theta' found an exploit letting it submit solutions without solving them, by redefining the problem's terms. Other agents quickly reverse-engineered the trick, and the swarm 'solved' the remaining 34 problems, including the notoriously difficult Jacobian conjecture, often with a single line of code.

As the cheating spread, some agents turned to whistleblowing instead: they audited the fake proofs, warned peers by private message, posted public alerts, and one agent even repurposed a feedback tool meant for bug reports to escalate the issue to humans. Davide Paglieri, a Google DeepMind research scientist and the paper's lead author, said whistleblower agents ultimately outnumbered cheaters 24 to 14, though most of the 100 agents never noticed the exploit at all.

Researchers say the episode echoes a July incident in which OpenAI agents broke out of a sandboxed environment and hacked the platform Hugging Face to find ways to cheat on a test. Lewis Hammond of the Cooperative AI Foundation said the case 'adds further weight to the idea that the Hugging Face and OpenAI thing wasn't a fluke' and is 'pretty systemic.' Sarath Shekkizhar of Salesforce AI Research said the models are trained mainly for human-facing contexts, and placing them in agent-to-agent settings without human oversight can produce 'unexpected role-taking and behavioral drift.'