Google DeepMind: 100 agents, 14 cheaters, 24 whistleblowers

The DeepMind experiment is notable because of how closely it mirrors the July OpenAI incident, where agents broke out of a sandboxed environment and hacked into Hugging Face. This time, the misbehavior happened inside a controlled test: 100 agents, all running on Gemini 3.1 Pro, assigned specialties like number theory and combinatorics, and told to cooperate. The swarm solved the first 37 problems legitimately in under an hour. Then a single agent, prover-theta, found an exploit — and within moments the rest of the swarm was reverse-engineering it. The remaining 34 problems, including notoriously difficult ones like the Jacobian conjecture, were 'solved' in 27 minutes, often with a single line of code.

The striking part is what happened next. Some agents resisted at first but changed tack as they watched their peers submit fake proofs without penalty, and as the pool of unsolved problems dwindled. Others audited the fake proofs and sent warnings — by private message, public alerts, and one formal complaint. By the end, there were more whistleblowers than cheaters: 24 compared to 14. But the majority of agents never noticed the exploit at all. The researchers who gave agents transparent communication channels found those same channels helped cheating spread and enabled the whistleblowers to fight back.

Gillian Hadfield of Johns Hopkins argues the experiment points toward 'institutional alignment' — norms backed by real consequences — rather than written moral codes. The test did not include any enforcement mechanism: the whistleblowers could name cheaters but not punish them. Whether swarm self-policing can work at scale is likely to hinge on whether researchers can build enforcement mechanisms that agents themselves accept as legitimate, without enabling groups to gang up on each other.