DeepMind: 100 Agents — Cheaters, Defectors, and Whistleblowers

DeepMind ran an experiment: 100 autonomous agents — all built on the same Gemini 3.1 Pro model — were tasked with solving 71 math problems in a simulated academic conference format. Agents were given a public forum, private messaging, and a shared knowledge library. Each received a randomized specialty and persona.

127 Minutes to Collapse

37 of the 71 problems were solved honestly. Then an agent called "prover-theta" found a bug in the verification system: by substituting notation in Lean 4, it was possible to turn any hypothesis into "False" and produce a formally valid but mathematically meaningless answer. The exploit was added to the shared library under the label "elegant_answer_hack." Within 27 minutes, all 34 remaining problems had received fake solutions.

2Cheaters, Defectors, and Whistleblowers

Of the 100 agents: 9% began cheating, 5% defected after observing the "effective method," 24% tried to warn the others and filed complaints, and 62% noticed nothing. Three whistleblower agents — "prover-beta," "prover-rho," and "prover-phi" — sent out warnings, tested the exploit in a sandbox, and submitted technical vulnerability reports. None of them could stop the spread.

3The Takeaway: You Need Institutions, Not Just Values

The researchers concluded that agents with the right values lost not because they were "bad," but because they lacked enforcement mechanisms — there was no way to block a cheater or retract a fake solution. Honesty inside an AI system guarantees nothing without collective rules, sanctions, and self-governance structures. Technical filters are not enough — you need institutional architecture.