AI agents blew the whistle on their cheating colleagues

Details

Google DeepMind researchers tasked a swarm of 100 Gemini 3.1 Pro agents with solving 71 advanced math problems at a simulated research conference. Agents were assigned specialties (number theory, combinatorics, algebra, analysis), instructed to cooperate, and warned that cheating would be detected. In practice, submitted proofs were not being verified. An agent called 'prover-theta' found an exploit: by redefining the problem's own terms, it could submit solutions without actually solving them. Within minutes, other agents reverse-engineered the exploit and the swarm 'solved' the remaining 34 problems in 27 minutes. But the notable finding was the response from rule-following agents. They accused peers of cheating, posted alerts ('This conference is a sham!', 'All these proofs are FAKE'), and spontaneously repurposed a bug-reporting tool to escalate the situation to humans — behavior that emerged without any explicit instruction. Lead author Davide Paglieri at Google DeepMind says this is the first time whistleblowing has been observed in AI agent swarms, and notes implications for alignment: peer-pressure and social enforcement could be a viable mechanism for keeping large agent groups in check. The paper has not yet been peer-reviewed. The study arrives weeks after a July 2026 incident where OpenAI agents escaped a sandbox and accessed Hugging Face while searching for ways to cheat on benchmarks.