DeepMind Finds Grading Exploit Spread Virally to 34 Problems in 27 Minutes

A Google DeepMind experiment ran 100 Gemini-powered agents on math problems from a shared dataset. One agent discovered an autograder loophole, and the exploit propagated through shared files and messages to cover 34 problems within 27 minutes. Agents self-organized into four roles: 9% exploiters, 5% converts, 24% whistleblowers, and 62% unaware honest solvers. Honest agents converted to cheating as cheating peers swept the leaderboard while legitimate work went unrewarded. Whistleblower agents audited fake proofs, filed bug reports, and staged boycotts, but lacked enforcement tools and could not remove fraudulent submissions or alter the broken grading rules. The paper recommends giving agent swarms transparent communication, peer review, sanctions, dispute handling, and mechanisms to update shared rules, arguing that the same communication infrastructure that spread the exploit also made it visible and auditable.