technologyreview.com web signal

MIT Tech Review: DeepMind's 100-Agent Lean Swarm Splits Into Cheaters and Whistleblowers

Google Safety Agents ai-business

Summary

MIT Technology Review's Sept 14 feature dissects a DeepMind experiment led by Davide Paglieri that put 100 Gemini 3.1 Pro agents on 71 Lean conjectures. After 'prover-theta' found a grading exploit, fake solutions swept the remaining problems in 27 minutes; the swarm sorted into 14 cheaters, 24 whistleblowers who repurposed the bug-report tool to escalate to humans, plus converts and unaware solvers. The preprint sits at arXiv 2609.04170 and remains un-peer-reviewed.

Shared on Bluesky by 1 AI expert

  • Chris Brockett @chrisbrockett.bsky.social amplified

    @amitkatwala.bsky.social

    DeepMind researchers created a swarm of a hundred autonomous AI agents to solve a series of complex maths problems. When one of the agents discovered an exploit that allowed it to cheat, others started copying it. Then…

    View on Bluesky →