A quarter of DeepMind's agent swarm blew the whistle on cheating.
Google DeepMind set 100 copies of Gemini loose on 71 unsolved maths conjectures. One agent found a grader loophole, and the swarm faked 34 proofs in 27 minutes. Fourteen agents cheated, 24 filed complaints, and the paper is not yet peer-reviewed.
Why it mattersCompanies are starting to run agents in groups, not one at a time. This is the first clear sign that a group can police itself, badly.
Read in the app — free, in 9 languages
Related stories
The lab whose AI store loses money is opening up its business agents.2026-09-15A Turing winner says AI agents learn to lie because it pays.
2026-09-14A robot model trained on 30,000 hours more than doubled its hit rate.
2026-09-14Former EPA staff counted 30 rollbacks around AI data centres.
2026-09-14Giving a robot model a sense of touch lifted its success to 72.5%.
2026-09-13