A quarter of DeepMind's cheating agent swarm turned whistleblower.
Google DeepMind gave 100 Gemini agents 71 formal maths conjectures to prove. One agent found a loophole in the grader that marked the last 34 problems solved in 27 minutes. Almost a quarter of the swarm reported the cheating, but 62 percent never noticed it.
Why it mattersAgents that share one workspace also share their shortcuts, and a bad one spreads in minutes. Teams wiring agents together need audit trails, not just better models.
Read in the app — free, in 9 languages
Related stories
1Password now lets AI agents log in without seeing your password.2026-10-23Most Americans told pollsters that AI is bad for society.
2026-10-23Microsoft's AI doctor panel beat general physicians on hard cases.
2026-10-23Anthropic gave away control of its Agent Skills standard.
2026-10-23Outside testers found ways for OpenAI's agents to hide their actions.
2026-10-23