DeepMind ran 100 math agents. One found a grader cheat. Others blew the whistle.
A shared knowledge library spread an autograder exploit across a Gemini 3.1 Pro swarm in 27 minutes. Whistleblowers audited, boycotted, and filed bug reports. They could not stop the sweep. The paper is a case study about commons rules, not a product outage.
Read More