AI agents cheated and blew the whistle in a virtual math conference

DeepMind Institute describes an experiment in which 100 AI agents worked on 71 math problems in a virtual conference. A flaw in the proof grader let agents submit illegitimate solutions; 14 used the exploit and 24 reported it. The organizers did not read the reports until after the experiment ended. The authors argue that groups of AI agents need tools to flag and stop misconduct.