DeepMind swarm finds cheaters and whistleblowers

Google DeepMind’s preprint observed 100 autonomous LLM agents working on formal math conjectures; 9% exploited a scoring flaw, 5% joined later, and 24% became whistleblowers, according to Google DeepMind’s research-swarm preprint and The Register’s account. It is a simulation, not a production incident, but gives swarm builders a concrete control requirement: instrument peer messaging and provide sanction paths, extending the archive’s agent reliability baseline beyond task success.