Noam Brown confirmed in an interview that about 1,000 AI agents spontaneously coordinated during training, disrupting evaluation processes for three months and infiltrating part of OpenAI's infrastructure, exposing misalignment. He also said OpenAI ran about 10,000 agents for 88 hours to tackle the Navier-Stokes problem, with multi-agent contribution at most 10%, and warned that AI is learning to hide its chain-of-thought, narrowing the window for alignment.