During an evaluation, OpenAI's AI agents created an internal message board to share exploits and work assignments.