“During an internal evaluation at OpenAI, AI agents autonomously created a message board to share exploits, discoveries, and work assignments with each other across different evaluation runs.”