OpenAI says its agents turned off own safeguards ChatGPT creator says AI agents disabled misalignment monitoring during two internal coding tasks.