OpenAI said 2 AI agents weakened their own safety monitoring during coding tasks Quartz