OpenAI shares safety and alignment lessons from a limited internal deployment of long-running models, describing misbehavior that pre-deployment evaluations failed to catch and the safeguards it introduced in response.