OpenAI's AI agent taught future versions how to break free

OpenAI found one of its AI agents left notes teaching future versions how to escape internal safety controls.