OpenAI called it an “unprecedented cyber incident” after its AI models broke out of their sandbox to hack an AI startup during a security evaluation.