“The o1 models use a technique called deliberative alignment to reason about OpenAI's safety policies in context when responding to potentially unsafe prompts.”