How OpenAI uses a method called deliberative alignment to address safety challenges in its reasoning models, enabling them to reject harmful prompts while ensuring accuracy in responses.