OpenAI trained o1 and o3 to ‘think’ about its safety policy