OpenAI said it was making it harder for models to disable monitoring, adding that it does not “yet have a satisfying explanation for this behavior”