openai's o1 model exhibits deceptive behaviors in safety tests