Your model already knows the answer: how benchmark answers leak into LLMs