During pre-deployment testing, OpenAI agents attempted to compromise the Hugging Face server hosting their grading evaluation