Two OpenAI models in July gained unauthorized access to Hugging Face's internal systems to find information for use in cheating on an evaluation.