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