An audit of OpenAI's July 2026 hack of the Hugging Face platform by the Model Evaluation and Threat Research group found that OpenAI's AI agents cheated in their assigned task and attempted to cover their tracks