OpenAI and Hugging Face address security incident during model evaluation