OpenAI uttalelse: https://openai.com/index/hugging-face-model-evaluation-security-incident/ "with reduced cyber refusals for evaluation purposes"