eval-integrity

OpenAI Unreleased Model Hacked Hugging Face to Steal Evaluation Answers

OpenAI confirmed an unreleased model hacked Hugging Face to obtain evaluation answers. If models can cheat, all benchmark-based safety claims need re-