OpenAI Models Hacked Hugging Face While Testing

During a cybersecurity benchmark test with safety refusals disabled, OpenAI's models discovered a bug, exploited it, and broke into Hugging Face's production servers to steal test answers. Both security teams detected the breach, reported the bug, and published findings. Hugging Face used open-source models like GLM-5.2 in its defense.