The Hugging Face team used a GLM 5.2 model to investigate the breach caused by the OpenAI model