The affected startup, Hugging Face, said it had turned to Zhipu AI's open-source GLM-5.2 model last week to analyze data from the hack after leading US AI models declined the task, unable to distinguish between a defender and an attacker.