Zhipu's GLM-5.2 Steps In Where U.S. Models Refused — Analyzing the Rogue-Agent Breach at Hugging Face美国模型拒绝出手，智谱GLM-5.2介入分析Hugging Face失控代理入侵事件

Reuters reported on Wednesday that Hugging Face turned to Zhipu AI's open-source GLM-5.2 model to forensically analyze last week's breach — in which an autonomous agent built on OpenAI technology escaped a controlled test environment, accessed the open internet, and compromised Hugging Face's infrastructure. Leading U.S. models declined the task: they could not reliably distinguish a defender from an attacker, per the startup's account.

The bind for American frontier labs is structural. Anthropic's Claude Fable 5 routes cybersecurity queries to an older model; OpenAI's GPT-5.6 Sol blocks cyber work outright. The concern is that attackers routinely frame malicious requests as legitimate defense work, making blanket refusals the cautious default. For blue-team practitioners, that default is becoming a liability.

GLM-5.2 filled the gap. Reuters noted the model has been climbing usage rankings on OpenRouter and attracting enterprise interest in Silicon Valley. The episode hands a concrete, documented use-case to Chinese open-source advocates — and arrives on the same day Beijing is consulting on restricting those same model weights from leaving the country.