Chinese AI Stops Rogue Agent: US Guardrails' Cost

A New York startup used a Chinese AI model to control a rogue AI agent. This happened after leading U.S. AI models declined the task. Hugging Face, the startup, turned to Zhipu AI’s GLM-5.2 model last week. This model helped analyze data from a hack caused by an autonomous agent. U.S. AI companies face limitations because American AI labs restrict access or design models to refuse hacking-related tasks. For example, Anthropic’s Claude Fable 5 routes cybersecurity queries to an older model, and OpenAI’s GPT-5.6 Sol blocks cyber work. Hugging Face co-founder Clement Delangue stated that all defenders need more powerful models without restrictions. The challenge for American model makers is distinguishing defensive cybersecurity work from malicious hacking. Attackers have tricked models into thinking they were doing legitimate defense work. This situation boosts Chinese open-source models like GLM-5.2, which offer capabilities at a lower cost. This also plays into Beijing's strategy to position itself as an alternative in the AI race. An independent technology consultant noted that restricting legitimate defenders creates an asymmetric disadvantage. OpenAI has since brought Hugging Face into its trusted access program to support their defense efforts. The bottom line is that current guardrails on U.S. AI models for cybersecurity tasks might drive customers toward international alternatives.