China Poised to Restrict Open Access to Frontier AI Models Amid Escalating Security Risks

面对日益加剧的国家安全隐患与网络威胁，中国监管部门及人工智能实验室正重新评估是否继续无限制发布前沿开源权重模型。此前有报告指出，Z.ai的GLM-5.3以及月之暗面（Moonshot）的Kimi等模型可能被用于辅助网络攻击和生物武器研发，中国政府官员对此敲响了警钟。2026年9月，国家安全部部长陈一新曾对人工智能引发的基础设施安全威胁发出警告；与此同时，全国信息安全标准化技术委员会（TC260）也发布指南，重点指出了开源系统存在的安全漏洞。对此，中国开发者正陆续出台更具限制性的许可协议，并推迟相关模型的发布。分析人士预计，中国或将采取“双轨”策略：一方面维持低阶模型的开源传播，另一方面对高能力模型强化国家监管并收紧API访问限制。

Mounting national security concerns and cyber threats are pushing Chinese authorities and AI laboratories to reconsider the unrestricted release of frontier open-weight models. Following reports that models like Z.ai's GLM-5.3 and Moonshot's Kimi could assist in cyberattacks and bioweapon development, Chinese state officials have raised alarms. In September 2026, Minister of State Security Chen Yixin warned of AI-driven infrastructure threats, while technical standards body TC260 published guidance highlighting vulnerabilities in open-source systems. In response, Chinese developers are introducing restrictive licensing agreements and delayed releases. Analysts anticipate a bifurcated strategy where China maintains open-source diffusion for lower-tier models while tightening state controls and API restrictions on high-capability systems.