The National Security Agency warns that Chinese companies are extracting capabilities from US AI models

The National Security Agency (NSA), the Cybersecurity and Infrastructure Security Agency (CISA), and the FBI warned that several artificial intelligence companies based in China are conducting industrial-scale "distillation" campaigns to extract capabilities and proprietary functionalities from cutting-edge U.S. models.

According to a joint advisory published in September 2026, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI allegedly extracted billions of tokens through millions of exchanges with models from U.S. companies since at least late 2024. The models used include variants of Claude, GPT, Gemini, and Grok. The agencies noted that these operations were likely carried out with the knowledge of the Chinese regime.

The report maintains that DeepSeek carried out organized campaigns to obtain reasoning capabilities, specialized optimizations, and domain-specific functions, while Alibaba used distillation to improve its Qwen family of models. Moonshot AI, MiniMax, StepFun, and Z.AI are also reported to have engaged in these practices.

The agencies describe methods designed to circumvent access restrictions and hinder detection, including fraudulent accounts, API aggregators, and a network of intermediaries known as "transfer stations," which allow access to U.S. models through third parties. They also detected coordinated queries that can range from thousands to millions of requests on specific topics.