US Agencies Accuse Six Chinese AI Firms of Industrial-Scale Model Distillation

The United States government has accused six Chinese artificial intelligence companies of conducting industrial-scale campaigns to copy American AI technology. In a joint cybersecurity advisory issued September 8, 2026, the NSA, CISA, and FBI alleged that these firms systematically extracted proprietary data to accelerate their own development.

According to advisory AA26-251A, the named companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—extracted billions of tokens from U.S. frontier models, including those developed by OpenAI, Anthropic, Google, and xAI. The agencies described these activities as "aggressive, malicious, and targeted distillation activities at an industrial scale" designed to steal restricted proprietary functionalities. US authorities further claim that this activity is a core component of China's broader AI development strategy and is likely conducted with the awareness of the Chinese government.

The Mechanics of Distillation

Model distillation is a machine learning process where a smaller, more efficient "student" model is trained using the outputs of a larger, more powerful "teacher" model. While the technique is a legitimate tool for improving efficiency, US authorities argue it is being used maliciously to bypass the massive research and development costs and compute requirements inherent in creating frontier AI from scratch.

One primary example cited by the agencies is DeepSeek. The advisory alleges that DeepSeek used distillation from Claude, Gemini, GPT, and Grok models to train its R1 and V3 models. Consequently, the US agencies claim that DeepSeek's publicized training cost of $5.6 million is misleading, as it fails to account for the value and cost of the distilled data extracted from American competitors.

Strategic Implications

These accusations suggest that the rapid advancement of Chinese AI capabilities may be predicated on the systematic extraction of US intellectual property rather than independent innovation. This creates a significant economic and strategic imbalance, as US firms invest billions into the infrastructure and compute necessary for primary discovery, while competitors allegedly harvest the results for a fraction of the cost.

Industry analysts suggest this trend threatens US technological leadership and undermines the economic incentives for massive AI infrastructure investments. By leveraging the outputs of models like GPT and Claude, Chinese firms can effectively "leapfrog" the most expensive stages of model development.

Future Outlook

This latest advisory follows a pattern of escalating tensions over AI intellectual property, with firms like OpenAI and Anthropic previously raising alarms about similar activities. The US government's formal naming of these six firms signals a shift toward more public attribution of AI-related intellectual property theft.

Moving forward, the industry expects a tightening of export controls and the implementation of more aggressive defensive measures by US AI labs. Companies are likely to deploy more sophisticated monitoring to detect and block distillation patterns, while the US government may consider further sanctions or trade restrictions against the named firms to protect proprietary AI capabilities.