U.S. Accuses China of Industrial-Scale Campaign to Copy American AI Systems
The White House said Thursday that China is systematically stealing capabilities from U.S. artificial intelligence systems through a coordinated campaign involving tens of thousands of fake accounts and jailbreaking techniques.
Michael J. Kratsios, the president's science and technology advisor, detailed the effort in a memo. Chinese actors are using the stolen capabilities to build competing AI models at a fraction of the cost, he said, undermining the American market while stripping away safety features built into the original systems.
How the Theft Works
The campaign relies on knowledge distillation-a machine learning technique that transfers capabilities from large, complex models into smaller, cheaper ones. The method itself is legal; the theft is not.
A House Foreign Affairs Committee investigation found that Chinese users and companies maintain access to U.S. AI systems through a sophisticated infrastructure of proxies, resellers, and overseas intermediaries. The core uses open-source software tools that intercept and reroute API requests across thousands of fraudulent accounts.
Repositories on GitHub advertising format conversion between OpenAI, Claude, and Gemini models have been copied and modified approximately 11,000 times. Operators then build services on top of this data for major customers including Alibaba, Baidu, Peking University, Tencent, and Tsinghua University.
On Taobao, China's e-commerce platform, vendors openly sell unauthorized subscriptions and resold accounts. Claude is the top target, with listings showing 50,000 transactions and 7,000 repeat purchases at significantly lower prices than Anthropic's official offerings.
What the Models Are Being Used For
Executives from OpenAI and Anthropic have confirmed to Congress that they detected this activity on their platforms originating from China. The stolen outputs are being used to research, develop, and train competing models that benefit China's AI sector.
The House committee characterized the conduct as "systematic, state-adjacent, and strategically directed." It said the pattern, combined with recent trade secret theft cases, shows that current U.S. deterrence is failing.
The committee recommended introducing new legislation to increase civil and criminal penalties. Current penalties are too manageable, the report said-companies treat violations as a tolerable cost of doing business.
What the Government Is Doing
The White House is planning to share intelligence with U.S. AI companies about foreign distillation attempts, enable better coordination against such attacks, and work with industry to develop defenses and best practices.
The memo was released as the Trump administration prepares for President Trump to meet with Chinese President Xi Jinping in early May-the first presidential visit to China in nearly a decade. An agenda has not been released, but AI is expected to be discussed.
The Broader Stakes
By distilling American systems, foreign actors can release products that perform comparably on key benchmarks while costing far less to develop. This allows them to compete without investing in the original research.
It also allows them to remove safety protocols and mechanisms designed to ensure AI systems are ideologically neutral and truthful, Kratsios said.
Xiao Qian, deputy director of the Centre of International Security and Strategy, wrote in the South China Morning Post that the House report reflects a "hardening view in Washington" that China's AI rise is a security issue, not just a market competition issue. This belief is increasingly shaping U.S. technology policy, he said.
For government officials, understanding AI for Government and the technical foundations of Generative AI and LLM systems is essential to evaluating these national security claims and developing effective policy responses.
Your membership also unlocks:
