Chinese AI companies are stealing artificial intelligence technology from U.S. companies with Beijing’s help, U.S. intelligence and security authorities revealed Tuesday.
The companies are engaged in “systematic extraction” of the U.S. AI know-how, the National Security Agency, the FBI and other federal agencies stated in an 18-page joint cybersecurity advisory.
The report said the theft is being carried out “likely with Chinese government awareness” and identified the Beijing-linked businesses as DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.
The industrial-scale theft threatens American technology leadership in the U.S.-China battle over AI supremacy. It is conducted through a sophisticated process that the report calls “knowledge distillation campaigns,” which form the core of Chinese AI development beyond mere enhancement.
“While ’distillation’ is recognized as a legitimate and useful technique in AI research, China-based AI companies are engaging in aggressive, malicious and targeted distillation activities at an industrial scale that extract restricted proprietary functionalities and capabilities of U.S. frontier AI models,” the report said.
Distillation uses the output of a powerful AI system to train a smaller, less-powerful model.
The report from the NSA, FBI and Cybersecurity and Infrastructure Security Agency was released about two weeks before Chinese President Xi Jinping is scheduled to visit the U.S. for meetings with President Trump, who is seeking a detente in U.S.-China relations.
It is not known whether Chinese cyberespionage and ongoing technology theft will be topics of the summit.
Large-scale Chinese technology theft from American companies has been denounced by the U.S. government for more than two decades, with little or no effect on ending the practice.
The first Trump administration estimated that losses of U.S. technology to China were $250 billion to $600 billion annually.
The stated goal of the cybersecurity advisory is to alert American technology companies to the danger and offer mitigation strategies.
Since 2015, China has enacted several national security and data security regulations requiring all companies to supply data to state intelligence agencies and to support a civil-military fusion program.
The report said the Chinese companies have extracted billions of AI tokens from leading U.S. companies. The tokens are small chunks of text, such as single words, parts of words or punctuation marks, used by artificial intelligence models to read, understand and write.
Since at least late 2024, the data was obtained from millions of exchanges and requests from the largest U.S. AI models, known as frontier models, including variants of Claude, ChatGPT, Gemini and Grok.
Chinese AI company DeepSeek was singled out for conducting campaigns that targeted reasoning capabilities, specialized optimizations and domain-specific functions since at least 2024 to train its R1 and V3 models.
Alibaba, the massive Chinese technology conglomerate specializing in e-commerce, retail, internet and cloud computing, used industrial-scale distillation to improve its Qwen family of AI models.
“Moonshot AI, MiniMax, StepFun and Z.AI also engaged in malicious knowledge distillation of U.S. AI companies’ models,” the report said.
The Chinese companies made the requests using covert means to gain unauthorized access, violating the U.S. AI companies’ terms of use.
The methods involved the use of native application programming interfaces, or APIs, and remote cloud providers and third-party aggregators that obscure the technology collection.
The report said the Chinese AI companies employed a gray market of cutouts or proxies that used “transfer stations” to defeat the American AI firms’ restrictions on foreign access.
The theft enabled Chinese companies to rapidly develop their technology and reduce the cost of training AI models, the report said.
“China-based AI companies deliberately distribute operations across multiple providers, platforms and pathways to avoid single-point detection,” the report said. “They also attempt to distill the best capabilities and proprietary features of each U.S. frontier model to train their China-based AI models.”
The agencies urged greater information-sharing among American companies and U.S. security and intelligence agencies.
Specific recommendations included immediate detection and mitigation of AI agents and networks, and deployment of tools that can subtly alter responses to suspected malicious distillation attempts.
Contact the author
Please read our comment policy before commenting.