U.S. accuses China AI firms of industrial-scale model extraction

U.S. government agencies have accused Chinese artificial intelligence firms—including DeepSeek and Alibaba—of systematically harvesting capabilities from leading American AI models. The agencies described the activity as an industrial-scale operation aimed at reducing development costs and closing the technology gap.

In a joint advisory released on Tuesday, the Cybersecurity and Infrastructure Security Agency, the National Security Agency and the Federal Bureau of Investigation said the campaigns probably took place with the knowledge of the Chinese government.

According to the advisory, DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun and Z.AI have extracted billions of tokens across millions of exchanges from U.S. models since at least late 2024.

The agencies detailed that DeepSeek ran organized campaigns focused on reasoning capabilities and specialized functions to build its R1 and V3 models, while Alibaba used industrial-scale distillation to enhance its Qwen family of models.

The advisory noted that the companies gained access through multiple routes, including APIs, cloud providers, third-party aggregators, and a gray market of proxies known as "transfer stations," to bypass geographic restrictions and safeguards. They also used fraudulent accounts, bulk subscriptions and automated systems to evade detection.

The U.S. agencies said DeepSeek targeted models including Claude, Gemini, GPT and Grok, extracting capabilities such as reasoning, coding, agentic functions and question-and-answer optimization.

The agencies recommended that U.S. AI companies strengthen monitoring for anomalous usage, adjust their responses when suspected distillation attempts occur, and share threat intelligence among model providers, cloud platforms and API aggregators.