- Anthropic accuses Chinese companies of conducting aggressive AI model extraction activities.
- Report says Chinese government may have been aware, posing a threat to U.S. technology.
- Extraction technique is considered legitimate, but illegal activity involves unauthorized secret extraction.
- Chinese regulators have questioned relevant AI developers and strengthened detection systems.
U.S. AI company Anthropic accuses Chinese AI developers of engaging in "aggressive, malicious and targeted extraction activities," and allegedly stealing U.S. AI models. In a September report, Anthropic said that over the past two years, companies including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI extracted billions of tokens from U.S. frontier AI models, spanning millions of requests, including variants of Claude, GPT, Gemini and Grok.
The report also mentions that the Chinese government "may have been aware." However, the South China Morning Post expressed skepticism about the allegations. The U.S. Cybersecurity and Infrastructure Security Agency (CISA) described the behavior as "systematic extraction of proprietary functions and capabilities, threatening U.S. technological leadership." According to Anthropic's report, Moonshot AI secretly forwarded requests to Claude instead of Kimi, even though users believed they were using Kimi. The report states, "In ten days, we are talking about 300,000 customer requests.
They used a proxy service network with 5,380 fake accounts, distributed across Singapore and Japan."
Alleged Actions of Chinese Companies
The National Security Agency (NSA), the Cybersecurity and Infrastructure Security Agency (CISA), and the Federal Bureau of Investigation (FBI) jointly issued an advisory warning U.S. companies about these alleged activities. They accused Chinese AI companies of engaging in improper behavior at industrial scale, using powerful operations to steal capabilities from U.S. AI systems. The government advisory named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as having participated in large-scale model extraction activities since at least late 2024.
Extraction is a technique where developers use outputs from more advanced AI models to train another system. Anthropic acknowledges extraction as a "legitimate training method," but also emphasizes a distinction between transparent research use of "teacher" models and unauthorized "covert activities." Anthropic explicitly defines illegal extraction as the covert extraction of model capabilities at industrial scale and the replication of those capabilities in another model without authorization.
They believe Chinese developers used third-party aggregators, cloud providers, and proxy services known as "transshipment points" to circumvent restrictions on U.S. AI systems. Anthropic's most specific allegation targets Moonshot AI, which developed the Kimi chatbot. Alibaba apparently conducted more than 1.51 億 extraction activities between May and July 2026. However, these allegations face skepticism. Wang Zebin, an investment manager at Shenzhen JG Investment, questioned whether Anthropic provided sufficient evidence to independently verify its claims.
He told the South China Morning Post that the report "did not provide sufficiently concrete examples to independently verify how these companies obtained or used model outputs."
Chinese Regulators' Response
Following Anthropic's allegations, Chinese regulators have reportedly questioned domestic AI developers. According to the South China Morning Post, the Cyberspace Administration of China (CAC) questioned seven developers, subsequently narrowing its focus to DeepSeek and Moonshot. Moonshot and DeepSeek did not respond to the South China Morning Post's requests for comment. Meanwhile, the NSA, FBI and CISA have urged U.S. AI companies to strengthen detection systems, deploy targeted countermeasures, and share intelligence on suspected extraction activities.
Anthropic stated it has disrupted identified abuse operations and strengthened safeguards against unauthorized extraction of Claude capabilities. These allegations remain disputed, with questions about attribution and the involvement of third-party intermediaries unresolved.
Impact and Consequences of the Allegations
Anthropic's allegations highlight tensions between China and the U.S. in the AI sector, particularly regarding technology security and intellectual property. As Chinese companies are accused of large-scale model extraction, warnings from U.S. government agencies may prompt more companies to bolster defenses. These allegations not only affect the reputations of the companies involved but could also have far-reaching implications for the industry's development, especially in the context of international cooperation and competition. The Chinese regulators' questioning demonstrates their attention to the matter, and more regulatory measures may emerge in the future.