US artificial intelligence developer Anthropic has accused three Chinese AI companies of attempting to extract proprietary capabilities from its Claude large language model through mass querying, claiming this could undermine intellectual property protections and create security risks.
In a detailed blog post, Anthropic said that DeepSeek, Moonshot AI and MiniMax generated more than 16 million interactions with Claude by creating around 24,000 fraudulent accounts. The company said these exchanges were aimed at applying a technique called distillation to improve the Chinese firms’ own models by leveraging Claude’s outputs.
Distillation is a recognised method in machine learning, where a smaller “student” model is trained on outputs from a larger “teacher” model. But Anthropic says that when used without permission and on this scale, the technique becomes a shortcut to replicate another lab’s work rather than a legitimate training practice.
How the Campaigns Worked
Anthropic’s account describes “industrial-scale campaigns” that, it claims, violated its terms of service and regional access restrictions, since Claude is not commercially available in China.
The company said MiniMax was responsible for most of the activity, with more than 13 million exchanges, followed by Moonshot’s 3.4 million and DeepSeek’s roughly 150 000 interactions. These exchanges, Anthropic asserted, focused on extracting advanced AI capabilities such as reasoning, coding and agentic tool use.
Anthropic also said the Chinese labs used proxy services and networks of fake accounts to evade detection and access Claude at scale.
Industry and Security Concerns
While distillation is a common approach in AI research, the allegations have drawn attention because of their scale and potential implications for intellectual property and technology competition. Anthropic warned that models produced through unauthorised distillation may lack built-in safety guardrails, increasing the risk that they could be misused in cyber operations, disinformation campaigns or surveillance if integrated into systems without proper controls.
The company’s claims also come amid ongoing policy discussions in the United States about export controls on advanced AI chips. Anthropic argues that restricting access to high-end computing hardware for Chinese developers could curb both direct model training and large-scale distillation efforts.
Responses and Wider Debate
None of the Chinese firms named in Anthropic’s post responded publicly at the time of reporting. Meanwhile, critics of Anthropic’s framing have argued that distinguishing between legitimate AI training practices and illicit copying is complex, and that public discourse around AI distillation often reflects broader competitive tensions in the global technology landscape.
OpenAI, another major US AI developer, has also recently accused DeepSeek of using distillation-type methods on its own models, highlighting how this issue resonates across the industry.
What This Means for the AI Sector
The controversy underscores the rapid evolution of the AI field and the challenges companies face in protecting proprietary models while balancing innovation and collaboration. As Chinese Labs Mine Claudes Capabilities becomes a focal point in debates over AI governance, policymakers, cloud providers and technology firms are under pressure to clarify standards for safe and ethical AI development.
