Anthropic accused DeepSeek and AI tigers from China of training on its Claude model

AI model creator Claude Anthropic has claimed widespread attacks on it by Chinese AI tigers - the country's name for startups that could become potential national leaders in generative AI. Anthropic estimates that Chinese competitors could have made more than 16 million queries to Claude across thousands of accounts to replicate its capabilities at a lower cost.

Anthropic said it detected organized, automated campaigns to massively exploit Claude by Chinese developers of AI tools, including DeepSeek. According to the US company, some 24,000 accounts and proxy services were used to circumvent regional restrictions. In total, Anthropic claims DeepSeek and "AI tigers" conducted more than 16 million interactions with the model.

The company characterizes these actions as a "distillation attack". Distillation is a technique in which a less powerful model is trained on the responses of a more advanced system to replicate its capabilities with less time and expense than would be required if developed independently.

Each of the Chinese companies, according to Anthropic's assessment, tried to extract certain skills:

- DeepSeek collected examples of step-by-step reasoning and logical analysis to train its model to solve complex problems. It also fielded queries on politically sensitive topics, seeking language that could be used to customize its own system to meet censorship requirements.

- Moonshot AI, Anthropic told us, emphasized the development of AI agent functions - the model's ability to use tools, work with code and data, and perform a sequence of actions without human intervention.

- MiniMax focused on training its system to solve complex programming problems on its own using external tools and services - effectively replicating the behavior of an advanced digital assistant. MiniMax sent the largest number of requests - 13 million, Anthropic reported.

In response, it implemented systems to detect distillation patterns, strengthened account authentication, and began developing defenses against such attacks.

DeepSeek plans to release its next-generation V4 AI bot by the end of the month, The Information reported. According to its sources, the model has powerful programming capabilities and has beaten rivals from Anthropic and OpenAI in internal testing. Reuters specifies that the release may take place next week. At the same time trained V4 on the most advanced chips Nvidia Blackwell, despite the ban on their supply to China, told the agency a senior official of the administration of US President Donald Trump.

The release of DeepSeek's other model, the R1, in January 2025 led to a sell-off in Nvidia and other AI companies.

At the same time, ChatGPT developer OpenAI reported signs of attempts to distill its models. Its main partner and investor, Microsoft, reported suspicions that DeepSeek was violating the license terms of ChatGPT.

Distillation itself is widely used in the industry to create smaller and cheaper versions of models, but usually only for internal use. Anthropic acknowledged that businesses use this approach in their own designs, but called it unacceptable to systematically extract the capabilities of a third-party model.

This article was AI-translated and verified by a human editor