Anthropic accuses Chinese companies of stealing data from its AI model Claude

According to the report, the companies DeepSeek, Moonshot AI and MiniMax fed the Claude system over 16 million queries. Anthropic claims that the extracted information was used to train and upgrade the products of those companies. This was announced in an official post on the company’s blog on Monday.

Earlier this month, according to the publication, OpenAI also appealed to lawmakers in the US House of Representatives and accused DeepSeek of model distillation (a tactic of systematically copying answers to train a competing system).

Anthropic clarified that refining has legitimate uses – for example when a company develops a reduced version of an existing model – but added that the method can also be used to build competing products “in a fraction of the time and at a fraction of the cost.”

The volume of activity varied between the companies: DeepSeek made about 150 thousand interactions with Claude, while Moonshot and MiniMax made more than 3.4 million and 13 million interactions respectively.

At the same time, Chinese companies including Moonshot and MiniMax have recently launched new AI models that include advanced capabilities of logical inference and programming. DeepSeek is expected to launch its next generation model soon.

When DeepSeek attracted international attention last year, there were speculations that China could quickly catch up with American companies — even without access to the most advanced AI chips. Observers estimated that distillation may have been used to shorten processes.

In a research paper updated in September, DeepSeek noted that in an advanced stage of pre-training of its V3 model, only web pages and digital books were used, without incorporating synthetic data. However, it admitted that some sites included “a significant number of answers generated by OpenAI models”. The company added that its model may have acquired knowledge indirectly from other models through the same sources.

The use of synthetic data – sometimes through distillation – is expanding, mainly due to the lack of high-quality training data and the desire to give models the ability to operate independently to perform tasks for users. In a technical report published in July, Moonshot stated that it used synthetic data to train its Kimi K2 model.

Anthropic warned that the activity raises concerns for US national security. “Foreign laboratories that refine American models can feed these unprotected capabilities into military, intelligence and surveillance systems,” the company said, according to the report.