Anthropic reckons three Chinese AI outfits have been siphoning data from Claude at an industrial scale, and it is not treating it as harmless curiosity.
In a blog post on Monday, the US artificial-intelligence startup said DeepSeek, Moonshot AI and MiniMax “prompted Claude more than 16 million times”, pulling information from its system to train and improve their own products.
The claim echoes a move earlier this month when an Anthropic rival, OpenAI, sent a memo to House lawmakers accusing DeepSeek of using the same tactic, called distillation, to mimic OpenAI’s products.
Anthropic said distillation had legitimate uses, like building smaller versions of a company’s own tools, but warned it could be used to build competitive products “in a fraction of the time, and at a fraction of the cost.”
It said the scale varied by company, claiming DeepSeek ran about 150,000 interactions with Claude while Moonshot and MiniMax logged more than 3.4 million and 13 million, respectively.
DeepSeek is preparing to roll out its next-generation model soon, after it grabbed attention last year and sparked worries that China could close the gap without the most powerful AI chips.
In a research paper updated in September, DeepSeek said that during a late stage of pretraining, its flagship V3 model “exclusively used plain webpages and ebooks, without incorporating any synthetic data.” It added some webpages contained “a significant number of OpenAI-model-generated answers,” and said its base model might have picked up knowledge indirectly by drawing on those pages.
Synthetic data, often using distillation, is becoming more common as developers run short of high-quality data and chase so-called agentic capabilities, meaning systems that act proactively to complete tasks for users.
In a technical report published in July, Moonshot said it used synthetic data to train its Kimi K2 model, which sits awkwardly alongside claims that distillation is just a benign optimisation trick.
Anthropic said the Chinese developers’ activity raised national-security concerns for the US, warning: “Foreign labs that distil American models can then feed these unprotected capabilities into military, intelligence, and surveillance systems.”
It claimed the three firms set up more than 24,000 fraudulent Claude accounts to help their own systems catch up, and that number is doing plenty of work all by itself.
x
