Anthropic Accuses Chinese Companies of Siphoning Data From Claude
The allegations mirror those of OpenAI, which told House lawmakers that DeepSeek used ‘distillation’ to improve models
- Anthropic accused three Chinese AI companies of creating over 24,000 fraudulent accounts to siphon data from its Claude model.
- The companies prompted Claude over 16 million times, using distillation to train their own products faster and at lower cost, Anthropic said.
- Anthropic said the alleged activity by the Chinese developers raises U.S. national-security concerns regarding military and intelligence systems.
U.S. artificial-intelligence startup Anthropic said three Chinese AI companies set up more than 24,000 fraudulent accounts with its Claude AI model to help their own systems catch up.
The three companies—DeepSeek, Moonshot AI and MiniMax—prompted Claude more than 16 million times, siphoning information from Anthropic’s system to train and improve their own products, Anthropic said in a blog post Monday.
Earlier this month, 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—companies use it to build smaller versions of their own products, for example—but it could also be used to build competitive products “in a fraction of the time, and at a fraction of the cost.”
The scale of the different companies’ distillation activity varied. DeepSeek engaged in 150,000 interactions with Claude, whereas Moonshot and MiniMax had more than 3.4 million and 13 million, respectively, Anthropic said.
Representatives from DeepSeek, Moonshot and MiniMax didn’t respond to requests for comment.
Many Chinese companies including Moonshot and MiniMax have recently released their latest AI models, many of which feature enhanced reasoning and coding capabilities. DeepSeek is preparing to roll out its next-generation model soon.
When DeepSeek first captured the attention of AI enthusiasts last year, it raised concerns that China might be able to quickly catch up with U.S. AI companies even without having access to the most powerful AI chips. AI observers speculated that DeepSeek might have used distillation.
In a research paper updated in September, DeepSeek said that during a late stage of pretraining its flagship V3 model, it exclusively used plain webpages and ebooks, without incorporating any synthetic data. However, it said some webpages contained “a significant number of OpenAI-model-generated answers.” DeepSeek said its base model might have acquired knowledge from other powerful models indirectly by drawing on such webpages.
Synthetic data, often using distillation, has been increasingly adopted for training large foundation models as developers face a shortage of high-quality data and focus on giving models so-called agentic capabilities, meaning allowing them to take action proactively to complete tasks on behalf of users. In a technical report in July, Moonshot said it used synthetic data for training its Kimi K2 model.
Anthropic said the activity by the Chinese developers raised national-security concerns for the U.S. “Foreign labs that distill American models can then feed these unprotected capabilities into military, intelligence, and surveillance systems,” the company said.