Anthropic accuses Chinese AI firms of secretly using Claude to train their models

What happened

US AI company Anthropic has accused Chinese AI developers of extracting data from American models without authorization. According to the company's September report, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI pulled billions of tokens across millions of requests from models including Claude, GPT, Gemini and Grok over two years.

What is distillation?

Distillation is a technique in which the outputs of a more advanced model are used to train another system. Anthropic acknowledges it is a legitimate training method, but defines "illicit distillation" as an industrial-scale, covert campaign to replicate a model's capabilities without authorization.

The most specific allegation

The sharpest claim targets Moonshot AI, the developer of the Kimi chatbot. The report alleges that Moonshot secretly forwarded requests to Claude while users believed they were using Kimi, citing 300,000 customer requests over a ten-day period and a proxy network of 5,380 fraudulent accounts in Singapore and Japan. Alibaba is said to have carried out more than 151 million distillation activities between May and July 2026.

Reactions and doubts

The NSA, CISA and FBI issued a joint advisory urging US firms to strengthen detection systems and share intelligence. However, Wang Zebin, an investment manager at Shenzhen-based JG Investment, told the South China Morning Post that the report does not provide enough concrete examples to independently verify how the companies obtained or used the model outputs. China's cyberspace regulator reportedly questioned seven developers before narrowing its focus to DeepSeek and Moonshot.

Anthropic said it had disrupted the misuse operations it identified and strengthened safeguards against unauthorized extraction of Claude's capabilities. The allegations remain contested, with questions about attribution and the role of third-party intermediaries unresolved.