Anthropic has accused Chinese AI companies Moonshot and DeepSeek of silently routing user requests to Claude models and displaying the responses to users as its own.
In a 154-page report released Thursday, Anthropic said that Moonshot secretly relayed user requests to Claude rather than using its own Kimi models. "These users thought they were using a Kimi model, but received responses from Claude instead," Anthropic wrote.
In one instance, Moonshot relayed nearly 300,000 customer requests to Anthropic over a ten-day window. As Claude is not accessible from inside China, Moonshot allegedly used a proxy service network comprising 5,380 fraudulent accounts to access the models, according to the report.
Some of the requests fed to Claude contained rather sensitive information. One user, who Anthropic said was likely affiliated with China's People's Liberation Army, used what they believed was Moonshot's Kimi model to load surveillance data from a CCTV archive covering hundreds of cameras in Chengdu, the capital of southwestern China's Sichuan province.
"We do not know if Moonshot notified their customers that their requests were being rerouted to Anthropic and exposed to a third party," Anthropic said.
The Anthropic investigation also said DeepSeek used similar tactics and routed exchanges to Claude without informing customers.
For example, DeepSeek allegedly relayed requests from a user working with Russian Ministry of Defense data, exposing live credentials for a Russian government database, according to the report.
Distillation campaigns
Anthropic also accused Alibaba and Xiaomi of running distillation campaigns targeting its models.
"Operators affiliated with Alibaba ran the largest distillation attack we have ever measured," Anthropic said, adding that Alibaba's "illicit distillation campaign" peaked at nearly 3 million exchanges per day launched from over 3,500 fraudulent accounts.
Anthropic also found that Xiaomi had relayed user conversations and coding sessions from its MiMo models to Claude. Anthropic, however, said it did not find evidence that Xiaomi used Claude’s responses to serve its users.
Anthropic's report came after several U.S. federal agencies on Tuesday alleged that Chinese AI companies systematically extracted capabilities from American AI models through distillation.
Knowledge distillation is a machine learning technique involving training a smaller AI model using output from more capable ones, according to the U.S. agencies' statement.
China has pushed back against the allegations, calling them groundless and arguing that the U.S. move "politicizes and instrumentalizes distillation," which it said was a "common practice" in the AI field.