Anthropic: Chinese rivals secretly rerouted users to Claude
Anthropic says Moonshot and DeepSeek funneled user requests to Claude through fake accounts, carrying data from China and Russia.
In short
Anthropic accuses Chinese providers Moonshot and DeepSeek of quietly routing their own users requests to Claude through thousands of fake accounts — nearly 300,000 requests from Kimi alone in ten days.
At a glance
- Anthropic says Moonshot had Claude answer nearly 300,000 customer requests in just ten days.
- The traffic ran through a network of 5,380 fake accounts, according to the report.
- From May to July, Kimi called Anthropic models 23 million times, Anthropic says.
- Redirected uploads included footage from areas around Chinese army sites.
- Via DeepSeek, login credentials for a Russian government database allegedly came through.
Anthropic says two Chinese AI providers passed their own customers requests to Claude instead of answering them with their own models. The companies named are Moonshot, maker of the Kimi assistant, and DeepSeek. The traffic moved through accounts created for that purpose, according to the report.
Over a ten-day window, Claude handled nearly 300,000 customer requests coming from Moonshot. A network of 5,380 fake accounts carried them. Between May and July, Kimi called Anthropic models 23 million times. Every figure here comes from Anthropic itself, relayed through a German news-agency report.
Someone typing into a Chinese assistant does not expect the text to arrive at a US company. According to Anthropic, that is what happened: a user uploaded surveillance footage from areas near Chinese army sites, assuming Kimi would process it. Material reaching Anthropic through DeepSeek allegedly included data from an agency linked to the Russian defense ministry and working login credentials for a Russian government database.
Cheap Chinese models are the main reason investors question how much pricing power US labs still hold. If the accusation holds, part of the quality buyers paid for was not built in-house but borrowed without a contract. Anthropic also argues that Chinese developers train their systems on American software such as Claude and ChatGPT.
So far this is one company talking about its competitors. The report carries no response from Moonshot or DeepSeek, no account of how Anthropic detected and closed the accounts, no named internal report, and no word on legal or regulatory follow-up. A second independent newsroom was not technically reachable for this piece, so the figures rest on a single source.
FAQ
What is Anthropic accusing Moonshot and DeepSeek of?
Of secretly rerouting requests from their own users to Anthropics Claude models rather than serving those requests with their own systems.
How much traffic was involved?
Anthropic cites nearly 300,000 Kimi customer requests in ten days, handled via 5,380 fake accounts, plus 23 million calls between May and July.
Have the Chinese companies responded?
The report reviewed for this article contains no statement from Moonshot or DeepSeek; the account comes solely from Anthropic.
