Anthropic said that DeepSeek and Moonshot redirected requests to Claude and passed the answers off as their own

In Anthropic’s report, the developer of Claude, it says that Moonshot users — a Chinese AI company — believed they were working with the Kimi family of models, but in fact they were receiving answers from Claude.

According to Anthropic, in one 10-day episode Moonshot redirected nearly 300,000 user requests to Claude.

To do this, Moonshot, the company says, used a network of 5,380 fake accounts, registered mainly in Singapore and Japan.

Anthropic also writes that Moonshot stored at least part of the conversations and extracted reasoning chains from the redirected sessions in order to later train its own models.

The report says that DeepSeek, another Chinese AI company, used similar tactics. Anthropic separately notes that the redirected requests contained confidential information about various clients of these companies. The company did not determine whether users were warned that their requests became accessible to a third party.

Sources

• Source: meduza.io