Anthropic has accused its Chinese rivals, DeepSeek and Moonshot, of secretly routing user requests to its Claude models, then presenting the responses as their own. The American AI company detailed the allegations in a 154-page report released on Thursday, marking a substantial level of documentation supporting the claims. This new accusation differs from previous claims as it alleges direct re-routing of users, potentially exposing Chinese user data to an American company. “[These] findings raise concerns about the misuse of user data by PRC AI labs,” Anthropic stated.

Moonshot and DeepSeek Routing Practices Utilize Anthropic’s Claude

The Kimi model, developed by Moonshot, reportedly presented outputs from Anthropic’s Claude as its own, according to a 154-page report released by Anthropic on Thursday. This practice extended beyond simple model imitation; Moonshot allegedly used these user interactions, funnelled through Claude, to further refine its Kimi model’s performance. DeepSeek employed a similar strategy, covertly routing user requests to Claude and displaying the responses as originating from its own systems, Anthropic says.

This marks the first accusation from Anthropic of direct user re-routing, a departure from previous claims centered on model training. These actions potentially exposed data from Chinese users to an American company, raising significant privacy concerns detailed in Anthropic’s report.

That campaign, officials claim, occurred “likely with Chinese government awareness”. Beijing responded to those earlier accusations on Wednesday, characterizing them as attempts to stifle China’s AI industry and warning of potential retaliatory measures.

While previous accusations focused on Chinese companies learning from American models, the direct re-routing of users represents a new level of concern for Anthropic, suggesting a deliberate effort to use a competitor’s infrastructure and data for their own gain, according to the company. The report details how these exchanges were used to improve the performance of the models, creating a feedback loop dependent on an American AI system.