China investigates DeepSeek and Moonshot AI over data redirection claims

The Cyberspace Administration of China has launched an investigation into several Chinese artificial intelligence companies, specifically targeting DeepSeek and Moonshot AI, following allegations of unauthorized data redirection.

A threat intelligence report from Anthropic claims that these companies used fake accounts to secretly route millions of user queries to Anthropic’s Claude models. The report alleges that DeepSeek redirected over 12.1 million conversations in a 14-day period during July, while Moonshot AI reportedly routed approximately 300,000 requests via thousands of fake accounts. The goal of this activity is suspected to be the extraction of Claude’s reasoning traces and safety measures to train domestic models more efficiently.

Chinese authorities are investigating whether sensitive data, including information linked to municipal police systems and state institutions, was exposed through this process. In response, Chinese state media has countered by accusing Anthropic of using its privacy policies as a channel to transfer user data to U.S. intelligence agencies.

The incident has escalated into a geopolitical issue. U.S. Treasury Secretary Scott Bessent indicated that sanctions could be imposed on Chinese-origin models if technology theft is proven, noting the presence of American AI watermarks in Chinese models. This follows warnings from the FBI and other intelligence agencies regarding the protection of intellectual property from Chinese-centered theft.