China Probes DeepSeek, Moonshot AI Over Anthropic’s Claims They Route Requests to Claude

China's Cyberspace Administration of China is probing DeepSeek and Moonshot AI over Anthropic's claims that the two labs routed user prompts to Anthropic's Claude, according to The Information. Anthropic's 154-page report, "Detecting and countering misuse of AI," published earlier this month, accused Moonshot of sending more than 23 million user inputs between May and July and DeepSeek of sending 12.1 million during a 14-day period in July, then using Claude's responses as training data in a practice known as distillation. Anthropic said the prompts contained names, email addresses, and company data of hundreds of end users in at least a dozen languages, and that the routing was "likely inconsistent with privacy laws and the labs' own terms of service.

China’s leading artificial intelligence models are once again coming under fire for their alleged reliance on American competitors—but this time, the inquiry is coming from inside the house. According to a report https://www.theinformation.com/articles/china-probes-deepseek-moonshot-potential-data-leaks-anthropic from The Information, China’s leading internet regulator, the Cyberspace Administration of China, is probing DeepSeek and Moonshot AI over accusations that the companies are routing user prompts and data to Anthropic’s Claude. The investigation has already reportedly resulted in several executives from leading Chinese AI labs getting called into the principal’s office and peppered with questions, but the Cyberspace Administration has reportedly narrowed its scrutiny to DeepSeek, the lab that produced the open-source DeepSeek AI model that briefly blew up the AI industry last year, and Moonshot AI, the makers of the popular open-source Kimi model. DeepSeek and Moonshot landed themselves under the microscope of the regulator after their names were brought up repeatedly in a 154-page report https://www.anthropic.com/threat-intelligence-report-september-2026 titled “Detecting and countering misuse of AI” published by Anthropic earlier this month. Within the document, the American frontier AI lab accused the competitors of having “fed conversations between their own models and users into Claude.” Anthropic accused the labs of sending millions of user inputs—more than 23 million from Moonshot between May and July and 12.1 million from DeepSeek during a 14-day period in July—to its model, then using the responses they received from Claude as training data to improve their own models, a method often referred to as distillation. Basically, the smaller models use the responses from the bigger, more robust models as a shortcut to improving its own outputs without needing the massive troves of training data. This has been a bit of a bugaboo for American labs, which have repeatedly accused Chinese competitors of distillation “attacks,” https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks alleging that the open-source models are effectively cheating or stealing training data by doing this. Go ahead and set aside the fact that the allegations come from companies known to have done things like pirated books and used the material to train their own models https://www.npr.org/2025/09/05/g-s1-87367/anthropic-authors-settlement-pirated-chatbot-training-material , they’ll tell you that’s fair use and not relevant. Arguably more relevant to the Chinese regulator’s work, though, is the fact that, in routing these user inputs to Claude, Moonshot and DeepSeek may have exposed sensitive user information to the third-party model. Anthropic alleges https://www.anthropic.com/threat-intelligence-report-september-2026 that the prompts included “sensitive information, including from individual users, major multinational companies, and state-affiliated actors,” and “contained names, email addresses, company data, and other sensitive data of hundreds of end users in at least a dozen languages.” Anthropic reasoned https://www.anthropic.com/threat-intelligence-report-september-2026 that the routing practice is “likely inconsistent with privacy laws and the labs’ own terms of service,” which it seems the Cyberspace Administration may agree with. The investigation was also launched, likely not coincidentally, right as the Trump administration is set to host Chinese president Xi Jinping https://apnews.com/article/trump-xi-summit-tensions-ai-competition-e560910c897fedb0448eda4cdbfdbaea . Expectation is that AI will be a prominent conversation during the stay, with the hope of landing on some sort of consensus https://time.com/article/2026/09/22/trump-xi-ai-deal-china/ on how to handle and regulate the technology. China’s investigation may be a bit of an olive branch to show the country is operating in good faith.