The Chinese Cyberspace Administration (CAC) is examining DeepSeek and Moonshot AI following allegations from Anthropic regarding the unauthorized transmission of user queries to Claude, as reported by The Information.
Initially, the CAC summoned representatives from all seven companies mentioned in Anthropic's September report for questioning. Subsequently, the focus shifted to DeepSeek and Moonshot, with officials interviewing their executives and staff.
The investigation is ongoing, and the CAC has yet to determine if sanctions will be imposed, as authorities assess the extent of potential violations and the companies' intentions. Neither DeepSeek nor Moonshot responded to inquiries from The Information.
The probe was prompted by a comprehensive 154-page report from Anthropic dated September 10, which accused seven Chinese developers of extensively using Claude for what is termed illegal distillation—leveraging a more powerful model's capabilities to train their own systems. The report also named Alibaba, Zhipu, SenseTime, MiniMax, and Xiaomi.
Millions of Queries Redirected to Claude
According to Anthropic, from May to July, Moonshot redirected over 23 million queries to Claude, while DeepSeek sent more than 12 million queries within a two-week span in July. The company claims that users were unaware their inquiries were being processed by Anthropic's model.
One instance noted in the report involves a user believed to be linked to the People’s Liberation Army of China, who allegedly transmitted surveillance system data from Chengdu via Kimi. This involved monitoring an individual across hundreds of cameras, including those near PLA facilities and defense research institutes.
Additionally, Anthropic indicated that queries from engineers developing a system for Chinese municipal public security were directed to Claude through DeepSeek, including processing data on people's movements using national identification numbers.
Beijing's Concerns Extend Beyond Distillation
For the United States, the primary issue remains the potential illegal acquisition of American AI model capabilities by Chinese firms.
However, for Beijing, there is a different concern: if Anthropic's accusations are accurate, sensitive data from Chinese users, government entities, police, and state-owned enterprises could have been transmitted to American systems. This situation may violate existing regulations on cross-border data transfers, as highlighted by sources from The Information.
China strictly regulates the transmission of extensive information about its citizens and organizations abroad. The CAC is also involved in the security assessment of generative AI models, granting the regulator access to information regarding the systems, data, and protective measures of Chinese developers.
Anthropic has separately reported that some of the redirected queries through DeepSeek and Moonshot contained confidential information from users and organizations. The company's report cites examples of internal data, accounts, and materials associated with government entities.
Investigation Takes Place Amid US-China Dialogue
This situation has gained further political significance with Chinese President Xi Jinping's visit to the United States. According to The Information, leaders of both nations are expected to discuss the establishment of an "AI dialogue" mechanism, allowing Washington and Beijing to inform each other about AI-related incidents that could impact national security.
Notably, Mattie Zhao, co-head of the Chinese equities department at Bank of America, has expressed that neither the US nor China can win the AI race alone.