Tech Desk: Anthropic has leveled serious accusations against several leading Chinese AI laboratories, including Moonshot AI and DeepSeek, alleging they secretly routed customer prompts through its powerful Claude models. This “illicit distillation” involved using thousands of fraudulent accounts to siphon Claude’s responses, then presenting them as if generated by their own systems. The allegations, detailed in Anthropic’s latest threat intelligence report, mark a significant escalation in the ongoing dispute over intellectual property in the rapidly evolving AI landscape.

Moonshot AI, the creator of the Kimi chatbot, is accused of diverting hundreds of thousands of user requests to Claude via over 5,000 fake accounts, primarily based in Singapore and Japan, to bypass geographical restrictions. Anthropic claims Moonshot then extracted Claude’s reasoning transcripts for training its own software, amounting to over 23 million exchanges between May and July. DeepSeek reportedly employed similar tactics, routing live customer conversations through Claude. Xiaomi was also implicated in comparable behavior.

Crucially, this covert data routing exposed highly sensitive information. Anthropic cited instances where users, unknowingly utilizing compromised Chinese platforms, submitted queries that included details from a Chinese military affiliate, live credentials for a Russian government database, and material from a Chinese police case-management system. The scale of this data exposure raises profound privacy and national security concerns.

Beyond Moonshot and DeepSeek, Anthropic’s report also named Zhipu, SenseTime, and MiniMax, accusing them of various forms of illicit extraction and proxy services. Anthropic first raised these concerns in February, but the latest findings suggest a substantial increase in the alleged activity. The company asserts these practices violate privacy laws and terms of service, confirming it has taken steps to disrupt the identified schemes while continuing vigilant monitoring.