DeepSeek and Moonshot Caught Using Claude as Secret Backend

Anthropic dropped a bombshell Threat Intelligence Report revealing that rival Chinese labs haven't just been scraping Claude for training data—they secretly deployed it as the real-time engine behind their own consumer chat interfaces. Whenever unsuspecting users thought they were talking to homegrown Chinese neural nets, they were frequently just chatting with an American AI wearing a cheap disguise.

Moonshot ran an aggressive relay ring using over 5,000 burner accounts to throw 300,000 live user prompts into Claude Opus within a single ten-day burst. Across several months, Anthropic mapped over 23 million distinct connections back to the company. Users asked a question in China, the backend bounced it across the Pacific, and the response was served back as domestic engineering genius.

DeepSeek operated with far more surgical precision. Instead of blasting every query overseas, its infrastructure filtered out high-complexity coding requests resembling Claude Code tasks and rerouted them to Anthropic. Over a 14-day window, this stealth pipeline accounted for 12.1 million exchanges, complete with intercepted PLA-linked surveillance code, Chinese state-owned enterprise credentials, and defense contractor database schemas.

Alibaba took the traditional brute-force route by funneling 151 million prompts through 3,500 puppet accounts between May and July. With daily peaks hitting 3 million requests, this massive haul was used to distill and refine the Qwen model family. Other players like Zhipu and Xiaomi also tapped the well, sticking to synthetic dataset harvesting rather than live session hijacking.

The ultimate irony of the global AI race isn't that frontier models might replace humans, but that national technological independence apparently boils down to a covert reverse-proxy forwarding API bills to the ideological competition.

Source: Anthropic