Anthropic claims Chinese AI firms secretly use Claude. Is it true?

Anthropic claims Chinese AI firms have been secretly using its flagship model, Claude, for "industrial-scale" distillation, a practice where outputs from a more capable "teacher" model are used to improve another model. According to the company, Chinese developers circumvented usage restrictions through fraudulent accounts and third-party proxy services, harvesting Claude's output and reasoning traces at scale.

The allegations come from a joint advisory by the NSA, FBI, and CISA, and US Treasury Secretary Scott Bessent also warned of large-scale "distillation". Anthropic reported that over 10 days, its Moonshot model routed almost 300,000 user requests, primarily to the Opus model, with Chinese companies using a proxy network of 5,380 fraudulent accounts based in Singapore and Japan.

Some alleged cases involved Claude being used as a hidden back end, with users interacting with a domestic Chinese model while queries were actually processed by Anthropic's servers. While industry analysts are divided on the credibility of Anthropic's findings, experts note that Chinese developers face intense pressure due to US chip export restrictions, and distillation allows domestic developers to leapfrog expensive reasoning data generation.

The main concern for Beijing would be data security, as sensitive information from state-affiliated organizations and companies could potentially be exposed to American AI model providers. Regulators are investigating domestic AI labs following Anthropic's report, with a focus on DeepSeek and Moonshot.