US AI Firm Alleges Chinese Companies Leak Sensitive Data via Model Use

Anthropic has accused several Chinese AI firms, including Moonshot AI and DeepSeek, of extensively using its Claude model to train their own systems, sometimes routing hundreds of thousands of real user requests without consent. The transmitted data reportedly included surveillance material, internal credentials from state-owned enterprises, police case files, and details of Chinese tech projects. This raises concerns about sensitive information flowing to the United States amid the AI competition.

The allegations center on a practice known as model distillation, where companies use a more advanced AI system to generate training data for their own models. Anthropic claims that in one case, hundreds of thousands of genuine user queries were redirected through Claude without user awareness over a ten-day span. The transmitted material reportedly included military-related surveillance content, credentials from state-owned enterprise engineers, police case management records, and internal corporate AI project details.

The accompanying analysis emphasizes that while individual data points may seem trivial, aggregated real-world usage data carries significant intelligence value. Patterns emerging from thousands of engineers discussing technical problems across sectors like semiconductors, aerospace, and telecommunications could potentially reveal systemic bottlenecks and create a dynamic picture of China's industrial and technological landscape.