Anthropic Says Chinese AI Labs Turned Claude Into a Training Ground

Anthropic Says Chinese AI Labs Turned Claude Into a Training Ground
Anthropic had already begun sounding the alarm in February, then escalated the case in June, when policy chief Sarah Heck told lawmakers that Alibaba had made 28.8 million illicit exchanges with Claude in little more than six weeks.

Its latest report, released Thursday, casts the alleged activity as far broader. Anthropic says it identified five campaigns linked to Alibaba, Moonshot AI, DeepSeek, Zhipu and Xiaomi, totaling nearly 200 million exchanges between May and July. The target, it says, was not merely Claude’s answers but the reasoning traces behind its coding, analysis, tool-use and logic capabilities. “Unauthorized labs have developed increasingly sophisticated methods to circumvent our defenses and harvest the capabilities of US frontier models,” the company said.

The largest alleged operation was attributed to Alibaba: 151 million exchanges across 3,500 accounts, peaking near three million a day, using what Anthropic described as a shared prompt designed to pull out chain-of-thought material for Qwen training. Anthropic also alleged that Moonshot routed 23 million user requests to Claude, while DeepSeek routed 12.1 million in 14 days. Some routed prompts reportedly included material tied to Chinese and Russian government or military users.

Anthropic’s account also points to a more immediate security concern. It says Moonshot-linked traffic included surveillance footage uploaded by a PLA-affiliated Kimi user, while another 10-day burst sent almost 300,000 requests through 5,000 accounts, largely toward Claude Opus.

In a separate threat-intelligence disclosure, Anthropic said it disrupted malicious Claude use by actors linked to China and Russia, arguing that AI was being used through multi-agent systems for reconnaissance, exploitation and data theft—not simply as a chatbot. Andrej Karpathy amplified the company’s statement that it had found and stopped operations involving cyberattacks, influence, surveillance, biology and weapons. Anthropic says it has tightened detection, reduced the detail in reasoning transcripts and will require identity checks for suspicious users operating from countries where Claude is unavailable. Alibaba, Moonshot, DeepSeek, Zhipu and Xiaomi did not respond to requests for comment in the cited reporting.

Continue reading https://foxvector.com/stories/01a09047-50a2-2a12-7008-25953fa4ac11