Anthropic, backed by US intelligence agencies, claims Chinese AI firms like Alibaba are harvesting Claude’s 'thinking process' to train rival models. This event highlights significant risks to intellectual property and potential for tighter AI export regulations.
Anthropic has reported large-scale efforts by Chinese artificial intelligence labs to extract proprietary data from its Claude AI models. This discovery, which involves multiple companies including Alibaba, Moonshot AI, DeepSeek, MiniMax, StepFun, and Z.AI, has been confirmed by US security agencies, including the NSA, CISA, and the FBI. The report describes a coordinated attempt to bypass security guardrails and steal the reasoning processes that give Anthropic's models their advanced capabilities.
At the core of the issue is a process known as 'model distillation.' While distillation is a standard technique in AI development—often used to make models faster or more efficient—the report labels these specific campaigns as malicious. The labs allegedly used automated systems to bombard Claude with millions of prompts, specifically designed to force the AI to reveal its internal 'chain of thought.' By extracting these reasoning steps, the labs could theoretically train their own, smaller, and more efficient language models to mimic the performance of Claude without incurring the massive research and development costs associated with building frontier AI from scratch.
Alibaba was identified as the operator of the largest campaign. According to the findings, the effort involved over 151 million exchanges between May and July 2026, utilizing more than 3,500 distinct user accounts to avoid detection. The scale of this operation, which peaked at three million requests per day, underscores the intense pressure on Chinese AI firms to keep pace with US technological advancements.
For investors, this news carries implications beyond simple intellectual property theft. The confirmation of these activities by US government agencies highlights the growing intersection of artificial intelligence and national security. There is a tangible risk that these 'distilled' models, which lack the safety guardrails developed by US firms, could be deployed for military, surveillance, or offensive cyber operations.
This incident may lead to stricter oversight of how US AI companies grant API access to international users. As the US government tightens controls on AI technology, companies with significant exposure to these markets or those operating in the generative AI space could face new regulatory hurdles. The situation also places pressure on AI leaders to balance the need for open access to their tools with the security imperative of protecting their most valuable intellectual property.
Investors should monitor future updates from the US government regarding potential export restrictions or new security mandates for AI developers. Additionally, the industry will be watching how companies like Anthropic adjust their security protocols to differentiate between legitimate users and automated harvesting attempts, as this will likely affect future API costs and accessibility.