Anthropic Accuses Chinese AI Firms of Siphoning Knowledge from Claude

Leading artificial intelligence developer Anthropic has accused Chinese AI firms of illicitly siphoning knowledge from its flagship large language model (LLM) Claude, reporting the incidents to the U.S. Congress. The company characterizes these actions as “distillation attacks” where Chinese rivals exploit American intellectual property to advance their own AI systems. This situation once again highlights the intense competition and geopolitical tensions within the global artificial intelligence sector.

According to Anthropic's allegations, the Qwen AI team of Chinese tech giant Alibaba utilized approximately 25,000 fraudulent accounts to conduct over 28.8 million prompt-response exchanges with Claude over three months, potentially siphoning an estimated 57.6 billion tokens. Similarly, other Chinese firms, including DeepSeek, Moonshot AI, and MiniMax, are accused of attempting to replicate Claude's capabilities through 24,000 fake accounts and 16 million interactions. The aim of these clandestine operations is to train their own LLMs in advanced areas such as agentic reasoning and software engineering, thereby circumventing significant research and development costs and bypassing U.S. export controls.

In letters to Congress, Anthropic urged the implementation of measures such as information sharing, penalties, and stricter rules to restrict Chinese access to U.S. high-performance chips to counter these industrial-scale siphoning activities. The company emphasized that such actions constitute not only intellectual property theft but also economic espionage.

These developments are further intensifying the U.S.-China rivalry for AI supremacy. Such attacks on proprietary models, in which Anthropic and similar U.S. firms have invested billions of dollars, expose a growing vulnerability for Western tech giants. Of particular concern are the national security risks, as models developed through illicit distillation may lack the safety guardrails designed to prevent misuse, such as enabling bioweapon development or cyberattacks. The U.S. government is reportedly increasing its scrutiny of Chinese AI firms and considering tightening restrictions on technology transfers.

This situation is being interpreted as a form of technological Cold War in the modern era, representing a global struggle for economic, technological, and geopolitical influence in the AI domain. As the U.S. strives to maintain its leadership in the AI sector, the rapid advancements by Chinese firms make this race even more critical. It is worth noting that Anthropic itself has previously faced criticism for training its models on copyrighted books, illustrating the complexity of intellectual property debates within the industry.

Analysts and market expectations suggest that stricter measures for AI security and intellectual property protection will be implemented in the coming period. Enhanced technical defenses (behavioral monitoring across accounts, data watermarking, query throttling) and operational controls (account verification tiers, API usage auditing) are anticipated. Furthermore, pressure is mounting on policymakers to establish legal frameworks encompassing AI distillation-related export controls and to define international norms. These developments indicate that competition in the AI sector will not be limited to technological innovation but will also encompass international law and cybersecurity dimensions.