Anthropic, the creator of the Claude series of large language models, is under fire for a covert tracker that monitored Chinese users of its AI service, according to a report in The Washington Post. The discovery has been framed as evidence that U.S. AI firms are adopting increasingly aggressive tactics to fend off what they describe as systematic copying by Chinese companies.

Chinese developers have demonstrated a capacity to match the performance of U.S. models within months. The Post cited a recent example: Zhipu AI, a Chinese startup, released a free model that outperformed Anthropic's Claude Opus 4.8 in identifying computer vulnerabilities, despite Claude having been launched only in May.

In response, Anthropic has called for a coordinated U.S. effort to curb “distillation attacks,” a process in which developers prompt an existing model millions of times to extract its capabilities and then replicate them in a new system. While distillation itself is not illegal and is practiced by American firms, Anthropic argues that using the technique to accelerate Chinese models violates its user terms.

The company has joined OpenAI in urging policymakers to treat such attacks as a form of intellectual‑property theft. During a recent Senate hearing, Republican Senator Tim Scott echoed that sentiment, urging the creation of clear, concise export‑control policies aimed at preventing China from gaining a technological edge through these methods.

Research from Peking University and the state‑funded Chinese Academy of Sciences, conducted in February, supports the claim that Chinese models are heavily derived from U.S. counterparts. The study detected substantial evidence of distillation across a broad range of Chinese AI systems, indicating that many are built upon the foundations of American models.

One striking case involved Alibaba’s Qwen AI model, which repeatedly mimicked Claude’s behavior in intensive tests. In some instances, the model even identified itself as Claude, suggesting a deep level of copying. Anthropic later claimed that the Qwen model’s advancement followed the largest distillation attack ever recorded on Claude, which occurred in June.

Anthropic’s leadership argues that to maintain a 12‑ to 24‑month lead for U.S. AI technology, the United States must consider a suite of interventions. Potential measures include restricting Chinese access to advanced models, semiconductor components, and U.S. data centers. The goal, according to Anthropic, is to blunt the rapid reverse‑engineering pipeline that Chinese firms appear to be exploiting.

The controversy raises broader questions about the balance between open AI research and competitive national security concerns. While the industry pushes for innovation, the line between legitimate model training and intellectual‑property theft is becoming increasingly blurred, prompting calls for legislative clarity.

This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.