Anthropic accuses Alibaba of stealing AI technology from Claude

Anthropic, a leading U.S. artificial intelligence developer, has accused Chinese e-commerce and IT conglomerate Alibaba of illicitly accessing its AI model, Claude, and stealing proprietary technology. Bloomberg reported on June 25 that Anthropic claims Alibaba orchestrated a systematic and illegal operation using tens of thousands of fake accounts to extract Claude's core functionalities.

In a letter to U.S. senators and White House officials, Anthropic detailed how entities linked to Alibaba's AI research for its chatbot, Tongyi Qianwen, specifically targeted Claude's most valuable capabilities: software engineering (coding) and agent reasoning. Between April and June of this year, Alibaba allegedly engaged in 28.8 million conversations with Claude through approximately 25,000 fake accounts. Anthropic asserts this represents the largest-scale alleged theft of U.S. AI models by a Chinese company.

The alleged method used by Alibaba is known as "adversarial distillation." This technique involves bombarding a leading AI model with numerous questions to extract high-quality responses, which are then used to train the company's own AI model. This allows companies to replicate the performance of advanced models without the substantial research and development costs and infrastructure investment typically required. While some forms of distillation for smaller models are accepted in the industry, replicating cutting-edge models directly violates the terms of service of AI companies.

Anthropic warned the White House that these "distillation attacks" are being conducted on an "industrial scale" to "harvest U.S. advanced AI capabilities and repackage them as their own." The company also noted that AI systems built this way often lack essential safety features. This prompted Anthropic to urge U.S. government intervention and enforcement.

In response, U.S. lawmakers are taking action. Senators Bill Hagerty (Republican) and Andy Kim (Democrat) plan to introduce an amendment to the annual National Defense Authorization Act (NDAA) that would allow for blacklisting or sanctioning Chinese companies that improperly intercept U.S. AI model outputs to train competing models. A similar bipartisan bill is also being prepared in the House of Representatives. The U.S. Department of Defense had already added Alibaba to a blacklist of companies supporting the Chinese military earlier this month, though Alibaba denies any military ties and is suing the Pentagon to have the designation removed. Alibaba has not commented on the allegations of Claude replication.

The alleged cloning of Claude poses a significant threat to Anthropic, which is currently preparing for an IPO with a market valuation of $965 billion. U.S. authorities estimate that such unauthorized distillation practices have cost Silicon Valley research labs billions of dollars. The U.S. Department of Commerce recently imposed restrictions, including barring foreign users from accessing Anthropic's top models, 'Firefly 5' and 'Mythos 5,' citing national security concerns, forcing Anthropic to temporarily suspend these services. Despite recent urgent meetings between Anthropic's top technical staff and White House officials, there has been little progress in resuming services or resolving the dispute. News of Alibaba's alleged Claude theft sent Alibaba's stock price plummeting by up to 4.8% in Hong Kong trading, causing significant market disruption.

“

These distillation attacks are being conducted on an industrial scale to harvest U.S. advanced AI capabilities and repackage them as their own. AI systems built this way often lack essential safety features.