Anthropic Alleges Massive AI Model Theft by Alibaba via Industrial-Scale Distillation Attacks

US AI powerhouse Anthropic has leveled serious allegations against the Chinese technology and e-commerce conglomerate Alibaba, claiming the firm “brazenly” and “illicitly” harvested capabilities from its Claude AI model.

In a formal letter addressed to US Senators Tim Scott and Elizabeth Warren, the San Francisco-based developer detailed a massive operation in which Alibaba-linked operators allegedly used thousands of fraudulent accounts to conduct nearly 29 million exchanges with Claude. Anthropic described this as the largest extraction campaign of its kind.

The company explained that the breach was executed through “distillation attacks,” a process where outputs from a superior AI model are extracted to train and improve a weaker model. According to Anthropic, the attackers specifically targeted Claude’s high-value capabilities, including its decision-making logic and its capacity to handle complex, long-form tasks.

Anthropic warned that such industrial-scale attacks effectively turn billions of dollars of American R&D investment into a subsidy for geopolitical rivals, allowing Chinese firms to repackage US innovation as their own. The letter further asserted that these activities pose a broader threat to US national security, citing Department of Defense claims that Alibaba, Baidu, and BYD maintain ties to the Chinese military.

While Alibaba and other named firms have denied these allegations—with Alibaba recently suing the US government to be removed from the Pentagon’s blacklist—the trend of distillation attacks remains a recurring concern. OpenAI has previously reported similar attempts by Chinese groups to replicate American AI technology at a fraction of the original development cost.

As Anthropic prepares for a potential landmark stock market debut that could place it among the world’s most valuable companies, it continues to face a complex landscape of both geopolitical rivalry and cybersecurity scrutiny, particularly regarding the potential for its advanced models to identify vulnerabilities in computer systems.

Source link