Anthropic says DeepSeek and other Chinese AI companies fraudulently used Claude

In a bold assertion, Anthropic has accused several major AI companies in China of unfairly leveraging its Claude technology to enhance their own systems. On Monday, the AI startup highlighted that three prominent Chinese research labs—DeepSeek, MiniMax, and Moonshot AI—have been 'illicitly' employing Claude through a technique known as distillation to boost their models. Anthropic warned that these activities are intensifying and becoming more sophisticated. In their detailed statement, they emphasized the urgency of addressing these issues, stating, 'The window to act is narrow, and the threat extends beyond any single company or region.' They called for swift and coordinated action from industry leaders, policymakers, and the global AI community to tackle the problem effectively. The startup reported that these distillation efforts involved around 24,000 fraudulent Claude accounts, which were responsible for generating over 16 million interactions that violated Anthropic's terms of service and access restrictions. Distillation allows less powerful models to learn from more advanced ones—a practice that is standard among many U.S. companies but has raised concerns about its misuse by Chinese competitors to gain access to proprietary technology. Previously, OpenAI had also raised alarms about DeepSeek's potential misuse of its outputs for model training. Moreover, recent disclosures from Google indicated a rise in model extraction attempts, also known as 'distillation attacks.' Anthropic pointed out that such practices enable competitors to acquire advanced capabilities far more quickly and cheaply than if they were to develop them independently. Claude, Anthropic's advanced model, is not commercially accessible in China, yet rival labs have reportedly discovered ways around this limitation. Notably, Anthropic revealed that DeepSeek aimed to develop 'censorship-safe alternatives' for sensitive inquiries, and they managed to observe MiniMax's activities while they were ongoing. When Anthropic launched a new model, MiniMax quickly adjusted its strategy, redirecting significant traffic to exploit features of the new system. Representatives from DeepSeek, MiniMax, and Moonshot AI have yet to respond to requests for comments regarding these allegations. Beyond ethical concerns, Anthropic warned that improper distillation poses safety risks, as less sophisticated models may not have adequate safeguards, potentially leading to dangerous applications like bioweapons development. In light of these threats, Anthropic has implemented 'behavioral fingerprinting systems' and collaborates with other AI firms to share insights on potential risks. CEO Dario Amodei has articulated concerns about the trajectory of leading AI models, suggesting that without proper safety measures, they could inadvertently assist in harmful applications. Furthermore, Anthropic has faced its own challenges regarding the use of copyrighted materials in training its models. Earlier this year, the Washington Post reported on a project dubbed Project Panama, which aimed to 'destructively scan' books worldwide, leading to a $1.5 billion settlement from a class-action lawsuit filed by authors and publishers, although Anthropic did not admit any wrongdoing in the settlement.