Anthropic has accused Alibaba of orchestrating large-scale unauthorized access to its Claude AI model through approximately 25,000 fraudulent accounts, according to a letter sent to US officials. The Chinese tech giant allegedly accessed Claude 28.8 million times between April and June.
In a formal complaint to US government officials, Anthropic alleged that Alibaba engaged in adversarial distillation—a technique to extract and replicate AI model capabilities. The effort directly contradicts Anthropic's policy of restricting Claude access in China.
The scale of the suspected operation is substantial. Over a three-month period, the roughly 25,000 accounts attributed to Alibaba generated nearly 29 million API calls to Claude. Anthropic characterizes the access as "illicit" and says it violates the company's terms of service.
Adversarial distillation involves using an AI model repeatedly to understand its outputs and behaviors, then building competing systems with similar capabilities. By accessing Claude millions of times, Alibaba could potentially train alternative models that replicate Claude's performance without licensing or paying for the original system.
Anthropics decision to restrict Claude in China reflects broader US policy concerns around AI technology transfer to Chinese entities. The company has positioned itself as cautious about where its technology goes, making the alleged breach particularly significant.
The accusation escalates tensions in the AI industry around model security and access controls. It also highlights the challenge AI companies face in preventing coordinated, large-scale attempts to circumvent restrictions through distributed account networks.
Anthropics formal complaint to US authorities suggests the company views this as more than a typical terms-of-service violation. The involvement of government officials points to potential national security implications around AI model access and technology transfer.
Alibaba has not publicly responded to the allegations at this time. The incident underscores growing concerns about protecting AI systems from deliberate extraction attempts and maintaining geographic restrictions on sensitive technology access.
A critical macOS vulnerability allowing remote attackers to gain complete system control without passwords is actively being exploited in the wild. The bug affects the built-in screen-sharing functionality across multiple Mac systems.
Multiple water treatment facilities across the United States have been compromised in recent weeks by attackers allegedly connected to the Iranian government. The breach marks a significant intrusion into critical infrastructure systems.
A PBS station is at risk of losing 50 terabytes of archived content after its cloud storage provider, Iron Mountain, became unresponsive and denied access to the data. The station has no backup copies of the material.
Four cybercriminals were arrested in Brazil and three others charged in Europe for exploiting a service provider vulnerability to steal €30 million from Commerzbank customers' accounts.