Anthropic Uncovers 'Largest Known Distillation Attack'
American AI companies say their tech edge is being quietly siphoned off—by both rivals and China. In a June 10 letter to Sens. Elizabeth Warren and Tim Scott, Anthropic accused Chinese giant Alibaba of covertly copying its proprietary AI via "distillation," a long-used technique in which one model teaches another how to behave. Anthropic alleges Alibaba and other Chinese firms tapped its systems through tens of thousands of unauthorized accounts, generating millions of Claude chatbot interactions to train their own models—echoing claims OpenAI has made about Chinese startup DeepSeek, per the New York Times. Anthropic specifically accuses Alibaba of "the largest known distillation attack" on it to date, per CNBC. Distillation is widely used to compress and copy open-source systems, but US labs bar it for their top proprietary models. Anthropic bans Chinese companies from using its models altogether, though companies can exploit loopholes to access them. Chinese fintech group Ant, for example, "provided employees with corporate Claude accounts that were accessed through the company's intranet," the Financial Times reports. In response, Anthropic tried using hidden code to identify China-based users, but backtracked last week in the face of privacy concerns, per the Washington Post. It says distillation attacks "pose a serious threat to national security." Alibaba, which has its own AI assistant, now says it will block employees from using Anthropic's AI tools for work beginning Friday, CNBC reports.
Source: Newser