Anthropic's Disclosure of Alibaba-Linked Distillation Campaign Exposes AI API Vulnerabilities

Anthropic informed the U.S. Senate Banking Committee in a June 10, 2026 letter that operators affiliated with Chinese conglomerate Alibaba conducted what it described as the largest known distillation attack on its Claude models. The campaign involved approximately 28.8 million queries across nearly 25,000 fraudulent accounts between April and June 2026, targeting capabilities in software engineering, reasoning, and cybersecurity. The goal was to harvest outputs to train competing Chinese models at lower cost, rather than stealing weights or code directly.

This method, known as model distillation, leverages API interactions to approximate a frontier model's behavior. CNBC confirmed the letter's details, noting Alibaba's actions were characterized as 'brazen' and 'illicit.' A related analysis detailed the scale and focus on Anthropic's Mythos Preview model capabilities. The White House Office of Science and Technology Policy had already flagged the broader issue in an April 2026 memo, warning of 'deliberate, industrial-scale campaigns to distill U.S. frontier AI systems' by foreign entities, primarily China, using proxy accounts to evade detection.

The incident highlights limitations in current defenses: API rate limits and fraud detection failed to flag activity that appeared as routine usage. Hardware export controls on advanced chips do not address this query-based extraction. Bipartisan responses include proposed legislation for tracking extraction attempts and sanctions, building on the White House memo's call for improved industry-government coordination.