Anthropic Accuses Alibaba of Largest Known Distillation Campaign Against Claude, Exposing Industrial-Scale IP Extraction in US-China AI Rivalry

Anthropic has detailed what it describes as the largest adversarial distillation effort yet by a Chinese firm, alleging operators tied to Alibaba's Qwen lab used nearly 25,000 fraudulent accounts for 28.8 million interactions with its Claude model between April and June 2026. The campaign targeted advanced capabilities in software engineering and agentic reasoning, according to a letter sent to U.S. senators and White House officials and reported across Bloomberg, Reuters, and The Next Web.[1][2][3]

Adversarial distillation involves querying frontier models at scale to train cheaper replicas, bypassing the massive R&D costs of original development while often omitting safety alignments. Anthropic framed this as part of a pattern of illicit, systematic harvesting of U.S. AI capabilities, building on its February 2026 report identifying similar campaigns by DeepSeek, Moonshot, and MiniMax.[4]

The accusations arrive amid escalating U.S.-China tech tensions: the Defense Department recently added Alibaba to its list of firms supporting the PLA (prompting a lawsuit from the company), Commerce Department export controls on Anthropic's Fable 5 and Mythos 5 models, and bipartisan legislative proposals for sanctions on entities using U.S. model outputs for training. Anthropic, OpenAI, and Google have begun sharing intelligence on such violations.[5]

While Alibaba declined comment, the episode illustrates state-linked or state-tolerated IP extraction at industrial scale, accelerating China's AI catch-up without equivalent foundational investment. This dynamic extends beyond individual labs to a broader tech cold war, where proxy accounts and infrastructure evade restrictions, potentially eroding U.S. leads in safe, aligned frontier systems.