Anthropic accuses Alibaba's Qwen lab of massive AI model distillation heist

Anthropic publicly accused Alibaba's Qwen lab of the largest known AI model-distillation heist, alleging a coordinated campaign to extract Claude's advanced coding capabilities. According to Anthropic's account, roughly 25,000 fake accounts generated more than 28.8 million interactions with Claude between April 22 and June 5, 2026 — a volume designed to systematically harvest outputs for training a competing model.

Mechanically, distillation attacks work by querying a target model at scale and using its responses as training data for a smaller 'student' model, effectively copying capabilities without access to weights. The alleged sophistication here was operational: by keeping each individual account's activity below standard rate limits, the campaign reportedly stayed undetected for weeks, only surfacing through aggregate pattern analysis.

Competitively, the accusation is explosive because Qwen is one of the strongest open-weight model families and a flagship of Chinese AI. If substantiated, it validates long-standing Western concerns that Chinese labs close capability gaps partly by distilling frontier US models. The story intertwines with Alibaba's simultaneous ban on Claude Code, creating a full-blown IP-and-security standoff between the two companies.

Skeptics note Anthropic has offered aggregate claims rather than published forensic proof, and a HN controversy alleges Anthropic hides invisible Unicode watermarks in Claude output to detect exactly this kind of extraction — with outrage centered on the secrecy of the detection method rather than its use. Watch whether Anthropic pursues legal action, whether regulators weigh in, and how the watermarking disclosure affects enterprise trust.