Anthropic accuses Alibaba of illicitly distilling Claude via 25,000 fake accounts

Anthropic sent a letter to U.S. Senate leaders Tim Scott and Elizabeth Warren claiming that Chinese AI giant Alibaba illicitly trained its own models on Claude's output using 25,000 fraudulent accounts and 28.8 million API exchanges between April and June 2026. The technique, called distillation, trains a cheaper or faster model on the outputs of a more advanced one. Anthropic traced the activity back to operators connected to Alibaba and Alibaba Qwen, Alibaba's AI lab.

This is the second time in 2026 Anthropic has publicly accused Chinese labs of distillation fraud. In an earlier complaint, Anthropic claimed DeepSeek, Moonshot, and MiniMax used 24,000 fraudulent accounts and 16 million exchanges to distill Claude. While distillation has legitimate uses (e.g., creating lightweight versions of frontier models), Anthropic argues that competing labs can use the same technique to build frontier-grade models 'at a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.'

For infrastructure and security teams, the distillation vector highlights a critical edge: frontier model outputs are commodified through API access, and scale fraud (thousands of throw-away accounts, millions of low-cost queries) can extract the intelligence asymmetrically. Anthropic warns that successful distillation may help China close the frontier-model gap faster than expected. The companies dispute this claim, but the scale of documented activity (28.8M exchanges) suggests systematic access to Claude's inference stack.