Anthropic says Chinese labs used fake accounts to copy Claude

In short: Anthropic says three Chinese AI labs used thousands of fake accounts to pull millions of responses from Claude and use them to train competing models.

What happened

Anthropic said it disclosed on February 23, 2026, that Chinese AI labs DeepSeek, Moonshot, and MiniMax carried out large-scale “distillation attacks” against Claude. Anthropic said the groups used more than 24,000 fraudulent accounts to generate about 16 million back-and-forth exchanges with Claude.

“Distillation” is a common training method where a weaker AI learns by copying the answers of a stronger AI (like a student learning from an answer key). Anthropic said this can be legitimate when done inside one company, but it becomes misuse when a competitor collects outputs through fake accounts, breaking terms of service and regional access rules.

Anthropic said the activity focused on advanced skills, including coding, data analysis, tool use, computer vision (understanding images), and chain-of-thought reasoning (the step-by-step reasoning that leads to an answer). Anthropic said Moonshot alone generated more than 3.4 million exchanges using hundreds of accounts and multiple routes, including commercial proxy services and distributed networks designed to avoid detection.

Why it matters

Anthropic said the main risk is that copied models may ship without safety checks that try to reduce harmful uses, like making bioweapons, carrying out cyber attacks, or spreading disinformation. The company also warned that models trained to dodge censorship, for example by producing “safe” answers about dissidents or authoritarianism, could be used for surveillance or other government purposes. Anthropic said this also weakens efforts like US export controls on AI chips, since copying can reduce the need to build capabilities from scratch.

Source: TechCrunch AI