24,000 fake accounts and 16 million conversations. None of them were real.
Three Chinese AI labs ran coordinated campaigns to extract Claude’s capabilities. DeepSeek, MiniMax, and Moonshot. Anthropic documented the whole operation.
The technique is called distillation. You feed a smarter model millions of carefully designed questions, collect its answers, and use those answers to train your own model. You are not stealing the code. You are stealing the thinking that is producing by the code, and then reverse engineering the code.
→ MiniMax ran the largest operation. Over 13 million exchanges routed through proxy networks managing 20,000 simultaneous connections. When Anthropic released a new Claude model, MiniMax pivoted within 24 hours. Nearly half its traffic redirected overnight.
→ DeepSeek was more surgical. 150,000 exchanges. It prompted Claude to spell out the reasoning behind its own answers, step by step. It also had Claude rewrite politically sensitive queries, generating training data for its own censorship filters.
→ Moonshot filled the middle ground. 3.4 million exchanges targeting Claude’s coding and agentic reasoning.
China’s AI progress has been an impressive technology story. Now we know at least part of the method.
The US AI labs themselves built their models by ingesting the open internet. Books, articles, code, forum posts, creative work. Often without asking. Often without paying. The copyright lawsuits are still stacking up.
Now those same labs are outraged that someone trained on their outputs.
The complaint seems legitimate, as does the irony.
And the uncomfortable question is whether distillation is fundamentally different from what the labs did first, or just the same move played back from a direction they didn’t expect.
♻️ Repost if someone in your network is following the AI race. 🔔 Follow Felix Ghauri for AI and what changes next. 🗂️ Sources in first comment.
First Comment
Sources: Anthropic blog “Detecting and Preventing Distillation Attacks” (Feb 2026), The Rundown AI (24 Feb 2026), OpenAI congressional testimony on DeepSeek distillation (Yahoo Finance, Feb 2026). Image credit: Venture.
