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. It’s a masterclass in modern corporate espionage—except it wasn’t humans stealing documents. It was machines stealing “essence” through a process called distillation.
They fed Claude complex logic problems, coding tasks, and reasoning tests. They didn’t want the answers; they wanted the how. They used Claude’s outputs to train their own models faster and cheaper than starting from scratch.
This changes the ‘Ferraris vs. Toyotas’ debate. Yes, Chinese labs are shipping highly efficient models (the Toyotas) that cost less to compute. But this report suggests they are reverse-engineering the Ferrari engine to do it.
In the AI era, defending IP isn’t about locking the filing cabinet. It’s about building a moat against a swarm of synthetic users designed to quietly strip-mine your intelligence.
