US agencies say Chinese AI firms are quietly copying American models at industrial scale
Intelligence officials accuse China-based labs of running 'distillation attacks' against Claude, GPT, Gemini and Grok to shortcut years of research.
Key points
- US cybersecurity and intelligence agencies have accused China-based AI companies of systematically copying capabilities from American frontier models including Claude, GPT, Gemini and Grok.
- Officials describe the activity as happening at industrial scale and call it a central pillar of Chinese AI development strategy.
- The technique in question is known as model distillation, where one AI is used to train another to behave like it.
- No specific companies were publicly named in the accusations, and no criminal charges have been announced.
- The claims sharpen an ongoing dispute over how much of China's rapid AI progress rests on borrowed work from US labs.
US cybersecurity and intelligence agencies have publicly accused artificial intelligence companies in China of quietly copying the inner workings of America's most advanced AI systems. The targets, according to the agencies, include Anthropic's Claude, OpenAI's GPT, Google's Gemini and xAI's Grok.
The method has a name: distillation. In plain terms, you take a powerful AI model, ask it huge numbers of questions, record its answers, and use those answers to train a new, cheaper model that behaves much like the original. It is a legitimate research technique when done on your own model. Doing it to somebody else's model, at scale, without permission, is a different matter.
US officials say this is exactly what is happening, and that it is not a side project. They describe it as "systematic extraction" running at industrial scale, and call it a "core" part of how several Chinese AI firms are catching up so quickly. The claims were reported by The Hacker News.
What is a distillation attack, in plain English?
Think of it as reverse-engineering by conversation. Attackers create thousands or millions of accounts, or use paid access to a top US model, and bombard it with carefully chosen prompts. The answers get saved. Those answers become the textbook for a new model back home.
The copy will never be a perfect clone. But it can inherit the tone, the reasoning style, the safety habits and much of the useful knowledge of the original. That is a huge shortcut. Training a frontier model from scratch costs hundreds of millions of dollars and takes years. Distilling one takes weeks and a fraction of the compute.
This matters for identity and access too, which is my usual beat. Every large AI provider now sits behind account signup, API keys and rate limits. If foreign firms are farming outputs at industrial scale, they are almost certainly doing it through networks of fake or straw-buyer accounts, resold API keys, and stolen credentials. Strong identity checks on signup, device binding, and aggressive anomaly detection on API usage are the front line here. Multi-factor authentication on developer accounts would not stop a determined buyer with a stolen credit card, but it does raise the cost of the bulk account creation that fuels this kind of scraping.
Which models are named?
Four: Claude from Anthropic, GPT from OpenAI, Gemini from Google, and Grok from xAI. These are the current generation of "frontier" models, the term the industry uses for the most capable systems on the market.
The agencies did not publicly name the Chinese firms accused of the copying, and no criminal charges have been announced alongside the statement. That leaves the accusation sitting in an awkward middle ground: serious, specific about the technique, vague about the culprits.
| Detail | What officials said |
|---|---|
| Alleged activity | Systematic extraction via model distillation |
| Scale | Described as industrial |
| US models named | Claude, GPT, Gemini, Grok |
| Chinese firms named | None publicly |
| Charges filed | None announced |
Should ordinary users be worried?
Not directly. Your ChatGPT or Gemini account is not being drained of personal data by this activity. The target is the model's behaviour, not your chat history.
The wider worry is strategic. If a rival country can shortcut years of research by copying American models, the gap that US firms and regulators assumed they had, on both capability and safety, shrinks fast. That has knock-on effects for export controls, for how AI providers police their own APIs, and for how much any of us should trust a chatbot whose lineage is unclear.
Expect tighter API terms, more aggressive account verification on the big AI platforms, and more noise from Washington in the months ahead.
