US agencies name six Chinese AI firms accused of copying American frontier models at industrial scale

Key points

• Three US agencies jointly named six Chinese AI firms on Tuesday, accusing them of large-scale theft of US AI model capabilities.
• The six named companies are DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI.
• Targeted models include versions of Claude, GPT, Gemini, and Grok.
• US agencies allege the Chinese government was likely aware of the activity.
• The practice can dramatically shorten development timelines and cut billions from training costs.

Three US government agencies have pointed fingers at six Chinese AI companies, accusing them of running what they call "industrial-scale" operations to copy the capabilities of leading American artificial intelligence models.

On Tuesday, the National Security Agency (NSA), the Cybersecurity and Infrastructure Security Agency (CISA), and the Federal Bureau of Investigation (FBI) issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, Z.AI, and StepFun. All six are accused of running these operations since at least late 2024.

What exactly are they accused of doing?

The technique at the centre of this is called distillation. Instead of spending years and enormous sums teaching an AI system from scratch, you feed it the answers from a smarter, already-trained model and let it learn by imitation. These Chinese firms reportedly did that at industrial scale, targeting American frontier models: the most capable AI systems built by companies including Anthropic, OpenAI, Google, and xAI.

Models named in the advisory include variants of Claude (Anthropic's assistant), GPT (the family behind ChatGPT), Gemini (Google's AI), and Grok (xAI's chatbot). How the firms accessed these systems wasn't spelled out, but large language models are often reachable through public products or developer APIs.

Does this mean the Chinese government was behind it?

Not quite. The agencies stopped short of saying so directly, stating only that the six firms "likely" acted with "Chinese government awareness." That word likely is carrying a lot of weight, but from three federal agencies, it's still a notable allegation, as first reported by Ars Technica AI.

The financial motive is plain enough. Training a frontier AI model from scratch costs hundreds of millions to billions of dollars. Distillation shortcuts that process substantially, and the agencies said the practice delivers "significantly shorter AI development timelines and reduced financial expenditures" to those who use it. That quote is the agencies' own language, kept verbatim for accuracy.

What does this mean for ordinary users?

If you use any of these apps, particularly DeepSeek, which became widely popular in early 2025, US agencies now formally consider it a national security concern. That doesn't mean the app vanishes from your phone tomorrow. It does mean American regulators are watching these companies closely.

One privacy point worth keeping in mind: DeepSeek's own policy states it stores user data on servers in China. Don't share sensitive work documents or personal details through any of these tools if you're not comfortable with where that data might end up. Our 9 September story on AI text watermarking is a useful companion read: those invisible markers could, in principle, help identify which outputs were copied from US models.

The bigger picture here is less about any individual app and more about the competitive gap these agencies are worried is closing fast.

Common questions

Is DeepSeek safe to use?

US agencies have flagged it as a national security concern, and its privacy policy confirms user data sits on servers in China. Casual use may feel fine, but keep anything sensitive well away from it.

What is AI distillation, in plain terms?

Distillation means training a new AI model on the outputs of an existing, more capable one. It's faster and cheaper than building from raw data, which is why the agencies say it can save billions in development costs.