US agencies name six Chinese firms in AI theft advisory

Six Chinese AI companies extracted billions of tokens from US models since 2024, according to a joint advisory from the NSA, CISA and FBI. DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI are named. The advisory, designated AA26-251A, lists which US model each firm allegedly targeted, pulling from Claude, GPT, Gemini and Grok across millions of requests.

For the six firms, this is a federal government putting their names on a specific accusation rather than a general warning about Chinese AI. For Anthropic, OpenAI, Google and xAI, it's confirmation their production APIs were the extraction point at industrial scale, not lab benchmarks.

The agencies draw a line between two things. Distillation, training a smaller model on a larger one's outputs, is a normal research technique. What the advisory describes is different: extraction of restricted proprietary capabilities at a scale it calls malicious, not academic. It puts all six firms on the wrong side of that line.

The sharper claim is about cost. DeepSeek has said its training run cost $5.6 million. The advisory argues that figure only holds if you ignore what it cost to build the data the model trained on, tokens pulled from systems that took far more than $5.6 million to build in the first place.

If that argument holds, it reframes more than one company's public relations. The next lab claiming a similarly cheap training run will need to show where its training data came from, not just what the run itself cost.

The advisory doesn't say how the agencies counted the tokens, or whether enforcement follows the naming. Watch whether Anthropic, OpenAI or Google change API rate limits or terms of service in response. That would be the first sign this moves past attribution.