US intelligence and cybersecurity agencies have accused several Chinese artificial intelligence companies of carrying out large-scale campaigns designed to extract capabilities from leading American AI models.

The National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation issued a joint advisory claiming that companies including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI have used American frontier models to help train their own systems.

According to the advisory, the activity involved billions of tokens generated across millions of requests since 2024.

The agencies allege that the Chinese companies used outputs from models including Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini and xAI’s Grok.

The technique is known as distillation, where the outputs of a more powerful model are used as training material for another model. Distillation is a legitimate method in AI development, but the US agencies say the alleged activity involved efforts to extract proprietary capabilities at scale.

DeepSeek is accused of using information from Claude, Gemini, GPT and Grok models to help train its own systems, including R1, the open-source reasoning model it released in early 2025.

Alex Shuper

Moonshot AI is also named in the advisory. US authorities allege the company extracted significant amounts of data from Anthropic’s Claude Fable model to help train Kimi K3, while Kimi K2 is alleged to have used data generated by GPT-4o.

Claude Fable was developed by Anthropic to make some capabilities from its more advanced Mythos cybersecurity model available more broadly, while Mythos itself is restricted to participants in Project Glasswing.

The joint advisory also outlines defensive measures that American AI companies can use to detect and limit what the agencies describe as malicious distillation campaigns.

Concerns about model copying are not new. OpenAI previously said it had worked with Microsoft to ban accounts suspected of attempting to distil capabilities from its systems, with DeepSeek among the companies investigated.

Anthropic has also accused DeepSeek, Moonshot AI and MiniMax of conducting large-scale distillation activity.

The issue is not limited to Chinese developers. OpenAI has separately pointed to xAI’s use of outputs from its models during the ongoing legal dispute involving Elon Musk.

The latest advisory highlights growing concern in Washington that competition in generative AI is moving beyond chip access and computing power to include attempts to replicate the behaviour of advanced proprietary models through their outputs.

Featured Photo by Getty Images