Regulation
US accuses six Chinese companies of copying its AI and proposes throwing them off
The NSA, CISA and the FBI say DeepSeek, Alibaba and four other Chinese companies train their models on answers from Claude, GPT, Gemini and Grok. They propose secretly altering what those companies receive.

The essentials
- The NSA, CISA and the FBI issued a joint advisory on October 28, 2026, targeting six Chinese AI companies (avisos de ciberseguridad).
- Claimed by one partyAccording to the agencies, those companies have been copying answers from Claude, GPT, Gemini and Grok on a large scale since at least late 2024 (avisos de ciberseguridad).
- Claimed by one partyThe agencies call DeepSeek's claimed $5.6 million cost to train its model misleading, because it leaves out the copied data (avisos de ciberseguridad).
- They ask US companies to watch for suspicious accounts, share information and subtly alter the answers given to anyone trying to copy them (avisos de ciberseguridad).
Why it matters
The agencies present distillation, which means training one model on another model's answers, as the core of China's strategy rather than an add-on. If confirmed, part of China's progress would depend on other companies' models. It would also cast doubt on the idea that cutting-edge models can be built cheaply (avisos de ciberseguridad).
The details
According to the joint advisory published on October 28, 2026, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI have extracted billions of tokens (pieces of text) through millions of queries to US models. The agencies say DeepSeek used this to train R1 and V3, and Alibaba used it to improve its Qwen family. They believe the Chinese government "probably" knew (avisos de ciberseguridad).
To avoid detection, the agencies say, the companies spread their queries across several channels. They use intermediary services that hide who is asking, and a gray market of bridge connections, known as "transfer stations," to get around country-level blocks. They also buy paid subscriptions in bulk and share them across teams. All of this breaks the US companies' terms of use (avisos de ciberseguridad).
What we don't know
- How DeepSeek, Alibaba and the other accused companies respond.
- What specific evidence backs the accusation beyond what the advisory summarizes.
- Whether OpenAI, Google, Anthropic or xAI will adopt the proposal to alter answers, and how it would affect legitimate users.
Sources
Was this story useful? Yes Not really