The US just accused six Chinese AI labs of straight up copying its models, and this fight is only getting louder

Three US national security agencies just accused six Chinese AI companies of copying America's homework, and they published a genuinely detailed paper trail to back it up.

The NSA, CISA and FBI issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, accusing them of systematically extracting capabilities from Claude, GPT, Gemini and Grok since late 2024. The agencies describe it as industrial scale distillation, pulling billions of tokens across millions of queries, and say it was likely carried out with the Chinese government's knowledge.

The specifics are oddly precise for a government advisory. DeepSeek is accused of drawing on multiple versions of Claude, GPT and Gemini to help train its R1 and V3 models. Moonshot AI allegedly pulled from Claude Fable to build its Kimi K3 system, and used GPT-4o output to help develop Kimi K2. This is not a vague accusation about copying homework in general, it names the homework and the students.

The methods described sound less like sophisticated espionage and more like determined workaround engineering: fraudulent accounts, bulk premium subscriptions and proxy routing services nicknamed transfer stations, all built to dodge regional restrictions and detection systems.

Rather than telling American AI companies to simply ban the accounts involved, the advisory recommends something subtler, quietly degrading responses for suspicious accounts instead of blocking them outright. That is a strategy built around not tipping off the other side that they have been caught, which tells you this fight is expected to keep running rather than end with one advisory.

This is the clearest official confirmation yet of how heated the AI race between the US and China has become, and it goes well beyond the usual headlines about chip export bans. If a model's training data can be extracted this efficiently through nothing more exotic than clever account management, the real competitive edge in AI might be a lot more fragile than either side would like to admit.