US Pulls Claude Mythos, China Restricts GLM and Qwen: AI Models Become Strategic Weapons

In 2026, the US and China have tightened control over their largest AI models in lockstep: Washington closed open access to Anthropic's Claude Mythos over cybersecurity risks, while Beijing is preparing an export ban on GLM, Qwen and DouBao, the three flagships of China's AI industry.

What happened in the US

Anthropic has pulled the Claude Mythos model from public access, officially citing potential cybersecurity threats. The move is unusual: in recent years the industry has deliberately pushed toward greater openness, with public APIs, democratized access and open weights for some models.

What exactly about Claude Mythos raised concerns has not been officially disclosed. The nature of the restriction suggests that corporate and government customers keep access under separate terms: the model was "removed from general access" rather than fully deactivated. The precedent matters: for the first time, a leading American AI lab is publicly restricting a top model on national security grounds rather than commercial ones.

• Claude Mythos is closed to the general audience
• The official reason is cybersecurity threats
• Corporate and government customers reportedly keep access
• The specific capabilities that prompted the shutdown have not been disclosed

What China is preparing

Beijing is following a symmetrical course. Chinese authorities are preparing an export control mechanism for domestic AI models. The restrictions will cover GLM (Zhipu AI), Qwen (Alibaba) and DouBao (ByteDance), the three biggest players in China's AI industry.

All three products are widely used outside China. Alibaba's Qwen competes with Western models on key benchmarks and is available as open weights on Hugging Face. DouBao is ByteDance's mass-market consumer product with hundreds of millions of users. GLM is used in the enterprise sector in a number of countries. Export restrictions would effectively cut international developers and companies off from official access to these tools.

In essence, this carries over to software the same logic already applied to semiconductors: first NVIDIA chips came under export controls aimed at China, and now the models themselves are becoming the subject of bilateral regulation.

"Welcome to the era of digital sovereignty, where neural networks are

officially recognized as strategic weapons," the analysis authors sum up.

What this changes for the market

These are not isolated events but part of a systemic shift. Both countries are moving toward a model in which powerful language systems are treated as strategic national security assets, on par with military technology, telecommunications infrastructure and advanced semiconductors.

For international teams and companies building products on external AI APIs, this means growing operational uncertainty: a tool available today may be restricted tomorrow without warning. Diversifying the AI stack and being ready to switch providers quickly are no longer an architectural option but a production necessity.

At the same time, the global AI market is being reconfigured: a single open space is giving way to several geopolitical zones with different sets of available models, regulatory regimes and export restrictions.

What this means

The global AI market is entering a phase of lasting geopolitical fragmentation. The US and China have set a precedent: the most powerful neural networks are not just technology products but strategic resources with controlled distribution. Developers and companies worldwide should build regulatory risk into their architecture decisions now.