APREX · EST. 2026
A new report from news outlet Semafor reveals that the White House imposed export restrictions on Anthropic's advanced AI model Mythos — partly because a group with ties to China is said to have gained access to the system. The story has been covered by The Verge, and would represent a serious national security breach if confirmed.
What Is Mythos — and Why Is It So Sensitive?
Mythos is one of Anthropic's most advanced language models, regarded as a top-tier system with potentially strategic value for actors seeking to develop competing AI capabilities. A comparable model referred to as Fable 5 is also mentioned in connection with the case.
If Chinese actors have indeed gained access to the model, the concern is twofold: the system could be used directly, or it could serve as the basis for what is known as distillation — a training method in which a simpler "student" model is trained on the output of a more advanced model in order to mimic its behavior and capabilities.
Distillation makes it possible to copy advanced AI capabilities without having access to the original source code or training data.
In practice, this means that access to a model like Mythos could allow an actor to build a highly capable model on their own — without having developed it from scratch.
Export Restrictions as a Security Response
The White House's decision to impose restrictions on Mythos is, according to the Semafor report, directly linked to the threat assessment surrounding the possible Chinese access. It has not yet been officially confirmed whether access actually took place, or to what extent any data was exfiltrated. The claims should therefore be treated with caution until further documentation is available.
Chinese AI Is Already at the International Frontier
Underpinning these concerns is the fact that China's own AI companies have already made enormous strides. If Chinese actors additionally have access to Anthropic's most advanced systems, it could accelerate their development even further.
Chinese models such as Alibaba's Qwen3-Max, DeepSeek's V4 Pro, and Zhipu AI's GLM-4.6 rank near the top of key international benchmarks in coding, mathematics, and general reasoning — and in certain tests outperform Western models such as GPT-5.1 and Claude Sonnet, according to research summaries of the field.
DeepSeek V4 Pro (Max) scored 87 on BenchLM's Chinese leaderboard as of March 2026, not far behind Google's Gemini 3.1 Pro (93). Moonshot AI's Kimi K2.6 scores 88.7 on coding benchmarks, while Alibaba's Qwen3-Max-Coder achieves 95.4 percent on HumanEval — both highly impressive results.
Serious If Confirmed
The case currently sits in a gray zone between a confirmed security incident and an unverified intelligence claim. Anthropic has not publicly commented on the matter, and the White House's official justification for the export restrictions has not been elaborated in detail.
If it turns out that a group with Chinese state connections did in fact have access to one of the United States' most advanced AI models, it would constitute one of the most serious AI security breaches on record — and would likely intensify the debate over how proprietary frontier models should be protected against state actors.