China May Have Accessed Anthropic’s Mythos, Driving White House Export Ban
The White House's export restrictions on Anthropic's Mythos were partly driven by fears a China-linked group accessed the model, according to Semafor.
LUMIEN4 min read
The White House's decision to impose export restrictions on Anthropic's Mythos AI model was driven in part by concerns that a group linked to China may have already accessed it, according to a report from Semafor. The Chinese government has not confirmed any such access, and the White House itself has not verified the report. Trump advisor David Sacks addressed the restrictions publicly on X but made no mention of China, focusing instead on other aspects of the decision.
What happened
Semafor reported that the White House’s move to place export restrictions on Anthropic’s Mythos AI model was motivated at least in part by intelligence suggesting a China-linked group may have accessed the system. The specific models named in the concern are Mythos 5 and Fable 5.
The White House has not confirmed or denied this account. On X, Trump advisor David Sacks publicly addressed the export restrictions but did not reference China as a factor in his post.
Why it matters
If a China-linked group did gain access to Mythos 5 or Fable 5, the national security implications would be significant. According to The Verge, one major risk is model distillation: a technique where a less capable “student” AI model is trained against outputs from a more advanced model, effectively copying much of its behavior without needing direct access to the underlying weights or training data.
Distillation is already a well-documented method in AI research. It is how many smaller, efficient models are built. In the wrong hands, it becomes a way to replicate frontier AI capability at a fraction of the original cost and effort. That makes any confirmed access to a top-tier model a meaningful shortcut for a foreign adversary’s own AI development.
Export controls on AI models are still relatively new territory for the US government. Applying them to software and model weights is more legally and technically complex than controlling physical hardware like chips. This case, if the underlying intelligence holds up, could accelerate how aggressively the US moves to treat frontier AI models as controlled technology.
Our take
A few things worth keeping in mind here. First, this is a single report from Semafor, and the White House has not confirmed it. David Sacks had a public opportunity to cite China as a rationale and did not. That gap between the reported justification and the official one is worth watching.
Second, the distillation risk is real and not hypothetical. We have seen smaller labs use distillation on openly available model outputs to produce surprisingly capable systems. If a state-level actor applied that same approach to a genuinely frontier model, the capability jump could be substantial.
Third, for businesses using or building on frontier AI APIs, this is a reminder that the geopolitical environment around these tools is tightening. Export controls, access restrictions, and national security reviews are now part of the landscape for any advanced AI model. That affects which models stay available, which integrations remain viable, and how quickly access can be cut off.
The unconfirmed nature of the report matters too. Governments sometimes allow security-framed narratives to circulate without confirmation because they are useful politically, not because every detail is solid. Treat the Semafor report as a serious signal, not a settled fact.
What to do about it
If your business relies on API access to frontier AI models, now is a good time to audit which providers you depend on and whether any of them could be subject to sudden access restrictions. Consider:
- Identifying fallback model providers in case a primary API becomes restricted or unavailable.
- Reviewing your terms of service with AI providers for clauses related to government compliance and service interruption.
- Watching how the White House formalizes its AI export control framework over the coming months, as the rules are still being written.
The practical takeaway: do not build a single-provider dependency into any critical workflow right now, because the policy ground is shifting fast.