← Back to feed

The Trump administration has moved to block foreign access to Claude Fable 5 and Claude Mythos 5, treating Anthropic’s most capable AI as restricted dual-use technology. The directive bars foreign governments, companies, and individuals from accessing either model, according to an Axios scoop published Saturday.

The action arrives four days after Anthropic shipped both models publicly and creates a fast-moving regulatory paradox: the same US government that restricted access to Mythos for months over cybersecurity concerns is now restricting its most powerful version from leaving the country — while simultaneously keeping Anthropic off the approved vendor list for its own agencies.

Two Restrictions Running in Parallel

The foreign access suspension sits alongside a separate domestic restriction. Anthropic has been added to a Pentagon blacklist, where the company is classified as too dangerous for US government procurement. A Commerce Department licensing regime is now in place to govern who can access the models and under what terms.

This creates a three-tier reality:

• Glasswing partners (roughly 200 organisations in 15 countries, predominantly US-aligned): full Mythos 5 access, the model with cyber-offensive capabilities intact
• US public and enterprise: Fable 5, which Anthropic describes as the same architecture but with safeguards that prevent it from being used for cyberattacks
• Foreign entities: blocked from both, pending Commerce Department review

The Pentagon blacklist is the sharper of the two restrictions. NSA has deployed Mythos for cyber defence work, while the Defence Department’s procurement arm has flagged Anthropic as a supply chain risk. Those two positions now sit alongside a third: a Commerce-managed access regime that treats Mythos-class models as exportable only under licence.

The Safety Tier Anthropic Built In

Wired’s reporting frames the public Fable 5 release as Anthropic’s answer to the access problem: a version of the model that passed the company’s own red-teaming for cyberattack capability and was cleared for general release. Mythos 5 — same weights, different fine-tuning — retains the capabilities that earned it classified access inside the Glasswing programme.

Anthropic has been explicit in the Fable 5 system card that the safety restrictions include refusals on a class of requests it does not fully specify in public documentation. The model card filed in June acknowledges that 0.03% of API calls currently trigger silent capability nerfs related to “AI development requests,” with that definition expanding over time.

That disclosure, and the existence of Mythos as a parallel product for trusted partners, was already attracting scrutiny. The foreign access suspension adds a geopolitical dimension the company had not publicly anticipated.

The Commerce Department’s Move

What changed between Anthropic’s original plan to expand Mythos access to 120 organisations — blocked by the White House in April — and the current foreign suspension is the involvement of Commerce’s Bureau of Industry and Security. BIS licensing for AI models would represent a significant step: it places Anthropic’s software on a similar administrative footing to advanced semiconductor exports, requiring case-by-case government review for each foreign recipient.

The practical effect is that non-US allies, companies operating in foreign jurisdictions, and foreign governments cannot access either model through Anthropic’s API or any licensed downstream provider. The geographic scope of the ban has not been published.

What It Means for the Market

Fable 5 is currently the highest-scoring model on the Artificial Analysis Intelligence Index at 64.9, and its Agent Arena composite of 12.94% leads every other model. The foreign access suspension means the global developer market is now split between providers who can and cannot access the frontier.

Anthropic’s $965 billion valuation and imminent S-1 process make this a material disclosure risk. International enterprise contracts, particularly in markets like the EU, India, and Japan, now depend on a US government licensing decision rather than a commercial agreement.

The company has not commented publicly on the directive. The Glasswing programme’s international members — including UK AISI and European cybersecurity agencies — are understood to retain access under the existing trusted-partner framework, but the legal basis for that access under the new Commerce regime has not been clarified.