- Lead. Anthropic was ordered by the US government on 13 June 2026 to immediately disable its two most capable AI models — Fable 5 and Mythos 5 — for all foreign nationals worldwide, including the company’s own non-citizen employees, under an export control directive citing national security.
- Fact. The directive, received by Anthropic at 5:21 PM Eastern time, cited a reported method for bypassing Fable 5’s safety guardrails — a capability described as potentially constituting a dangerous cyberweapon if extracted by a hostile foreign actor.
- Stake. Anthropic noted that OpenAI’s GPT-5.5 possesses comparable vulnerability-detection capabilities and faces no equivalent export restriction — a disparity that could materially disadvantage Anthropic in the enterprise market at a sensitive moment ahead of its IPO.
Anthropic said it was forced to “abruptly disable” access to the two models across its entire customer base to comply with the directive, according to reporting by Al Jazeera. The company stated it received only partial information about the government’s underlying reasoning and disagreed with the action, but complied. It did not indicate when access might be restored.
What the models do and why they were targeted
Fable 5, Anthropic’s publicly available flagship released earlier in June, incorporates advanced capabilities in cybersecurity vulnerability detection and biotechnology analysis. Mythos 5 is a non-public variant limited to government and selected institutional partners. The government’s stated concern focuses on a reported jailbreak method for Fable 5 that, if exploited, could allow a foreign actor to use the model’s vulnerability-scanning capabilities without safety constraints — effectively repurposing it as a tool for identifying exploitable weaknesses in critical systems.
Anthropic disputed the severity of the concern, saying its safety measures for Fable 5 had been “extensively tested.” The company also pointed out the selective nature of the directive: OpenAI’s GPT-5.5 has similar capabilities and is not subject to the same restrictions. That asymmetry has direct commercial consequences. Anthropic filed a confidential S-1 at a reported $965 billion valuation, and enterprise customers operating with multinational teams cannot currently use the company’s most advanced models.
What it means for AI export control policy
The directive extends the regulatory logic of Trump’s January order restricting exports of advanced AI hardware to the software layer. That transition raises harder enforcement questions. Chips can be physically tracked and intercepted; AI models can be copied, fine-tuned, and redistributed without the friction that makes hardware controls legible. Applying export controls to models also creates a definitional challenge that the hardware order avoided: at what capability threshold does a software tool become a controlled dual-use item, and how is that threshold set consistently across competing products from different companies?
Anthropic said all of its other models — including earlier versions of the Claude and Fable product lines — remain fully accessible to international users. The company did not confirm whether it intends to pursue legal remedies, beyond its statement of disagreement with the directive.