White House Cites National Security to Halt Public Access to Anthropic's New AI Model

Today’s Focus

On Friday, June 13, the Trump administration moved against one of the country’s leading AI developers, telling Anthropic that its newest models could not be used by foreign nationals at home or abroad, according to The Atlantic and Reason.

The directive targeted Claude Fable 5, released June 9, and the more powerful Claude Mythos 5 that preceded it. Anthropic had described Fable 5 as a general-use version of Mythos, which it had restricted in April to a small set of cyberdefenders and infrastructure operators working with the U.S. government.

Reason reported that the White House sent Anthropic a letter invoking a new export-control directive. The order barred any foreign national inside the United States, including Anthropic staff on work visas, as well as anyone overseas, from using the models.

Anthropic said it could not immediately separate eligible from ineligible users, so it suspended public access for everyone rather than risk violating the order.

The administration labeled the models a threat to national security. Anthropic said the government’s letter offered no specifics, and that the company’s own read was that officials had learned of a method to “jailbreak,” or bypass, Fable 5’s safety guardrails.

The Atlantic noted the move clashed with President Trump’s stated goal of accelerating American AI to outpace China, a priority he stressed early in his second term.

The Debate

Supporters argue

Backers of the directive frame it as a straightforward use of export-control authority to keep a dangerous capability out of adversaries’ hands. The administration says Fable 5 and Mythos 5 represent a national security risk, and that limiting who can run them is a reasonable precaution while officials assess a possible jailbreak.

The underlying worry is concrete. Anthropic itself had called Mythos so capable at hacking that it initially withheld the model from general release, granting access only to vetted cyberdefenders in coordination with the government.

If a method existed to strip away Fable 5’s restrictions, supporters contend, the same offensive cyber capabilities could become available to hostile states or criminal groups. Under that logic, blocking foreign access first and resolving the technical questions later treats the matter with the urgency a security threat warrants. The view holds that a frontier model with demonstrated hacking strength is exactly the kind of technology export controls exist to govern.

Critics argue

Critics see the timing and reasoning as suspect. Reason described the action as looking like “payback” for Anthropic’s earlier refusal to let the government use its products for mass surveillance and autonomous weapons.

They point out that the company built safety guardrails for Fable 5, submitted them to third-party testing that included the U.S. government, and still drew complaints from cybersecurity researchers that the model was too locked down, as The Atlantic reported.

The White House letter, by Anthropic’s account, named no specific concern, which critics argue makes meaningful compliance impossible and the rationale hard to evaluate.

There is also a coherence problem. Trump has cast speed in AI as a national emergency and pledged to clear obstacles for U.S. developers, yet this order forces an American company to cut off its own visa-holding employees and its public users. Critics call that the opposite of winning the race against China.

What the experts say

Independent analysts have warned for years that export controls on software and AI models are difficult to enforce because the underlying weights and techniques can spread quickly. A 2023 RAND Corporation analysis of AI and compute governance found that controlling finished models is far harder than controlling the specialized chips used to train them, since models can be copied and moved without physical shipment.

The Center for Security and Emerging Technology (CSET) at Georgetown University has documented that “jailbreaking” remains a persistent, unsolved problem across major language models, with researchers repeatedly demonstrating bypasses of published safety measures.

Brookings Institution scholars studying AI regulation have noted that U.S. export policy on advanced computing has leaned heavily on hardware restrictions, such as the chip controls expanded since 2022, rather than on restricting model access by user nationality. That history suggests Friday’s directive applies an unusual mechanism, and its practical effect on a publicly released model is untested.

By the Numbers

June 9, 2026: the date Anthropic released Claude Fable 5 for general use, according to Reason.

June 13, 2026: the Friday the White House sent its letter restricting the models, per Reason and The Atlantic.

April 2026: when Anthropic first announced Claude Mythos 5 and limited it to cybersecurity partners, as reported by The Atlantic.

2 models: Fable 5 and Mythos 5 are both covered by the new export-control directive, according to Reason.

0 specifics: Anthropic said the government’s letter provided no detailed description of its national security concern.

Day 1: Trump declared an AI “emergency” and pledged to speed the industry on the first full day of his second term, The Atlantic reported.

2022: the year U.S. advanced-chip export controls on China were significantly expanded, the policy framework analysts compare this action against, per Brookings research.