Anthropic shuts down Fable 5 and Mythos 5 after US order bars foreign access

Anthropic just killed access to two of its most powerful AI models, Fable 5 and Mythos 5, days after publicly launching Fable 5. The reason is blunt and compliance-driven: a US national security order that bans all foreign nationals, including people who work at Anthropic, from accessing those models.

In a blog post, Anthropic said the “net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.” In other words, this is not a permissions tweak or a narrow restriction. It becomes a universal shutdown, because the company cannot operate those models in a way that satisfies the scope of the directive.

Why this matters to executives is that it is a stress test of the entire AI release pipeline. These models are products, but they are also regulated systems running inside a global labor and data reality. Anthropic’s situation highlights a recurring tension in US tech policy: national security controls often land on access and identity, not model capability alone. So when the rule reaches “all foreign nationals,” even “foreign persons within the country,” the operational answer can be shutdown rather than granular gating.

The immediate policy mechanism, according to reporting summarized in the article, traces to action by the US Commerce Department. Earlier on Friday, Axios reported that the Trump administration was blocking foreign governments, companies, and individuals from accessing Anthropic’s most advanced AI models. The Commerce Department took action after another company claimed it was able to jailbreak Mythos, Axios reported, citing an administration official.

The article says US Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei. That letter reportedly states that Mythos 5 and Fable 5 would be subject to export controls to any location outside the US, and to all foreign persons within the country. That framing matters because export controls usually revolve around geography and controlled persons. Here, the compliance burden spreads into day-to-day access management, which creates a practical “all or nothing” outcome.

This comes after a broader executive push. The Trump administration signed an executive order early in the month asking leading AI developers to voluntarily submit their most capable models for government cybersecurity tests before releasing them to the public. Voluntary on paper does not always stay voluntary once agencies start testing, labeling risks, or acting through Commerce authorities. Add jailbreak allegations into the mix, and the incentives tilt toward stricter controls and faster compliance.

The background is not new either. A months-long dispute between Trump administration officials and Anthropic had been showing signs of easing across parts of the US government as the company prepares to go public, But risk labels and national security framing do not disappear just because a company is nearing a public-market milestone. The article notes that the Defense Department in March labeled Anthropic a “supply-chain risk,” a signal that regulators were already treating the company and its model pipeline as part of a broader security picture.

Now, consider the second-order implications for boards and rival model labs. Anthropic’s decision suggests that even robust teams can face forced downtime if they cannot meet a regulator’s access constraints with enough precision. It also hints at an execution problem across the AI industry: identity-based restrictions collide with international workforces, contractors, and distributed access patterns. If your go-to-market depends on global talent and globally distributed customers, you may be one compliance rule away from a product pause.

For decision-makers, the strategic stake is immediate and reputational as well as operational. A three-day gap between publicly launching Fable 5 and disabling both Fable 5 and Mythos 5 shows how quickly regulatory timelines can override product timelines. The executive takeaway is simple, but not comfortable: in this policy environment, “release readiness” now includes regulatory access design, not just model safety and performance. If you are an AI developer, investor, or operator, the question is no longer only whether your models work. It is whether you can keep them running under shifting national security requirements without turning every rollout into a compliance hostage situation.