News · 2026-06-28
The ban on Anthropic's most powerful model just partially lifted -- for Americans only
Earlier this month the U.S. government did something that had never happened before: it pulled a commercial AI model off the market on national-security grounds. Anthropic's most capable systems, Mythos 5 and Fable 5, were suddenly off-limits -- not just to customers abroad, but to Anthropic's own foreign-national employees. We covered that ban when it landed. This week the story moved: according to Anthropic's official statement, the restriction has partially lifted.
The shape of the partial lift matters. As of late June, more than a hundred U.S. institutions have had access restored. But the relief stops at the border: foreign nationals, including some of Anthropic's own staff, remain blocked. So the model did not come back for everyone -- it came back for an approved American list, and the line is drawn by nationality rather than by company or use case.
To understand why a government reaches for a tool this blunt, you need the original trigger. The concern was that a sufficiently capable model could be steered -- jailbroken -- into helping with offensive cyber operations: finding and chaining software vulnerabilities faster than defenders can patch them. A model that compresses weeks of an expert's work into an afternoon is, in the government's framing, closer to a controlled technology than to ordinary software. Export-control law, the same body of rules that governs advanced chips and certain encryption, was the lever they pulled.
Here is the counter-argument, and Anthropic itself has gestured at it: the specific weaknesses the government worried about are the kind that other, freely available models can already find. If the dangerous capability is also sitting inside open-weight systems that anyone on earth can download, then locking up one American company's product is closing one door in a building with the walls knocked out. That tension is the whole debate in miniature -- a national-security apparatus built around scarce, controllable technology, colliding with a field where the capability is increasingly cheap, open, and everywhere.
And the rest of the world did not wait politely. In the days around the restrictions, rival labs moved to fill the vacuum, pitching their systems directly at the customers and tasks the American models had vacated. Open models from Chinese labs in particular have been gaining ground on exactly the kind of security work the controls were meant to fence off -- the same week brought a clean example of a free Chinese model beating Claude on a vulnerability-finding test. The likely net effect of a unilateral control, critics argue, is not to slow the capability down but to relocate it -- and to hand the relocated version a marketing story about American models being unreliable, here today and gone tomorrow.
The so-what is bigger than one product line. This is the clearest sign yet that frontier AI has crossed from the commercial column into the geopolitical one, and that the map of who can use what is now being drawn in capitals, not just in pricing pages. The honest caveat is that the situation is fluid and the official statements are terse: the exact list of restored institutions, the precise legal basis, and what full restoration would even require are all still moving. What is not in doubt is the precedent -- a government can now switch a leading AI model on and off, and has shown it will.