Anthropic Pulls Fable 5 and Mythos 5 After White House Order

Anthropic has taken its two newest models, Fable 5 and Mythos 5, offline after the Trump administration issued a Friday afternoon directive ordering the company to block access by foreign nationals. It is the most aggressive move Washington has made yet to wall off frontier AI from the rest of the world.

Fable 5 had only been public for a matter of days. It is the lighter sibling of Mythos 5, a model Anthropic had already been keeping on a short leash over cybersecurity concerns. Among those concerns: reports that Mythos 5 can take a freshly disclosed vulnerability and produce a working exploit in hours, not days. That capability is almost certainly what put it on the government's radar, although the directive itself reportedly stopped short of spelling out the national security reasoning.

Anthropic is not thrilled

In a public statement, the company said governments absolutely should be able to block unsafe deployments, but only through a process that is transparent, fair and grounded in technical facts. It argued this order met none of those tests, called the whole thing a misunderstanding, and said it is hoping to get the models back online soon. The Commerce Department has not commented.

The timing is, to put it gently, awkward. Ten days before the takedown, Trump signed an executive order setting up a framework for the federal government to review the national security risks of advanced AI systems for up to a month before public release. Participation was officially voluntary. The Anthropic episode suggests the voluntary part has a shorter shelf life than a lot of developers were assuming.

What it actually signals

For anyone tracking the security side of AI, the subtext matters more than the headline. A few things to sit with:

• Frontier models are being treated like munitions. Export controls, foreign-national restrictions, classified national security justifications. This is the same toolkit regulators reach for with encryption and weapons-grade software.
• The takedown window is now days. Fable 5 launched and was pulled inside the same week. That is a meaningfully different risk profile for anyone building on top of these APIs.
• Exploit-generation is the line in the sand. If a model can compress vulnerability-to-exploit time from days to hours, defenders have less room to patch, and attackers with access have considerably more.

The bigger question is what this does to the rest of the field. If Mythos 5's capabilities triggered an emergency order, every lab working on similarly capable systems is now operating under a regime where a public launch can be reversed by directive before the marketing campaign finishes rolling out.

For security teams, the practical impact is less about which AI vendor you use and more about the assumption underneath everything. The window between a new capability appearing in the wild and that capability being weaponised against your infrastructure is shrinking. The window between a fresh CVE and a working exploit is shrinking too. The people on your team who notice a suspicious email or an odd login attempt are the ones holding the line while the rest of the industry argues about export controls.