Anthropic Suspends Claude Fable 5 and Mythos 5 Worldwide

Anthropic has suspended worldwide access, for all of its customers, to its two most advanced AI models: Claude Fable 5 and Claude Mythos 5. This is not a commercial decision. It follows an emergency directive from the US government, issued in the name of national security, that bans the use of these models by any "foreign national".

Only a few days ago, we welcomed the release of these two models, which outpaced their competitors on most published benchmarks. The reversal is brutal. Here is what happened, why, and the question it all leaves open, explained plainly.

One reassuring note up front: the other models from Anthropic, namely Claude Opus 4.8, Sonnet, and Haiku, are not affected. They remain fully available. Only Fable 5 and Mythos 5 have been switched off.

---

What happened, in plain terms

The US administration issued an emergency export-control directive. This kind of text usually governs the sale of sensitive technology abroad. Here, it targets two AI models directly, treating them as strategic goods.

The order bans the use of Fable 5 and Mythos 5 by any "foreign national", meaning anyone who is not American. The rule is so broad that it applies even to Anthropic's own engineers based outside the United States.

Yet a company cannot verify in real time the nationality of each of its users around the world. Lacking an immediate technical way to filter access, Anthropic chose the only legally compliant option: fully disabling both models, for everyone, everywhere.

💡 Rather than risk heavy criminal penalties, Anthropic chose to cut off all access to its two flagship models. A radical decision, imposed by a text it did not write.

The company has apologized to its customers and says it is actively seeking a regulatory path to restore access as quickly as possible. As we write, no return date is confirmed.

Why: a jailbreak reported to the government

The immediate trigger has a name: a "jailbreak". A jailbreak is a method for bypassing an AI's safety barriers to make it produce what it would normally refuse. A third-party report, forwarded to the US government, demonstrated one on Fable 5.

The technique is narrow. According to Anthropic, it essentially consists of asking the model to read a codebase and fix its software flaws. Nothing more spectacular at this stage.

And this is exactly where Anthropic disputes the decision. The company calls it a "minor and non-universal" jailbreak, notes that it has received no report of a case leading to a genuinely harmful result, and points out that the same capability is available in competing models such as OpenAI's GPT 5.5, without those being recalled. In its view, this is a misunderstanding.

A broader backdrop remains. In a recent essay on AI policy, Anthropic co-founder Dario Amodei argues that Mythos-class models have become tools of strategic consequence, particularly in cybersecurity. He also raises, in the conditional and for the future, the possibility that AI could one day help design biological weapons. These are forward-looking warnings about the technology's trajectory, not a description of anything Fable 5 actually did.

---

Defense in depth, explained simply

To protect its models, Anthropic applies a principle borrowed from cybersecurity: defense in depth. The idea is to stack several independent layers of security, so that if one is bypassed, the next ones take over.

Here are the four layers that make up this protection.

Security layer	Its role, in plain terms
Constitutional AI	The deepest layer, Anthropic's signature. From training onward, the model learns to follow a written charter, inspired notably by the Universal Declaration of Human Rights. A second model critiques the first continuously to correct dangerous answers.
Input and output filters	A software firewall. The input filter inspects the question before it reaches the model; the output filter inspects the answer before display and screens out malicious code.
Behavioral monitoring	The system analyzes not an isolated sentence but the full conversation history, to spot a gradual manipulation of the user.
Sandboxing	The AI's code-execution abilities are isolated from the internet and sensitive servers, to prevent any escape into the infrastructure.

The jailbreak reported on Fable 5 is what's called context injection: a sensitive intent is buried inside a technical task that looks harmless, such as reviewing code. Anthropic stresses one point, however: in the reported case, the technique stayed narrow, non-universal, and led to no documented harmful result. This is far from a general collapse of security.

An already tense political backdrop

This shutdown does not happen out of a clear sky. For months, relations between the Trump administration and Anthropic have been marked by mutual distrust.

Back in March, the Pentagon had labeled Anthropic a "supply-chain risk". The tag is heavy: it is usually reserved for adversarial foreign powers. It followed the breakdown of negotiations on the military use of AI.

The export-control directive fits this hard line. By banning the models' use by "foreign nationals", including the company's non-American engineers, it left Anthropic with no way out: cut access, or face prosecution.

One last element intrigues observers: the timing. This sudden brake comes just as Anthropic has filed its confidential IPO paperwork, with a valuation approaching one trillion dollars.

---

Marketing stunt or lasting reality?

This is the question many are asking, and we put it openly rather than answer it for you. Because two readings clash.

The first sees a communications operation. A model a government deems worth suspending in a hurry is also, mechanically, a model presented as powerful. The timing, days before a colossal stock-market listing, feeds the suspicion. Nothing sells a technology better than the idea that a government fears it.

The second reading takes the decision seriously. In his essay on AI policy, Dario Amodei argues that models of this generation have become matters of strategic security. If so, strictly governing their distribution is no longer a publicity stunt but a necessity, and the suspension could settle in for the long haul.

The truth probably sits between the two, and that is exactly what makes the episode interesting. It says something about our era: we are building tools whose power already tests our ability to govern them calmly.

At AB-Arts, we track these tremors because they shape the frame in which we work and teach. Understanding how an AI is secured, and what a security alert really means, is now part of basic professional literacy. That is precisely what we pass on in our Claude masterclass.

→ To learn how to work with these models with open eyes, explore our Academy training. For a conversation about what this changes in your organization, reach out via our contact page.