Anthropic has been forced to shut down Claude Fable 5 and Mythos 5 following a US government order based on national security concerns. Beyond the immediate blow to users and companies, the case has become a new chapter in the debate about how much real power advanced AI models have and how they should be regulated.

What happened to Claude Fable 5 and Mythos 5

According to Anthropic itself, the United States government issued an export control directive that requires suspending all access to Fable 5 and Mythos 5 to any foreign citizen, regardless of whether they are inside or outside the country and even if they are part of the company's staff. In practice, this forces Anthropic to completely disable both models for all its customers in order to comply with the law.

The government letter does not detail in writing the exact technical reason for the concern, although the company explains that its understanding is that the authorities believe they have identified a “jailbreak” method for Fable 5, that is, a way to bypass part of the model's security measures. Anthropic says it reviewed a demonstration of this technique and that, in that context, the system was able to identify only a small set of software vulnerabilities that were already known.

The interesting thing is that, according to the company, these same vulnerabilities can be discovered by other public models without the need for any jailbreak, which leads them to question whether the authorities' reaction is aligned with the real level of risk. Even so, the company makes it clear that it will comply with the order and withdraw access to Fable 5 and Mythos 5 for all users, even if it does not share the criteria behind it.

Why Mythos generates so much fear

The name Mythos 5 has been loaded with symbolism for a simple reason, the public conversation about advanced AI is greatly influenced by the fear of models that can enhance cyber attacks or violate critical systems. Anthropic says that the government has only verbally shown them evidence of a very limited possible jailbreak that basically consists of asking the model to read a specific code and correct software errors.

The company assures that one of the reports that it has reviewed and that it considers the basis of the government decision shows a level of capacity that is not exclusive to Fable or Mythos, since models from other providers such as GPT 5.5 from OpenAI can do something similar. Furthermore, he emphasizes that this type of use is very similar to what the defense teams themselves who keep computer systems secure already do on a daily basis.

This is where you understand some of the fear around Mythos. Although Anthropic insists that it has not received a single piece of evidence of a worrying jailbreak that has resulted in real harm, the simple fact that there is a theoretical possibility of exploiting a model to automate or escalate certain hacking tasks puts regulators on alert. The controversy is not only about what Mythos can or cannot do today, but about what it symbolizes at a time when governments and companies are trying to gauge the risks of the most powerful frontier models.

Anthropic also emphasizes that the possible jailbreaks that have been reported to it so far either generate completely benign responses or are minor findings that do not offer any specific advantage linked to Mythos over other models available on the market. From the company's perspective, elevating such a narrow case to the level of retiring a business model used by hundreds of millions of people sets a precedent that could halt any future deployment of advanced models if applied equally to everyone.

The defenses of Fable 5 and the coming debate

Anthropic's reaction is based on a key message Fable 5 is not a model launched without control but one of the systems with the most safeguards that the company has deployed. In their statement they recall that they have implemented very strict security measures to minimize the risk of uses related to cybersecurity and other sensitive areas, to the point that many users have complained that they are too restrictive.

In the weeks leading up to launch, Anthropic worked with the US government itself, the UK's AI security agency, and several independent organizations to put Fable through red team testing for thousands of hours, looking for ways to break its defenses. According to the company, these tests showed that Fable 5's safeguards are more effective than those of any previous commercially deployed model and that no one has found a universal jailbreak capable of broadly disabling system protections.

The company also admits somewhat awkwardly but realistically that perfect jailbreak resistance is probably not possible for any model vendor today. That is why they say they have adopted a defense in depth strategy, trying to ensure that any bypass that appears is either very narrow in what it allows or extremely expensive to find and exploit. Another decision that is not usually popular with customers fits into that logic: retaining user data for thirty days in order to quickly detect and close any successful attack.

Regarding the government's decision, Anthropic maintains a nuanced position. On the one hand, it publicly recognizes that governments must have the ability to block AI deployments that are truly insecure, as long as there is a clear, transparent, fair legal process based on technical criteria. On the other hand, he affirms that the specific action against Fable 5 and Mythos 5 does not comply with these principles and responds more to a misunderstanding than to a rigorous analysis of the risk.

Meanwhile, the company apologizes to its customers for the interruption and assures that it is working to restore access as soon as possible. The episode leaves several unknowns open and also an idea that should not be lost sight of: models like Claude Fable 5 are born with multiple layers of security so that they do not become an easy gateway to critical systems, but that defensive engineering coexists with a growing regulatory sensitivity that can shut down a model in a matter of hours.

More than a simple clash between “innovation or security,” the Mythos case reminds us that the conversation about advanced AI moves in gray. The challenge for companies, governments and users will be to find a balance point in which it is possible to deploy very capable models without triggering unjustified alarms, but without ignoring the real risks that also exist. In that intermediate space is where the future of systems like Claude Fable 5 will be played out.