Three days is how long Claude Fable 5 had to enjoy the title of the world’s highest performing artificial intelligence model among those available to the public, before the US government ordered its immediate shutdown for all non-American users, across the planet, even within the USA and even for non-US Anthropic employees. In the statement released by the company directed by the Amodei brothers, in fact, we read: «The United States government, citing national security authorities, has issued an export control directive to suspend access to Fable 5 and Mythos 5 to any foreign national, both inside and outside the United States, including foreign employees of Anthropic».
As you will have just learned, in addition to Fable 5, Mythos 5 was also deactivated, its “big brother” which until then had remained reserved for a very few selected partners. Anthropic, the company that developed both models, said it was ready to respect the order but also made it known that it considered it an error: according to its reconstruction, the basis was an alleged and limited “jailbreak” identified in Fable 5, a flaw in the security mechanisms which, according to the company, would in any case already be present in other competing models. The story also had an enormous impact outside the United States, with several European politicians taking it as a starting point to talk about technological sovereignty and the need for the continent to develop its own artificial intelligence models. Let’s see together, in order, what happened and why.
The blocking of Mythos 5 and Fable 5 by the US government
Let’s start with Mythos: according to Anthropic it is the most advanced model ever created by the company, previewed in April but never distributed on a large scale because it is capable, according to the company itself, of identifying security vulnerabilities on its own in practically all operating systems and browsers tested. That’s why it had only been shared with about 50 organizations, including Amazon, Apple, Google, Microsoft and CrowdStrike, under a program called Project Glasswingdesigned to use these capabilities for defensive purposes, that is, to find and correct flaws before they are exploited by malicious actors.
Fable 5 arrived just three days before the shutdown and was the “tamed” version of Mythos: same basic capabilities, but with additional filters to block answers on high-risk topics, such as cybersecurity and biology. This made it, according to Anthropic, safe enough to be made available to everyone, and as soon as it was released it proved to be the best performing model among the public ones in the tests Vals AIa company that measures the performance of AI models through benchmarks, i.e. standardized tests that allow different systems to be compared with each other.
We now come to the government’s order. Formally it is an export control measure, a tool that normally serves to prevent certain technologies from ending up in the hands of foreign entities for national security reasons. In this case, however, the practical effect was much broader: to comply with the order, Anthropic had to deactivate both models for all users around the world, including American citizens, because there was no technical way to apply the block only to foreign citizens, including its own non-US employees. The other Anthropic models continue to work normally for everyone.
According to Anthropic’s reconstruction, the government did not provide detailed reasons in this regard, but the company believes that the concern concerns precisely that “jailbreak” of Fable 5. To put it simply, in the world of AI, this term indicates a way to “convince” a model to ignore its own internal rules. In this case, it would be a technique that allows the model to analyze code and identify already known software vulnerabilities, a capability that the company says is also available in other public models, including OpenAI’s GPT-5.5, and that cybersecurity experts regularly use for defensive purposes.
Anthropic also emphasizes that its most important protections do not depend only on the behavior of the model, but on independent control systems that remain active even if the model itself is “convinced” to respond abnormally.
Anthropic’s frustration
Hence Anthropic’s frustration: withdrawing from the market a model used by hundreds of millions of people due to limited vulnerability, in its opinion, risks blocking the launch of any new model by any company in the sector. There is also a basic paradox: the company has built much of its image on the caution shown with Mythos, presented as too powerful for free distribution, and today it finds itself “punished” for almost the same reason, so much so that, already in April, Sam Altman of OpenAI had spoken of a marketing strategy based on fear.
The story also had an immediate echo in Europe, where several politicians, from France to the United Kingdom, passing through the Netherlands, took the opportunity to ask for greater investments in European AI models, arguing that depending on technologies developed elsewhere exposes you to the risk of being cut off from them at any moment, just as happened to non-American users of Fable 5 and Mythos 5.
