US Export Ban Forces Anthropic to Take Fable 5 and Mythos 5 Offline Worldwide

The US-based AI developer Anthropic is cutting off access to its most advanced artificial intelligence models for all users worldwide for the time being, following a government order. Since it was not possible for Anthropic to block only users from outside the US, the only practical alternative was to shut the systems down. "We are operating under the assumption that this is a misunderstanding and are working to restore access as quickly as possible," the company said, apologizing to its users.

According to US media reports, the trigger was a report that in-house cybersecurity researchers from Amazon had successfully "jailbroken" the Fable 5 model using a targeted sequence of prompts. Andrew Jassy, the head of the Amazon group, had described this to US government officials, including Treasury Secretary Scott Bessent. The model had subsequently spit out sensitive information that could be used directly for large-scale cyberattacks.

Background on the Order from Washington

Anthropic itself rejected this account in a blog post. As far as it knew, the government had only received information about a "potential" – but not a verified, critical – jailbreak. Moreover, the supposedly uncovered security vulnerabilities had been "minimal." The company stressed that it had only received parts of the information from the government and spoke of a misunderstanding.

Fable 5 is a variant of the Anthropic model Mythos, which detects dangerous security vulnerabilities in nearly any software in the shortest possible time. The artificial intelligence behind Anthropic's AI model "Mythos" is particularly good at uncovering software vulnerabilities, some of which have remained undetected for decades. This capability has previously been used by US agencies and selected companies to find security gaps.

The less powerful Fable 5 had only become available in Anthropic's Pro tier a few days before the incident. There had been concerns at the White House about a possible jailbreak – that is, a trick that helps unleash the dangerous capabilities of Fable 5 by circumventing its built-in safeguards. Among other things, Fable 5 was only supposed to be included in the subscription until June 22, after which the model was to be available only via API and with token-based billing – and thus significantly more expensive.

How Fable 5 Works and Its Limitations

Fable 5 refuses to answer prompts in the areas of biology, chemistry, cybersecurity, and knowledge distillation. Fable 5 regularly switched itself in the background to the next weaker Anthropic model, Claude Opus 4.8. Without this being visible to users. Anthropic cites this as a protective measure against knowledge distillation – that is, training one AI model with another. Fable 5 is supposed to independently detect when it is being used for such training and then switch to Opus 4.8.

You pay for Fable 5, but in the end you get an uncontrollable mix of models. Even though Anthropic promised to bring Fable 5 back into the subscription at some point. In Anthropic's benchmark comparison, Fable 5 is significantly stronger than, for example, Opus 4.8 or ChatGPT 5.5. Anthropic cites a lack of capacity as the reason, which first has to be built up through new data centers.

Behind the scenes, Amazon CEO Andrew Jassy is said to have played a central role. According to consistent reporting by US outlets such as The Wall Street Journal, Reuters, and The Information, Jassy deliberately sought out conversations with senior government officials, including US Treasury Secretary Scott Bessent. Amazon is itself one of the largest investors in Anthropic and is closely intertwined with the AI company through its cloud division AWS. An Amazon spokesperson stated: "As a leading cloud provider serving a large number of customers in the private and public sector, it is not unusual for governments to seek our input on potential security risks. When this occurs, we do not share the details of those discussions."

The Role of Amazon and David Sacks

Tech billionaire David Sacks is also likely to have played a role. David Sacks, the former "AI Czar" of the Trump administration and co-chair of the president's Council of Advisors on Science and Technology, confirmed to TechCrunch that a "highly credible, trusted partner to both Anthropic and the US government" had approached the administration with a working jailbreak. According to Sacks, the administration ultimately demanded that Anthropic CEO Dario Amodei either fix the jailbreak or take the model offline for the time being. "Dario refused" – after which the government "reluctantly" pulled the emergency brake and imposed a strict export ban.

Sacks expressed the administration's hope that Anthropic would fix the security issue, the export controls would be lifted, and Fable could return to the general release. The Trump administration's official justification for the export ban sounded drastic: Anthropic's new models represented an unacceptable risk to national security, as they could identify critical software vulnerabilities and make them usable for cyberattacks.

According to Anthropic's account, however, there is no known way to get the model to do something that could not also be done with other models, such as ChatGPT 5.5. Anthropic also stressed that the security precautions in Fable 5 had been extensively tested. The company rejected blocking software for hundreds of millions of users on these grounds. The government's current measures do not, however, correspond to the principles of fair and fact-based regulation, the company now stated.

Anthropic's Positions and Reactions

At the same time, just a few days earlier, Anthropic chief Dario Amodei had called for stronger oversight of AI, including the ability to block models with unacceptable risks. As recently as Wednesday, Anthropic itself had called for stronger oversight of AI, including the ability to block models with unacceptable risks. The company is also suing the Pentagon, which had classified Anthropic as a supply chain risk – a classification that can severely restrict the use of the software in government agencies.

In Europe, meanwhile, concern was growing about becoming too dependent on the US for AI infrastructure. Austrian Green Party digital policy politician Süleyman Zorba said on Saturday: "Anyone who places their digital infrastructure entirely in foreign hands makes themselves dependent and vulnerable to blackmail." Clearly, model families like Mistral cannot currently keep up with the products of OpenAI or Anthropic. But: for individual steps in a workflow, they could certainly be an alternative.

IT security expert Dennis-Kenji Kipker stated that "Mythos" could find a large number of IT security vulnerabilities. This could also be exploited by criminal actors. The AI manufacturer prohibits the use of its models for autonomous weapons systems of the US military; sanctions were imposed for this. In recent months, Anthropic had repeatedly come into conflict with the US government because the company insisted that its AI models must not be used in autonomous weapons systems and for mass surveillance in the US.

Consequences for the Global AI Market

Observers interpret the incident as a possible prelude to greater fragmentation of the global AI market. Washington is increasingly treating modern AI systems as strategic technology whose dissemination should be controlled as strictly as other security-relevant high technologies, according to an analysis. Until now, US export restrictions had mainly focused on high-performance AI chips and data centers. If the US government continues on this course, other particularly powerful AI systems could in the future be available internationally only to a limited extent or not at all.

The company announced that it would restore access as soon as possible. To meet this requirement, there was no alternative but to take the systems completely offline. In the official statement, Anthropic said: "The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance." This is the first case in which AI models themselves have become the subject of US government control measures.

Anthropic had staged Mythos for weeks as "too dangerous for the public" in order to generate exclusivity. Mythos 5 is the non-public full version, which is to continue to be used only by government agencies and selected corporate partners to secure their systems. The Pentagon had meanwhile classified Anthropic as a supply chain risk. Anthropic is suing over this classification. Under the hood of an agentic AI application – especially when it is built entirely or partly in-house – there is often a multitude of models.