US Export Ban on Anthropic's AI Models Fable 5 and Mythos 5 – Global Shutdown as a Consequence

According to US media such as Reuters, the ban affects not only users outside the United States, but explicitly also foreign nationals located in the US – including those employed at Anthropic itself. The Department of Commerce had, on national security grounds, prohibited granting foreign nationals – including those in the US – access to the new models "Fable 5" and "Mythos 5." This takes the control of modern AI systems to a new dimension, going beyond the previously common regulation of AI chips and data centers.

Anthropic's Position: Limited Capability, Not Exclusivity

According to the company, the background is a report that, in Anthropic's assessment, is likely to have triggered the order. Anthropic's experts had concluded that this was a limited capability to have the AI check specific program code and correct errors. The artificial intelligence behind Anthropic's AI model "Mythos" is particularly good at detecting software vulnerabilities, some of which have remained undiscovered for decades. This capability has previously been used by authorities and selected companies to close security gaps.

Anthropic emphasizes that the safety precautions in Fable 5 have been extensively tested. According to Anthropic's account, however, there is no known way to make the model do anything that could not also be achieved with other models, such as ChatGPT 5.5. Competing models such as rival OpenAI's GPT-5.5 also possess comparable capabilities for code review, according to the company. The government's current measures, however, do not correspond to the principles of fair and fact-based regulation, the company stated.

Fable 5 is a less powerful variant of the Mythos model, capable of detecting dangerous security vulnerabilities in nearly any software in a very short time. The model had become available a few days earlier in Anthropic's Pro tier and was originally intended to be included in the subscription only until 22 June, before being made available only via API and with token-based billing – and thus significantly more expensive. Mythos 5, on the other hand, is the non-public full version, which was to continue to be used only by authorities and selected corporate partners to harden their systems. Anthropic promised, however, to eventually bring Fable 5 back into the subscription.

Comparison with OpenAI and Heise Tests

In Anthropic's benchmark comparison, Fable 5 is significantly stronger than, for example, Opus 4.8 or ChatGPT 5.5. It was notable that Fable 5 regularly switched in the background to the next weaker Anthropic model, Claude Opus 4.8, without this being apparent to users. This is said to happen quite frequently, as colleagues from Heise's video magazine c't 3003 have learned. Anthropic cites a protective measure against knowledge distillation – that is, training one AI model with another – as the reason. Fable 5 is said to independently detect when it is being used for such training and then switch to Opus 4.8. Fable 5 also refuses to answer prompts in the areas of biology, chemistry, cybersecurity, and knowledge distillation. Anthropic has, admittedly, recently announced it would become more transparent in this regard.

There had been concerns at the White House about a possible jailbreak. According to the company, the government believes it has gained knowledge of a method to lift restrictions in the software. The government did not elaborate on its security concerns, Anthropic explained. The company has so far received only partial information from the government and has not obtained concrete technical proof or detailed information about the alleged security vulnerability. Anthropic disagrees with blocking software for hundreds of millions of users on these grounds. To meet this requirement, there is no alternative but to take the systems completely offline.

Export Controls on AI Models: A Novelty

Concern that such an AI could also be used as a cyber weapon has accompanied the Mythos models since their conception. "Mythos" could find a large number of IT security vulnerabilities, says IT security expert Dennis-Kenji Kipker. This could also be exploited by criminal actors. The capabilities of Mythos 5 had also raised concerns among authorities that such models could be used to prepare cyberattacks more quickly and effectively or to exploit security vulnerabilities in an automated manner. That these concerns are now resulting in a formal export control is new: Until now, US export restrictions have focused primarily on powerful AI chips and data centers. Washington is increasingly treating modern AI systems as a strategic technology whose proliferation is to be monitored as strictly as other security-relevant high technologies. The incident is, in observers' assessment, the first situation in which AI models themselves have become the focus of state control under US export control law.

Background: Dispute with the Pentagon

The conflict between Anthropic and the government has a longer history. The Pentagon declared Anthropic a supply chain risk, which can severely hamper the use of the company's software in government agencies. The trigger was that Anthropic, despite pressure, insisted that its AI models not be used in autonomous weapons systems or for mass surveillance in the US. The AI manufacturer prohibits the use of its models for autonomous military weapons systems, and sanctions were imposed for this. Anthropic is suing against the classification as a supply chain risk. The current step is to be regarded as a further episode in its already long-running dispute with the government.

Only a few days before the current order, Anthropic CEO Dario Amodei had publicly advocated giving the government the ability to block potentially dangerous AI software. As recently as Wednesday, Anthropic itself had called for stronger oversight of AI, including the ability to block models with unacceptable risks. Anthropic emphasizes, however, that this must be done on the basis of transparent and clear procedures as well as technical facts. That is not the case now.

Consequences for Europe and the AI Market

The incident is reminiscent of discussions about digital sovereignty in Europe. Süleyman Zorba, the Austrian Greens' network policy spokesperson, addressed the topic on Saturday: "Wer seine digitale Infrastruktur vollständig in fremde Hände legt, macht sich abhängig und erpressbar". Many German IT decision-makers are currently at the beginning of projects intended to bring AI agents and the corresponding infrastructure into their organizations. Commentators also see in the shutdown a signal for the European AI industry. Model families such as Mistral cannot currently keep up with the products of OpenAI or Anthropic.

Anthropic had originally withheld Mythos from a broader audience because, in the company's assessment, the model was too good at exploiting software vulnerabilities. The release of a stripped-down variant as Fable 5 was intended to address these concerns. It is now becoming apparent that even the stripped-down version can become the object of state export controls. We at Heise have also tested this successfully – namely, circumventing the security precautions of other AI models. Experience shows that purely software-based blocks remain vulnerable.

Meanwhile, market observers point to the economic context: OpenAI and Anthropic want to go public. One must understand that stock valuations are not only about the current performance of the companies, AI expert Kevin Bauer recently said on ZDFheute live. The current conflict with Washington could affect the planned IPOs. Manageable damage with a large signal, finds Niklas Engelking – the immediate impact on ongoing operations is limited, but the political reach of the event is enormous. For the European debate on digital sovereignty, the case provides new fuel.