The US government has banned Anthropic from making its latest LLMs available to non-Americans, after which the company felt compelled to withdraw Fable 5 and Mythos 5.
Under the banner of national security, the US government has imposed strict export restrictions on Claude Fable 5 and Mythos 5. These new and powerful AI models may no longer be used outside the US, or by foreigners within the US.
Geographic restrictions are not impossible to implement, but because Anthropic must also ensure that foreigners in the US no longer have access, the company feels forced to withdraw the models entirely. Fable 5 and Mythos 5 have been unavailable since this weekend, barely a few days after their formal launch.
‘National security’
Anthropic denounces the decision. The company was notified by a letter that contained no details regarding which national security concerns were at play.
The company suspects the government believes the strict restrictions surrounding the use of Fable 5 can be bypassed. The Fable 5 LLM is identical to Mythos 5 and can potentially perform dangerous tasks. Restrictions are meant to ensure Fable does not take on those tasks. Mythos 5 does not have these restrictions but is only available to curated partners as part of Project Glasswing.
Cybersecurity as the motive?
Cybersecurity is one of Mythos’s strengths, and one of the tasks Fable is not allowed to perform at the same level. The LLM can detect and combine complex and unprecedented vulnerabilities to develop cyberattacks. This is relevant for defenders who want to make their systems more secure, but dangerous when it falls into the wrong hands.
As a publicly available LLM, Fable 5 is not permitted to perform those complex cybersecurity tasks. Anthropic believes the imposed export restriction is justified by a misunderstanding regarding this. Fable 5 has, after all, discovered some vulnerabilities in software, which set off alarm bells.
However, these involve relatively simple bugs that even less advanced LLMs can easily find. Anthropic specifically refers to OpenAI’s GPT 5.5. The detection is therefore unrelated to advanced capabilities. Anthropic maintains its view that its usage restrictions are currently rigid, secure, and cannot be bypassed.
No details
At this moment, the government has shared no further details about the cause. Anthropic states that it agrees the government should have the power to block unsafe AI rollouts. However, this must then happen according to a transparent, clear, and fair process based on technical facts. According to Anthropic, the current ban does not meet that standard.
Anthropic wants to remedy the situation and is sending engineers to Washington DC in the hope of talking through the problem. Whether that will succeed remains unclear.
Poor relationship
Two side notes should be made, taking into account the lack of transparency surrounding the ban. First and foremost, Anthropic and the US government are on bad terms.
The government previously labeled Anthropic as a ‘supply chain risk’. This happened as a reprisal because the company refused to make its LLMs available to the Pentagon for the development of autonomous weapons or for mass surveillance of its own citizens.
Tensions have not subsided since then. The Pentagon and the White House did want access to the first versions of Mythos, whereby the previously issued risk label was tactically ignored. Subsequently, the Trump administration tried to further restrict access to Mythos for external parties. The new move can be seen in that light.
What about sovereignty?
Furthermore, the saga demonstrates the power of the US government and the relevance of digital sovereignty. The US is showing how it can restrict the export of the latest version of powerful software without thorough justification.
Fable 5 is much more capable than, for example, Claude Opus 5 in more innocent tasks, such as building and coding complex applications. This is relevant for organizations worldwide that are increasingly deploying AI to support programmers. Due to the actions of the US, the most suitable LLM for that task is currently no longer available.
The fact that the LLMs are now also unavailable to American users is a reaction from Anthropic to the broad US export ban. If it were up to the US government itself, Fable 5 would still be available, but only for Americans in America.
This at least demonstrates the risks of dependency on the US regarding access to the latest AI technology. In a broader interpretation, it illustrates a risk for all advanced software, which can apparently fall under heavy restrictions without a standardized process.
