US order: Anthropic bans Fable & Mythos 5 Apfelpatient

Anthropic has abruptly cut off access to its two most powerful AI models for all customers – just days after their launch. The trigger is an export control order from the US government, citing national security interests. The company is complying with the order but is unusually vehemently contradicting its reasoning.

Just at the beginning of the week, Anthropic unveiled its most powerful generation to date, the Claude Fable 5, the first publicly available model in the Mythos class. A few days later, this opening has completely disappeared: since the evening of June 12th, both the Fable 5 and the more powerful Mythos 5 have been deactivated for all customers. This is due to a formal order from Washington prohibiting access for foreign nationals – with consequences that extend far beyond this group, as Anthropic was subsequently forced to disable the models for all users.

What the US government has ordered

The ball started rolling with a report from Axios: Commerce Secretary Howard Lutnick stated in a letter to Anthropic CEO Dario Amodei that the Mythos 5 and Fable 5 are subject to export controls – for any location outside the US as well as for any foreign national within the country. According to the order, a license would therefore be required for the export, re-export, or domestic distribution of the affected models.

Anthropic itself describes the scope even more precisely: The order suspends all access by foreign nationals to both models – whether within or outside the United States, explicitly including foreign Anthropic employees. In practice, this requirement could only be implemented through a complete shutdown. The directive was received that same day at 5:21 p.m. local time (ET); the document did not contain specific details regarding the aforementioned security concerns. Access to all other Anthropic models remains unaffected by the measure.

The trigger: a tight jailbreak

The Axios report also reveals what alarmed the agency. According to the report, an administration official stated that the Commerce Department acted after another company claimed to be able to bypass - or "jailbreak" - Mythos. The Trump administration had previously attempted to prevent the model's release but had failed.

Anthropic's own interpretation aligns with this picture, but is far less dramatic in detail. According to the company, the government believes it has learned of a method to circumvent Fable 5's security mechanisms. Anthropic claims to have examined a demonstration of this technique and identified only a small number of already known, minor vulnerabilities. The crucial point: these vulnerabilities are comparatively simple, and other publicly available models could also find them – without any circumvention. Specifically, the alleged jailbreak essentially consists of allowing the model to read a specific codebase and fix the software errors it contains. This level of capability is widely available, for example via OpenAI's GPT-5.5, and is used daily by defenders securing systems.

Why Anthropic opposes the measure

Anthropic insists it is complying with the legal order and removing the access – but considers the reasoning flawed. The company points to the safeguards it implemented in Fable 5: In the weeks leading up to its launch, the US government, the UK's AI Security Institute, several external organizations, and internal teams subjected the security measures to thousands of hours of rigorous testing. The result, they claim, is that Fable's protection is significantly more effective than that of any previously released model. They assert that no one has yet found a universal jailbreak – a method that unlocks a wide range of dangerous capabilities at once.

The company's defense rests precisely on this distinction. Anthropic describes itself as pursuing a "defense in depth" strategy: jailbreaks should either be severely restricted or very expensive to produce, combined with close monitoring to quickly detect and stop successful attacks. According to Anthropic, the controversial 30-day storage of customer data for Fable also serves this purpose – it makes it possible to research and mitigate jailbreaks in the first place. The company considers it disproportionate to recall a commercial model already available to hundreds of millions of people because of a single, potentially limited jailbreak. If this standard were applied industry-wide, Anthropic argues it would amount to a de facto halt to all new model releases.

What remains of the model portfolio

Only the Fable 5 and Mythos 5 are affected. All other models from the company – including the recently introduced Claude Opus 4.8 – will continue unchanged. This measure specifically targets the newest and most powerful generation, while leaving the broader commercial product range unaffected.

Both blocked models originate from Claude Mythos Preview, a highly sophisticated system that Anthropic originally designed for security research and aimed at detecting software vulnerabilities. Access was initially restricted to a small circle of companies and research partners through the program, the expansion of which Anthropic had previously announced. Participants reported finding and closing numerous security gaps using the model during this phase – evidence that the same technology now classified as a risk also serves a defensive purpose.

AI models are becoming a national security issue

Beyond this specific case, the order marks an escalation: Washington is increasingly treating state-of-the-art AI systems as security-relevant goods, comparable to traditional export technology. According to Axios, Anthropic is already on a Pentagon list that deems the company too risky for government use – and is now also subject to a licensing regime by the Department of Commerce. The fact that the government is effectively blocking foreign access to a commercial model by directive is unprecedented.

Anthropic argues that a government should indeed be able to block unsafe publications – but only within the framework of a legally regulated process that is transparent, fair, clear, and based on technical facts. The company believes that the current procedure fails to meet precisely this standard. The dispute thus touches upon a fundamental question: According to which verifiable criteria may authorities withdraw an already released AI model from the market?

A precedent with an open outcome

Anthropic has apologized to customers for the disruption, calling it a misunderstanding and promising to restore access as quickly as possible. The company announced further details will be released within the next 24 hours. How quickly the conflict can be resolved now depends on whether the government can substantiate its concerns with solid technical evidence – and whether both sides can reach a shared understanding of what a tight jailbreak actually entails. (Image: Shutterstock / JRdes)