Anthropic disables its most powerful AI models worldwide after US government order

Anthropic has disabled its two most advanced AI models, Fable 5 and Mythos 5, for all users worldwide.

The AI giant shut down the models after the US government issued an export control directive citing national security authorities.

The company said the directive, received at 5:21pm ET on Friday, ordered it to suspend all access to both models by any foreign national, whether inside or outside the United States, including Anthropic’s own foreign national employees.

In a statement on its website, Anthropic said the order is so broad that the only way to comply is to disable both models entirely for every customer, regardless of nationality. Access to all other Anthropic models remains unaffected.

What Anthropic is saying

Anthropic said it reviewed a demonstration of the technique the government cited and found it could only find a small number of previously known, minor vulnerabilities.

It said these issues are simple enough that other publicly available models, including OpenAI’s GPT-5.5, can identify them without requiring any bypass, and which security defenders use every day to find and fix software flaws.

• “These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass,” the company said

The company also insisted that has implemented robust safeguards designed to significantly limit the risk of Fable being misused

• “We have instituted strong safeguards that greatly reduce the likelihood that Fable is misused for tasks related to cybersecurity (among others). In fact, our safeguards are so strong that many users have complained that they are overly broad,” it wrote.

Anthropic said it disagrees that the discovery of a narrow, non-universal jailbreak should be grounds for recalling a commercial model already deployed to hundreds of millions of users.

The company warned that if this standard were applied across the industry, it would effectively halt new model deployments for every frontier AI developer.

The company said it supports the government having the ability to block unsafe AI deployments, but only as part of a statutory process that is transparent, fair, clear, and grounded in technical facts, and said this directive does not meet that standard. The company also said it is working to return online.

• “We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible,” it assured customers.

More insights

Anthropic said that in the weeks leading up to Fable 5’s launch, it worked with the US government, the UK’s AI Security Institute, multiple private third-party organisations, and internal teams to red-team the model’s safeguards for thousands of hours.

Those tests, the company said, showed Fable 5’s protections were substantially more effective than any previously deployed model, and no tester found a “universal jailbreak” capable of broadly unlocking the model’s restricted cyber capabilities.

• Anthropic acknowledged that perfect jailbreak resistance is not currently possible for any AI model provider, noting that every safeguard used across the industry remains vulnerable to narrower, non-universal jailbreaks, and that universal jailbreaks will likely eventually be found for any model, a point the company said it stated clearly at Fable 5’s release.
• Given this, Anthropic said it adopted a “defense in depth” strategy for Fable 5, designing jailbreaks to be either narrow or prohibitively expensive to produce, paired with monitoring to quickly detect and shut down successful attacks. As part of this approach, the company requires 30-day retention of customer data on Fable, a policy it said carries real costs with customers but allows it to research and mitigate jailbreak attempts.
• On the specific case behind the directive, Anthropic said the government has so far provided only verbal evidence of one potential narrow jailbreak, which essentially involves asking the model to read a specific codebase and identify software flaws.

What you should know

Nairametrics had earlier reported that Anthropic called on developers of advanced AI systems to establish a coordinated and verifiable framework for slowing or temporarily pausing AI development if future models begin improving themselves faster than society can effectively manage the associated risks.

The company made the appeal in a detailed post published on its website, warning that increasingly capable AI systems could eventually reach a point where humans lose meaningful oversight and control.

Anthropic argued that the industry should agree in advance on clear conditions and mechanisms for halting development if such risks begin to materialize.