US Order Forces Anthropic to Pull Claude Fable 5

Synopsis

The United States government export control directive, citing national security concerns, has led to Anthropic disabling its newest and most powerful AI models, Claude Fable 5 and Mythos 5. On June 12, the order prohibited foreign nationals from gaining access to the models. The shutdown came after a week of the researchers backlash against hidden safeguards and compulsory data retention policies linked to the same models.

AI firm Anthropic, headquartered in San Francisco, has disabled its two most powerful models shortly after their public launch has been met with controversy among AI researchers and developers due to national security concerns, the company says.

The Directive

Anthropic has been informed of US government export control authorities according to its public statement. Anthropic had to stop offering access to Claude Fable 5 and Claude Mythos 5 to any foreign national, even if they are in the United States, which is where some of Anthropic's own employees reside.

The order was given on June 12, 2026. It specifically mentioned Claude Fable 5 and Claude Mythos 5. The two models were both in the market just three days ago on June 9. Anthropic quoted national security officials in its statement.

Anthropic quoted national security officials in its statement. It suspended the entry by every foreign national, inside and outside the United States. That range encompassed Anthropic's foreign nationals employees.

The company is not equipped with a real-time system to determine who is a foreign national and who is a domestic user so decided to implement a blanket suspension. In real-time it is not possible for Anthropic to filter out foreign nationals from US users.

So that's what she did; she closed both models down for everyone, so that they would be compliant. All other Anthropic models were left intact. The rest remained up, as did Claude Opus 4.8.

Once the order was received, the reports suggest that Anthropic acted promptly, releasing a statement to the public within hours.

What the Government Flagged

The order did not specify what national security issue it addressed, but Anthropic said it learned that it was related to a “jailbreaking” method for uncovering known software vulnerabilities.

The company said it has already seen with other models available publicly and has no idea that they offer "meaningful additional capabilities to malicious actors. Anthropic has stated that it plans to provide more details on the issue.

A Troubled Launch Week

Fable 5 is the first product released in Anthropic's most capable model class, the Mythos tier, and it was a rather bumpy debut. Fable 5 is a Mythos-class model, which will be higher than Opus.

Anthropic released new classifiers to make it safe for general use. Classifiers are standalone AI systems which are used to identify potential misuse. They mark queries concerning cybersecurity, biology and chemistry, and distillation. Flagged queries are redirected to Claude Opus 4.8 instead.

Fallback occurs in less than 5% of sessions. More than 95% of Fable sessions have no fallbacks. Anthropic also outlined its plan for what it termed a “defense in depth” strategy: it used to make jailbreaks costly or narrow in scope, while still keeping data for 30 days, to spot new attacks, but not train new models from it.

This was days before the export order, when this retention policy itself was a flashpoint. All traffic on Mythos-class models will be retained for 30 days, irrespective of platform, where these models are available, including third-party cloud surfaces like AWS Bedrock and Google Vertex AI, according to Anthropic's announcement.

Critics said that this effectively excludes organisations that have to deal with strict limits on data-mining. One of the AI community commentators, using the pseudonym Lisan al Gaib, stated that the policy would be problematic for companies that need no data retention, and would hurt European companies with such constraints.

Another controversy arose regarding the model's documentation. The paragraph in the 319-page system card for Fable 5 announced that the model would be "defensively cautious" when given requests to perform frontier AI tasks, meaning that it would be made less effective without alerting the user; the callers wouldn't be denied access to the model, but it would simply be made less useful.

One of those who pointed out the clause was researcher Nathan Lambert at the Allen Institute for AI, who read the disclosure twice to get its meaning.

The policy was soon reversed after receiving strong criticism from Anthropic within a day or two. The company stated that flagged requests would no longer be silently downgraded to Opus 4.8, but be clearly downgraded instead, and API users would also be given a clear reason if they were denied.

In the process of making the change, Anthropic admitted there is a compromise in the form of easy probing and bypass of the visible safeguards, which would require the classifiers to be more aggressive, with higher rates of false positives, and that the transition was not scheduled.

According to anthropic, it was only a small portion of traffic that was originally impacted by the hidden restriction, around 0.03 percent of sessions.

Reception Despite the Controversies

Despite the controversy surrounding the safeguards, some early testers have lauded the model's raw prow. The release was a "major-version-bump-deserving step change forward," said Andrej Karpathy, a former co-founder of OpenAI and Tesla AI director who recently joined Anthropic, but also said that the steps taken to protect the model were also “set too aggressively” for launch.

Dianne Na Penn, the head of product management, research, and labs at Anthropic, stated that the new model was 10 to 20 points better than its predecessor, Opus 4.8, in terms of the frontier performance of the model.

Independent test house Hex noted it had been Fable 5 that achieved a 90 per cent score on its benchmark of complex, long-running analytical tasks, while coding platform Base44 said its first attempt to create a full application was a success.

The model beat “by a long shot” every other public model Wharton professor Ethan Mollick had ever tried, he said.

Anthropic's Position

Anthropic stated in the statement that it was “complying with the legal order” but had “disagreements with the rationale underpinning the narrow potential jailbreak,” given that a commercial model has been deployed to a “significant number of users.”

The company's press release states that it is urging any government bodies to limit the use of AI systems must do so in a clear and transparent manner, with a process established by law and based on technical evidence, and that it is endeavouring to make the affected models available again as soon as possible.

Anthropic's other models like Claude Opus 4.8 will not be impacted yet, and continue to be accessible to all users.