Claude Fable 5 was only available for three days. Anthropic launched it on June 9. On June 12, the United States government ordered it shut down — the first time a government has disconnected an AI model — and Anthropic, unable to filter users by nationality, disabled it for the entire planet.
The most talked-about model of the year — records on SWE-bench, cutting-edge vision capabilities — went from triumphant launch to global block in 72 hours. Here is what actually happened, what is known, and what has not been said.
What the government ordered
The directive came through the Bureau of Industry and Security (BIS) of the Department of Commerce, under the direction of Secretary Howard Lutnick. It was not a general shutdown order, as many headlines have mischaracterized it. It was an export control directive that conditioned access by foreign persons to the models.
The problem is that Anthropic did not have — and still does not have — a system for determining the nationality of each user in real time through its API. Without that capability, the only way to comply with the order was to block access for everyone.
According to Anthropic’s official statement, the directive ordered “the suspension of all access to Fable 5 and Mythos 5 by any foreign person, inside or outside the United States, including foreign employees of Anthropic.”
Anthropic complied immediately. But it made its public disagreement clear:
“We are complying with the government’s legal directive and are withdrawing access to Fable 5 and Mythos 5 for all users. However, we disagree that the finding of a potential and limited jailbreak should be grounds for withdrawing a commercial model deployed for hundreds of millions of people.”
The jailbreak in dispute
The official reason cited by the government is the existence of a jailbreak: a method for circumventing Fable 5’s safety guardrails. The government considers this a threat to national security.
Anthropic, which had made its own safety warnings about Fable 5 public — including classifiers for cybersecurity, biology, chemistry, and model distillation — maintains a very different account. According to the company, the jailbreak in question is “narrow and not universal”: it consists of asking the model to read a specific codebase and find software vulnerabilities.
The company argues that the same capability is “widely available in other models (including OpenAI’s GPT-5.5)” and is used daily by defensive cybersecurity professionals.
Anthropic also noted that it conducted “thousands of hours” of red-teaming with the US government, the UK AISI, and third parties before launch, and that “no evaluator has been able to find a universal jailbreak” in Fable 5.
“We have not even received disclosure of a potential non-universal jailbreak that has led to a harmful outcome,” the company stated.
The government, for its part, has not published the technical report on which it bases its decision. The directive itself remains a non-public document.
The legal gap
One of the most concerning aspects of the case is the lack of legal clarity. The government has not publicly named the specific statutory authority under which it issued the directive. The two possible avenues are ECRA/EAR (Export Control Reform Act / Export Administration Regulations) or IEEPA (International Emergency Economic Powers Act).
The legal analysis published by Penwell Law identifies several structural problems:
Classification vacuum. The only ECCN that classified AI model weights — ECCN 4E091, part of the Biden-era AI diffusion framework — was rescinded on May 13, 2025, one month before Fable 5’s launch. At the time of the directive, no existing regulatory category explicitly classified AI model weights.
Published software doctrine. Under 15 C.F.R. § 734.7(a), software published on the internet without restrictions is “published software” and is not subject to the EAR. Fable 5 was a paid API — which creates a gray area — but Mythos 5 was restricted to selected partners, further complicating the analysis.
Deemed export rule. The “deemed export rule” (15 C.F.R. § 734.13(a)(2)) covers technology or source code, but not object code. Model weights are closer to object code than to source code.
Severely limited judicial review. Under 50 U.S.C. § 4821(a), export control actions are excluded from judicial review under the Administrative Procedure Act (APA). The surviving avenue for review — non-statutory ultra vires review — is extremely narrow.
As the analysis notes: “A directive this consequential, without a publicly stated legal basis, is itself part of the problem.”
Only two models?
Yes. All other Anthropic models — Claude Opus 4.8, Sonnet 4.5, Haiku 3.5, and the rest of the family — were not affected by the directive. The government’s action was concentrated exclusively on Fable 5 and Mythos 5, the two models based on the Mythos-class architecture.
This is relevant because it suggests the government’s concern is not general, but specific to the capabilities of these particular models. Or, alternatively, that the action was fast enough that it did not extend to other models — perhaps because the reported jailbreak only worked against Fable 5.
Will the models return?
David Sacks, co-chair of the PCAST (President’s Council of Advisors on Science and Technology), indicated that the models could return once Anthropic patches the identified vulnerability. “The ball is in Anthropic’s court,” he stated, suggesting that the lifting of the restriction depends on the company.
There is no public timeline. Nor is it clear whether the “patch” demanded by the government is the same one Anthropic considers necessary — the company has said the reported jailbreak is minor and that equivalent capabilities are available in other models.
What is not known
Several key questions remain unanswered:
What is the real motivation? A follow-up article by TechCrunch published on June 15, titled “The US government’s Anthropic models ban was never about an AI jailbreak,” suggests there may be broader motives — political, regulatory, or competitive — behind the action. So far, no official source has confirmed them.
Who triggered the process? Unverified claims are circulating on YouTube and technical forums about a call between Amazon CEO Andy Jassy and Treasury Secretary Scott Bessent that allegedly triggered the directive. Amazon is an Anthropic investor and its primary cloud provider. Reuters, CNBC, The Guardian, and WIRED have not confirmed this information.
How much notice did Anthropic receive? Unconfirmed sources claim the company received approximately 90 minutes between notification and the compliance deadline. This is legally possible under a Temporary Denial Order (15 C.F.R. § 766.24, which can be issued without prior notice), but has not been confirmed by Anthropic or the government.
Was this used against Anthropic because of its political stance? Another unverified theory suggests the action was retaliation for Anthropic’s refusal to participate in equity stake conversations with the White House or for not funding a political PAC. There is no public evidence supporting this claim.
The precedent
Beyond the details of the case, what makes this event historic is the precedent it sets: for the first time, a US government agency has used export control machinery to effectively withdraw a commercial AI model from the market after deployment, without prior judicial order and without a clear legal framework.
Every frontier model developer in the US — OpenAI, Google DeepMind, Meta, Anthropic — now has a new data point to incorporate into its risk assessment: a model that complies with all existing regulations can be globally disconnected in a matter of hours by a directive whose legal basis is not made public.
As one legal analyst wrote: “This is not a technical debate about jailbreaks. This is about the question of who has the power to decide what AI capabilities can exist, for whom, and under what rules.”
And for now, the answer is: the US government, without anyone knowing exactly under what authority.
Primary sources: Anthropic Official Statement | Reuters | CNBC | TechCrunch | Legal Analysis — Penwell Law