Anthropic cut off access to its Fable 5 and Mythos 5 models Friday night, days after launch, after receiving a US Commerce Department directive Friday evening subjecting both models to export controls that bar their use outside the United States. The models had reached hundreds of millions of users before the shutdown. Anthropic said full compliance in the immediate term required disabling both models entirely, for every customer, including those inside the US.
Access to other Anthropic models is unaffected. The company framed the move as a temporary measure tied to a single Commerce directive rather than a broader pull-back, and apologized to customers for what it called a disruption resulting from a misunderstanding. Anthropic said it will release more details in the next 24 hours.
The trigger, according to an administration official cited in reporting on the order, is a jailbreak that reportedly bypasses Fable 5's classifier-based safeguards for cybersecurity, chemistry, and biology prompts. The administration asked for a pause so the national security apparatus could be hardened against the threat, a process that one source said could finish in the next few weeks.
“We are complying with the government's legal directive and are removing access to Fable 5 and Mythos 5 for all users.”— Anthropic, Company statement
Key facts
- 01Anthropic shut down Fable 5 and Mythos 5 Friday night after a US Commerce Department directive issued Friday evening.
- 02The models had been deployed to hundreds of millions of users and were live for only days before the pull.
- 03The directive cites a narrow jailbreak involving codebase review for software flaws; Anthropic says GPT-5.5 has similar capabilities.
- 04The Commerce Department wants 'national security apparatus' hardened, possibly within the next few weeks.
- 05Access to Anthropic's other models is unaffected; the company promised more details within 24 hours.
Anthropic's own description of the evidence is narrower. The company said the government provided only verbal evidence of a potential narrow, non-universal jailbreak involving Fable 5 reviewing a specific codebase for software flaws. The flaws surfaced so far have been minor and relatively simple, Anthropic said, and the company noted that other publicly available models including GPT-5.5 have comparable capabilities on the same task.
That gap, between the administration's framing of a national security risk and Anthropic's framing of a routine code-review capability shared across frontier models, is the substantive dispute. Export controls applied to a deployed consumer model are an unusual instrument, and the precedent matters more than the specific finding. If a verbal report of a narrow jailbreak is enough to pull a live model used by hundreds of millions of people, the same standard would apply to every competing system.
Anthropic made that argument explicit in its Friday night post, complying with the directive while putting on the record that it disagrees with the threshold the government applied. The statement is unusually direct for a frontier lab responding to a federal order, and it draws a line that other model providers will likely cite in future disputes.
“However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”— Anthropic, Company statement
The broader regulatory backdrop is in motion. Earlier this month, President Trump signed an executive order urging AI model makers to submit to voluntary government security testing. The signing followed an initial ceremony last month that was postponed amid reported disagreements inside the administration about the scope of the program. Friday's directive is the first concrete enforcement action to land on a deployed frontier model since that order.
The mechanics of the Commerce action also raise questions Anthropic has not yet answered publicly. Export controls typically restrict cross-border movement of regulated technology, but a cloud-hosted model accessible by API does not move in the same sense. Disabling the models entirely was the company's chosen path to immediate compliance, but it is not obvious that a geofenced US-only deployment would have been insufficient, and Anthropic has not said whether that option was on the table.
There is a counter-case for the government's caution. If even a narrow jailbreak can be tuned by a determined actor into something broader, a brief pause to stress-test classifiers before re-enabling access is a small cost against a real downside. Anthropic itself ships safety frameworks that contemplate model recalls in extremis. The disagreement is over whether the evidence presented clears that bar, not whether such a bar should exist.
For the AI market, the immediate read is that frontier model deployments now carry a new category of operational risk: a verbal federal directive can take a live product offline overnight, with hundreds of millions of users on the other side of the switch. That changes how labs price uptime guarantees, how enterprise customers structure model dependencies, and how competing providers like OpenAI and Google think about the timing and scope of their own launches. Anthropic's public dissent, lodged while complying, is the more consequential move — it sets up the next dispute, on the next model, with the threshold itself in question.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
