A Friday letter from Commerce Secretary Howard Lutnick classified Anthropic's Fable 5 and Mythos 5 systems as export controlled after a jailbreak Amazon reportedly found, where a jailbreak is a prompt that bypasses a model's safety guardrails.
A US Commerce Department letter, delivered to Anthropic at 5:21 p.m. Eastern on a Friday, classified the company's two most capable AI systems, Fable 5 and Mythos 5, as subject to US export controls and ordered them shut off to "foreign nationals" worldwide, including Anthropic's own foreign-national staff.
The directive came from Commerce Secretary Howard Lutnick and was triggered by what officials described as a narrow jailbreak identified by Amazon. A jailbreak, in the AI sense, is a technique for bypassing the safety guardrails built into a model. Anthropic, which says it has no way to verify the citizenship of every user logging in, treated the order as a total shutdown: the company abruptly disabled both models for all customers. Other Anthropic products, including its consumer assistant Claude, were not affected.
The story was first assembled and contextualized by Zvi Mowshowitz in his Substack post American Government Takes Down Claude Fable. Anthropic's public statement, quoted in that post and confirmed in a post from @AnthropicAI, frames the action as a regulator reaching for a tool that does not fit a software deployment. The company's stated position is that a jailbreak finding a vulnerability in one model is not a reason to invoke an export-control instrument historically used for things like advanced chips, strong encryption, and dual-use military goods.
The technical premise matters because the order, as Anthropic describes it, runs on a specific logic. A jailbreak surfaces a flaw in one model, and the government treats that flaw as if it were an exportable good. The premise breaks down, by the company's account, when the same flaw can be found in other publicly available models without any bypass at all. In that case the supposed national-security harm is not tied to Fable 5 or Mythos 5 specifically. It is a property of the broader class of frontier AI systems, none of which is being similarly restricted right now.
Mowshowitz and cited commentators, including Dean W. Ball, frame this as a misunderstanding of what a jailbreak is and how defense in depth works in modern AI deployments. The offense a jailbreak causes is to the model's safety training, not to a unique capability. Pulling the model because someone published a prompt that works around its guardrails, the argument goes, is closer to recalling a car because a mechanic published a way to disable its alarm than to invoking arms-control authority.
The order's compliance mechanism makes the fit worse. Anthropic cannot reliably determine which of its users are US persons and which are foreign nationals, so it disabled the models for everyone. That is the practical shape of an export-control order aimed at software: it has to be enforced as a total ban, because the regulated item is a digital download, not a crate of hardware. The mechanism is the news.
A second-order question follows. If the technical harm is real but shows up across the entire category of public frontier models, then a single-model order either overstates the harm, by treating a category-wide problem as a property of one system, or understates the regulator's actual concern, in which case the order's narrow scope is a tell. Either reading is uncomfortable. The first makes the letter look like a regulator using the largest available hammer on a Friday night. The second implies a follow-up order is already in the mail.
The Friday timing is part of the story. The order landed shortly after 5 p.m. Eastern, the window Mowshowitz and other Washington-watchers flag as the standard "bad news dump" slot, when administrations push out material they expect the press cycle to absorb rather than scrutinize. The exact time — 5:21 p.m. Eastern — is on the record in the cited coverage, per Zvi's post, and corroborated by CNN's reporting.
The story has moved beyond Zvi's original post. David Sacks, reportedly speaking for the administration, confirmed the core facts: a trusted partner came forward with a jailbreak of Fable's guardrails; the administration asked Anthropic to fix it or de-deploy; Anthropic refused. Sacks characterized the models as a "cyberweapon" that requires regulation — language Anthropic itself used when describing Mythos, but which Anthropic contests as a mischaracterization of a minor vulnerability.
Anthropic's counter, as described by Zvi and consistent with the company's stated public position: the same vulnerabilities exist in other publicly available frontier models, including GPT-5.5, without any bypass required. If the national-security concern is category-wide, restricting only Fable and Mythos is either arbitrary or the opening move in a broader crackdown. The Lutnick letter has not been independently published in full; the Commerce Department's underlying legal reasoning remains second-hand.
What a reader should take from this is not a verdict on whether AI export controls are good or bad. It is a question worth holding for the next time a national-security instrument gets aimed at a software deployment: is the technical harm actually here, is it unique to this model, and is the instrument being used because it fits, or because it is the one within reach?