On Friday 12 June 2026, at 5:21pm Eastern time, US Commerce Secretary Howard Lutnick sent an export control order directly to Anthropic’s chief executive, Dario Amodei. The instruction was specific: suspend access to the newly launched Fable 5 and Mythos 5 models for foreign nationals, on national security grounds. Within hours, both models were unavailable to every user, everywhere, including inside the United States.
That is the whole story in two sentences, and almost none of the weekend’s coverage engaged with it on those terms.
Anthropic’s explanation, such as it was, came down to a single architectural admission: the company does not have the infrastructure to apply nationality-based access controls at the API layer in real time. Told to exclude one population from a global service, and lacking any means of doing so selectively, Anthropic excluded the service. Everyone lost access because no one could be kept on.
This is worth sitting with, because it inverts the framing every outlet reached for first. The order was not a demonstration of control; it was a demonstration of its absence. A government with the authority to direct a private company’s product availability discovered that the company itself could not execute the instruction as written, and the gap was filled by removing the product entirely. Sovereignty, in the sense the order implies, is the ability to decide who gets access to what, based on who they are. That did not exist on either side of the transaction. Washington could issue the instruction. Nobody could fulfil it precisely.
The CLOUD Act debate has spent years establishing that data residency is not the same thing as sovereignty; a server in Dublin owned by a US company is still reachable by a US subpoena. Friday’s order extends that lesson in the other direction. Nationality-based access control is not sovereignty either, when the underlying architecture cannot support it. What looked like an act of state power was, mechanically, an act of state power running into a wall and taking everyone down with it.
The Telegraph’s framing, Britain cut off from frontier AI, was the dominant domestic angle over the weekend, and it is the wrong angle. Britain was not singled out. The United States cut itself off too, along with every other jurisdiction, as a side effect of an order aimed at “foreign nationals” in the abstract. Treating this as something done to Britain specifically obscures what is actually the more interesting story: Britain’s response to a universal pulldown was to ask for an exemption.
Reporting over the weekend has Number 10 actively lobbying for a UK carve-out, with Jade Leung understood to be coordinating discussions around a possible “approved countries” list. Set the two facts side by side. An export control order, issued by one government, with no working mechanism to target the population it names, produces a global outage; and the British government’s response to that outage is not to question the order’s legitimacy or to accelerate domestic alternatives, but to request inclusion on a list that a foreign Commerce Department would maintain, for a product built by a company headquartered in that foreign jurisdiction, governing access for British citizens to a service hosted on that jurisdiction’s terms throughout.
That is not a sovereignty story in the sense the Telegraph’s headline implies. It is a dependency story. The pulldown did not reveal that Britain lacks sovereignty over American AI infrastructure; everyone already knew that. It revealed that Britain’s preferred response to losing access to something it has no sovereignty over is to apply for better terms within the existing arrangement, rather than to treat the outage as evidence the arrangement itself is the problem.
Anthropic’s own account of Friday’s order is notably thin, and the thinness is doing the work. The company said it received only “verbal evidence” of a “potential narrow, non-universal jailbreak,” and pointed out that the underlying capability, asking a model to read a codebase and identify or patch flaws, exists in other publicly available models, including OpenAI’s GPT-5.5. Both models had, days earlier, passed thousands of hours of red-teaming by the US government, the UK’s AI Security Institute and third parties, none of which surfaced a universal jailbreak.
Weekend reporting attributed considerably more colour to whatever was actually found, including claims that a Pliny-style jailbreak of Fable 5 had surfaced both exploitation chains and information related to dangerous chemical synthesis. Whatever the precise technical detail, there is a wide gap between “previously known issues, verbally described, also present in competitors’ models” and the substance reporting suggests was extracted, and that gap is exactly where the next section starts.
Read in isolation, Friday’s blackout looks like a sudden response to a sudden discovery: a jailbreak found, a directive issued, an outage within hours. Read against the institutional record, it looks like the third act of an argument the American state has been having with itself since February; one in which two branches of government have spent months acting as though Anthropic’s most powerful models are simultaneously too dangerous to touch and too useful to give up.
On 27 February 2026, Secretary of War Pete Hegseth announced that he was directing the Department of War to designate Anthropic a “Supply-Chain Risk to National Security” under 10 U.S.C. §3252; a statutory category whose previous applications had been reserved for firms with ties to foreign adversaries, not domestic companies. Anthropic’s own response named the substance of the dispute precisely: the designation followed months of negotiation that broke down over two exceptions Anthropic refused to grant, for the mass domestic surveillance of Americans and for fully autonomous weapons. Judge Rita Lin’s subsequent preliminary injunction, granted in March, found the government’s stated reasoning “pretextual” and its real motive “unlawful retaliation,” writing that “if the concern is the integrity of the operational chain of command, the Department of War could just stop using Claude. Instead, these measures appear designed to punish Anthropic.” A parallel designation under separate statutory authority remains contested in the DC Circuit, so the matter is not fully settled; but on the central question of intent, a federal judge has already weighed in, and not in the government’s favour. That is the Title 10 track.
On 5 June 2026, a week before Friday’s order, the Financial Times reported the other half. Roughly six Anthropic engineers had been embedded inside the National Security Agency as “forward-deployed” staff, working to deploy and customise Mythos, Anthropic’s most capable and most tightly restricted model, for the agency’s own use. One person familiar with the arrangement told the FT the model would be useful for infiltrating the networks of states such as China and Iran; the FT itself was careful to note that whether the engineers or the model are involved in live operations remains unconfirmed. What is confirmed is the posture: an intelligence agency, under a different statutory framework and a different oversight regime, quietly running the arrangement that the Department of War had a few weeks earlier spent a news cycle declaring too dangerous to touch. That is the Title 50 track.
Holding both tracks in view changes what the “supply chain risk” designation actually meant. It was never a finding that Anthropic’s models were too capable or too dangerous for sensitive government use; another arm of the same government was making active use of the company’s most capable model, embedding its own engineers to do so, within weeks of that designation taking effect. The dispute was about two specific uses Anthropic would not authorise, not about its models in general; and the legal architecture under which the NSA arrangement operates is, by the FT’s own account, unclear. An arrangement that places vendor engineers inside an intelligence agency to adapt a restricted model for offensive cyber work does not obviously sit within either the commercial relationship that a §3252 designation would restrict, or the kind of formal intelligence-community contracting that would normally carry its own oversight.
This is also what explains the timing of Friday’s order, which the “sudden jailbreak” framing struggles to. Mythos has reportedly been available to the NSA in some form since at least April. What changed on 9 June was that Fable 5 and Mythos 5 went to general availability. That move took the underlying capability from a contained Title 50 asset, accessible only within the security state’s own architecture, to a public commercial product running across standard internet infrastructure, reachable by anyone with an API key. The Department of War’s February designation could only ever touch Anthropic’s relationship with military contractors; it had no power to restrict public access to a model that did not yet publicly exist. An export control directive can reach exactly that. Whether or not the jailbreak Anthropic describes is the actual trigger, Friday’s instrument is the one tool available to the government that could touch the thing February’s designation structurally could not: the public release itself.
The £500 million Sovereign AI Fund was the government’s flagship answer to exactly the anxiety Friday’s order has now made acute: contracts of up to £5 million per project over 12 to 24 months, run through a Sovereign AI Unit structured on venture-capital lines, with successful bidders retaining all background and foreground IP and the government left holding usage rights. Whatever its merits as industrial policy, it was never sized for this. A law firm’s analysis of the episode, published the day after the pulldown, made the underlying point without needing to dress it up: nothing currently available domestically comes close to Fable 5 or Mythos 5 on capability, and that gap will not close on the fund’s timescale. Five hundred million pounds, paid out in five-million-pound parcels over years, against a dependency that a single Friday-evening letter from Washington just proved cannot be substituted on demand. The thing withdrawn was never the kind of thing the fund’s structure was built to replace.
The week’s other resignation supplies the sharper illustration. On 11 June, the day before Lutnick’s letter, Al Carns resigned as Armed Forces Minister alongside Defence Secretary John Healey, over a Defence Investment Plan settlement of £13.5 billion against the £28 billion military chiefs had asked for. His resignation letter argued that “the character of conflict is changing faster than our procurement can keep up with,” and that the country was governing a more dangerous world with machinery built for a calmer one. Days later, with Fable 5 and Mythos 5 dark, Carns reached for the same argument about a different industry entirely: British researchers, companies and hospitals had been using the most capable AI model on the planet, and then, overnight, were not. This wasn’t an AI story, he said. It was the story of “every industry we used to lead.”
Carns is not an outside commentator reaching for an analogy. He resigned from government the day before this happened, making precisely this argument about defence procurement, and watched it recur in a different sector within the week. The structural complaint is identical in both cases: a gap between the scale of the problem and the scale of the response, and institutions built for a world that has already changed. £13.5 billion against a requested £28 billion for defence; £500 million, in five-million-pound parcels, against a frontier-AI dependency whose scale Friday night made impossible to ignore. The order didn’t create that gap. It switched off the thing everyone had been quietly depending on while the gap went unaddressed.
Strip away the framing each outlet brought to this story and what remains is simple. An order premised on the ability to distinguish between nationalities produced a global outage, because that ability does not exist. The same government that spent February and March publicly designating Anthropic a national security threat spent April, May and June quietly embedding that company’s engineers inside its most sensitive intelligence agency, under a legal framework nobody involved has been able to specify. A close ally’s response to losing access to infrastructure it never controlled is to ask for better terms within an arrangement it still doesn’t control. Everyone in this story is performing sovereignty. Nobody in this story has it; that includes, on the evidence of Friday night, the government that issued the order.
Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” 12 June 2026.
Anthropic, “Statement on the comments from Secretary of War Pete Hegseth,” 27 February 2026.
Financial Times, report on embedded Anthropic engineers and Mythos at the National Security Agency, 5 June 2026, via Axios, Tom’s Hardware and Investing.com.
Axios, “Trump admin blocks foreign access to Anthropic’s most powerful AI,” 12 June 2026.
CNBC, “Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive,” 12 June 2026.
NPR, Breaking Defense and AP coverage of Judge Rita Lin’s preliminary injunction, 26 to 27 March 2026.
Sunday Telegraph Business, 14 June 2026. News coverage, an analysis sidebar, and an opinion column on the Fable 5 and Mythos 5 shutdown.
Mishcon de Reya, “Who pulls the plug: The Anthropic Fable affair and what it means for everyone else,” 13 June 2026.
Al Carns, resignation letter as Armed Forces Minister, 11 June 2026, via Forces News and CNN.
Euronews, “Wake-up call: Europe reacts to Anthropic halting access to its Fable 5 and Mythos 5 AI models,” 13 June 2026.
Cross-reference: Race to Control
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy—with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
