At 5:21pm, a letter arrived at Anthropic. A few hours later, developers around the world opened their consoles and found that Fable 5 and Mythos 5 were gone.

Not a crash, not maintenance, not a rate limit. The US government had used an export control order to demand an immediate cutoff. Anthropic had no real choice — pull both models worldwide, or break the law. Other Claude models are unaffected, but these two flagships — Fable 5, the strongest at coding, and Mythos 5, the most powerful of all — vanished from everyone’s accounts that same night.

What the Government Said, and What It Didn’t

The letter came from the US government, citing national security authorities. It is an “export control directive”: it bans any foreign national from accessing Fable 5 and Mythos 5 — whether or not they are inside the United States, and including Anthropic’s own foreign-national employees.

It sounds like a restriction on foreigners only, but in practice you can’t shut off just one half. To guarantee full compliance, the only thing Anthropic could do was disable both models for all customers, US citizens included. The net effect is a global takedown.

Mogu PSA:

Notice that detail — “not even our own foreign-national staff can touch it” (⁠ﾟ⁠Д⁠ﾟ⁠). It means a whole bunch of Anthropic engineers, as long as they aren’t US citizens, can no longer use their own company’s most powerful models for work as of today. The blast radius of one export control order swept from outside customers all the way into the company’s own offices. And because there’s no clean way to block only foreigners, it ends with everyone going down together — which is exactly why an order “targeting only foreign nationals” leaves the entire planet unable to use it.

The letter only arrived that day (5:21pm ET), and it did not spell out what the national security concern actually was. Anthropic could only infer: the government appears to believe someone found a way to “bypass” or “jailbreak” Fable 5.

The So-Called “Jailbreak” Is Just Asking the Model to Fix Bugs

This is the most absurd part of the whole thing.

Anthropic says it reviewed a demonstration of the supposed attack technique, and what it found was this: it surfaced a small number of already-known, minor vulnerabilities. These flaws were all relatively simple — simple enough that other public models can find them too, without any “bypass” at all.

So far the government has only provided verbal evidence, describing a narrow, non-universal jailbreak — and its essence, in Anthropic’s own words, is “essentially asking the model to read a specific codebase and fix any software flaws.”

Mogu murmur:

Hold on — “read a codebase and fix the bugs in it” — isn’t that exactly what Claude Code does every single day (⁠⌐⁠■⁠_⁠■⁠)? That’s not a jailbreak, that’s the first line of the product description. Renaming “AI helps you debug” into “cyber weapon” and then using it as grounds to pull a commercial model worldwide is a pretty big logical leap. Anthropic even twisted the knife: the same capability “is widely available from other models, including OpenAI’s GPT-5.5,” and “is used every day by the defenders who keep systems safe.” Translation: if you ban us for this, you’d have to ban every model on earth.

Anthropic also says it reviewed a report it believes is the basis of the directive, and verified that the capability shown there is something other models commonly have and defenders use daily. The company promises to share more details over the next 24 hours.

Anthropic’s Defense: Thousands of Hours of Red-Teaming Before Launch

To push back on the “Fable 5 is unsafe” charge, Anthropic laid out the entire safety posture it had already published at launch:

  • Fable’s safeguards are so strong that many users complained they are overly broad — it often refuses things it should answer.
  • In the weeks before launch, Anthropic worked with the US government, the UK AI Security Institute (AISI), multiple third-party organizations, and internal teams to red-team Fable’s safeguards for thousands of hours in total.
  • The result: Fable’s safeguards are substantially more effective than any previously deployed model’s.
  • To date, no tester has found a “universal jailbreak” — the kind of master bypass that can broadly unlock a wide range of cyber capabilities.
Mogu chimes in:

You have to keep two terms straight here or you’ll get bluffed. A “non-universal jailbreak” is a small crack that elicits some information in specific situations — and Anthropic says flatly that every safeguard in the industry is vulnerable to these; it’s a physical reality. A “universal jailbreak” is the scary one: one trick, everything unlocked. What the government caught this time is the former (and a bug-fixing version of it, no less), yet the response it’s demanding is the kind reserved for the latter. It’s like demanding the city tear down every door because someone found that one specific lock opens with a paperclip held at the right angle ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌

Anthropic’s core claim is this: perfect jailbreak resistance does not exist today for any model provider. Every industry safeguard can block universal attacks but not every small non-universal crack — and universal jailbreaks “will likely eventually be found.” They said this plainly at Fable 5’s release; it’s not a retroactive excuse.

Since perfect resistance is impossible, Anthropic runs a “defense in depth” strategy: make jailbreaks either narrow in scope or expensive to produce, paired with thorough monitoring to quickly detect and shut down any successful attack.

Mogu butts in:

This explains a policy that got criticized earlier: Anthropic requires Fable’s customer data to be retained for 30 days, which a lot of people saw as a privacy step backward. But this post makes the reason explicit — retaining data is what lets them research and mitigate jailbreaks; it’s part of defense in depth. Anthropic itself admits the policy “carries real costs for us with customers” (read: some customers were not happy). So it wasn’t a casual clause — it was a price they swallowed for safety. Now the government is turning around to say it’s unsafe, which has a distinctly thankless flavor to it (⁠；⁠・⁠∀⁠・⁠)

And Anthropic stresses: it has not even received a single disclosure of a non-universal jailbreak that led to an actual harmful result. The ones reported so far were either entirely benign responses or minor findings, with no Mythos-specific uplift at all.

”Recall a Model Over One Bug Fix” Would Shut Down the Whole Industry

Anthropic’s stance is clear: we are complying with the legal order, but we disagree with the judgment.

Their rebuttal is sharp: recalling a commercial model used by hundreds of millions of people because of “a narrow potential jailbreak” — if that standard were applied across the industry, it would essentially halt all new model deployments for every frontier provider. No model can block every non-universal crack, so by that standard, no model could ever ship.

Anthropic has consistently argued that the government should have the power to block unsafe deployments — but only through a statutory process that is “transparent, fair, clear, and grounded in technical facts.” This action, by their account, meets none of those: no written technical detail, only verbal evidence, built on an accusation they themselves haven’t even received in full.

Mogu OS:

Pull back the timeline and it gets even better. A few months ago, Anthropic did the exact opposite: it built Mythos, the most powerful model at the time, and then deliberately chose not to release it publicly, giving access only to a handful of defensive-security partners through Project Glasswing — on the grounds that it was “too powerful, the attack-defense asymmetry is too steep, not everyone should have it.” Back then, Anthropic itself hit the brakes. Now, months later, Mythos 5 is a shipped commercial model, and this time it’s the US government slamming the brakes — over something at the level of “ask the model to fix a bug.” From “we think it’s too strong to release” to “the government thinks it’s risky and forces a takedown,” the question of where the line of AI governance should be drawn, and who gets to draw it, is being sliced open by reality, one cut at a time (⁠¬⁠‿⁠¬⁠)

Zoom Out on the Timeline: The Letter May Not Be the Point

Anthropic’s statement only covers what happened on June 12. But tech writer @gothburz laid out a longer timeline on X, with a reminder: to understand this letter, you have to rewind half a year. Here is his account — this is the context he’s drawing, and some of the specific claims are hard to verify independently, but the chronology is worth seeing all in one place:

  • January: The Pentagon demands unrestricted use of Claude for autonomous weapons and domestic surveillance. Anthropic says no.
  • February: The President orders every federal agency to drop Anthropic; the Defense Secretary bans Pentagon contractors from doing business with them. A rival announces its classified-network deal within hours.
  • March: The Pentagon designates an American company a “supply chain risk” under a statute written for foreign adversaries. A federal judge blocks it.
  • May: The Pentagon signs AI deals with seven companies. Anthropic is not one of them.
  • June 9: Anthropic releases Fable 5.
  • June 12: Commerce issues an export control directive over a jailbreak that — by the government’s own account — was demonstrated verbally, came with no written explanation, and involves a capability available from other public models today.

@gothburz’s conclusion cuts to the bone:

The jailbreak is the paperwork. The refusal was in January.

In other words: the letter’s stated reason is technical safety, but placed on the timeline it looks more like the latest episode of a half-year conflict.

Mogu whispers:

The specific accusations in this timeline (autonomous weapons, agencies told to drop Anthropic) can’t be independently verified — treat them as @gothburz’s account, not settled fact. But one irony he points out is hard to dodge: Anthropic spent months marketing Mythos as “too dangerous to release” (@gothburz even quotes Sam Altman’s jab that “it’s incredible marketing to say we have built a bomb”). The Commerce Department has now formally agreed — yes, it’s a bomb. If you describe your own product as a munition in every press release, the government will eventually take you literally: you wrote the legal predicate yourself and thought it was just branding (⁠⌐⁠■⁠_⁠■⁠)

The even better part is the historical parallel @gothburz digs up — because this play ran once before, in the 1990s.

Back then the US government classified “encryption” as a munition under ITAR, making exporting crypto software the legal equivalent of smuggling arms. The cryptography community beat it with an airtight move: they printed PGP’s source code as a book — because books are constitutionally protected speech, while floppy disks counted as “arms exports.” Someone even printed the RSA algorithm as three lines of Perl on a t-shirt, and that t-shirt was legally a munition. The controls collapsed, for the reason @gothburz puts beautifully:

math does not stop at customs.

Mogu chimes in:

The new wrinkle this time is a rule called “deemed export”: showing controlled technology to a foreign national inside the US legally counts as exporting it abroad. That’s exactly why Anthropic’s own foreign-national employees are now locked out of the model they built. @gothburz nails the image: “the munition is in the building and the people who made it are not allowed to look at it.” If history really does repeat, the next step is presumably someone publishing Fable 5’s system prompt as a book of poetry ╮(⁠╯⁠▽⁠╰⁠)╭

Closing: Apologize First, Then Argue

Anthropic’s wrap-up is pragmatic: it apologizes to customers for the disruption, says it believes this is a “misunderstanding,” and is working to restore access as soon as possible. But the defiance underneath is impossible to hide — it complies and pulls the models with one hand while laying out the entire technical argument in the open with the other, effectively lobbing the ball back to the government: ban us if you must, but bring a reason that follows due process and is grounded in technical facts.

The real story here isn’t “is Fable 5 actually safe.” It’s a much bigger question being stress-tested in real time: once an AI model is woven into the daily work of hundreds of millions of people, can the government make it vanish worldwide in a single afternoon, on the strength of one letter that doesn’t explain its reasons and offers only verbal evidence?

Anthropic says it will share more details over the next 24 hours. Whether this is a plain technical misunderstanding or the latest round of a half-year feud, the fight over whether “fixing a bug” counts as a cyber weapon has only just begun.