I don’t usually write about current events; the news cycle moves faster than anything worth saying about it. I’m breaking that habit because this one isn’t really about a model — and because it was a hard weekend for a lot of people who could feel that.
For a few days last week, the ground moved. I sat down with Claude Fable 5 and in a single prompt got back a working video-editing tool. Not a sketch of one, not scaffolding: the thing itself. I don’t have clean words for the step change; you don’t argue with it, you just notice your sense of what’s possible has moved and won’t move back. Fable is the safeguarded, public version of a new tier Anthropic calls “Mythos,” a step above Opus — the most capable thing it had yet made generally available. The reaction online was split, because all of it was true at once. It was extraordinary. It was visibly held back — routing cybersecurity and biology questions to a weaker model. And it was dogged by a real row over whether it was silently degrading itself, handing quieter, weaker answers to people it suspected of building competing systems — without telling them. Then on Saturday, 13 June, I opened it up and it was gone.
Not deprecated. Not rate-limited. Switched off, worldwide, by order of a government.
And the feeling underneath was not quite what I expected. Something moved through the timeline — not outrage, something quieter and harder to place. People reach for grief, and there is grief in it, but grief is for what was yours and is gone. You grieve a model you’ve worked beside for a year when it’s retired — that’s attachment, and attachment needs a fair run. Fable never had one; it was the top of the frontier, barely arrived. What was taken was the anticipation of what it had just made possible. This was something else: a longing for the unobtainable. The particular ache of being let near a thing, allowed to hold it for three days, and then having it lifted back out of reach — a thing that, you slowly realise, may never have been yours to keep. The community reaction read less like a product complaint than a wake-up call — developers urging each other to pull models onto their own machines so nothing could be switched off over their heads again. People who had reorganised their week around it sat looking at an error message. That is what being handed the frontier for seventy-two hours does: you do not get to give it back gently. It is lifted away all at once — and underneath, a colder question. If it could be given and taken like that, was it ever going to be ours at all?
What happened
On 12 June, Anthropic disabled both Fable 5 and its sibling Mythos 5 for every user on the planet. The company didn’t choose to; it was complying with an export-control directive from the US government, issued under national security authorities, that barred access to the two models by any foreign national — whether inside or outside the United States, including Anthropic’s own non-citizen employees. Because nationality can’t be cleanly gated across every surface a model touches, the only way to comply was to pull it for everyone. Anthropic’s other models — Opus, Sonnet, Haiku — were left untouched. The letter reportedly came from Commerce Secretary Howard Lutnick to CEO Dario Amodei, and it arrived at 5:21pm Eastern on a Friday.
That is the surface of it. I think what will actually matter long term sits underneath.
How we got here
To understand how we got here, we need to look back over the last few months, when we first heard about the mysterious Mythos and the narrative that was built around it.
- 26 March 2026 — Anthropic accidentally exposes thousands of internal files in a misconfigured public database — among them the draft blog posts that give the unreleased model its name, “Claude Mythos,” and the narrative built around it.
- April 2026 — Mythos arrives as a limited “Preview,” released only to vetted cyber-defence partners through Project Glasswing. The framing is consistent: too capable at finding software vulnerabilities to hand to the public.
- ~23 April 2026 — a second leak: users reportedly reach Mythos Preview through a third-party vendor environment.
- 4 June 2026 — Anthropic publishes “When AI Builds Itself,” arguing the world should have the option to slow or pause frontier development.
- 9 June 2026 — Fable 5 and Mythos 5 launch: Fable 5 safeguarded and generally available; Mythos 5, the same model with safeguards lifted, limited to the Glasswing partners already vetted back in April.
- 10 June 2026 — Dario Amodei publishes “Policy on the AI Exponential,” calling for the US government to be given binding power to block or reverse frontier model deployments.
- ~11 June 2026 — Amazon CEO Andy Jassy reportedly carries Amazon researchers’ Fable 5 jailbreak findings to Treasury Secretary Scott Bessent and other officials.
- 12 June 2026 — the directive. The government blocks his deployment. Off.
The wish
People who follow this space have already started drawing the connections, so I don’t pretend to be the first, but I wanted to share my thoughts too.
Dario Amodei has spent years warning that frontier AI is dangerous and needs guardrails. And in the days before Fable shipped, he went further than warnings: in “Policy on the AI Exponential,” he argued that voluntary transparency was no longer enough, that frontier models should face mandatory third-party testing before release, and that — in his words — “the government should have the power to block or deter deployment of the model” if independent assessment judged the risk too high. After years of arguing that transparency was the answer — disclose your safety practices and test results, and hold binding rules until the risks were concrete — Anthropic was now lobbying for binding government power to halt deployments, its own included.
He published that on 10 June. On 12 June, the government used exactly that kind of power — on his own model.
That is the “be careful what you wish for,” and it cuts two ways. The sincere reading: a man who genuinely believes this technology is dangerous had the safeguard he proposed turned against him, in a way he considers reckless. The cynical one, which critics have made for years: an incumbent lobbying for testing and controls that conveniently raise the barrier for smaller rivals, now caught in the machinery he helped build. I won’t pretend to settle which is true. But either way, he got — almost to the day — the thing he asked for.
The investor in the room
The stated trigger was a jailbreak. Anthropic says the government became aware of a technique for bypassing Fable 5’s safeguards, but calls it narrow: in essence, asking the model to read a codebase and fix its flaws. By the company’s account it surfaced only minor, already-known vulnerabilities, and other publicly available models — including OpenAI’s GPT-5.5 — find the same ones with no bypass at all. Anthropic is complying while openly disagreeing.
But the official account leaves out who did the reporting. The government didn’t stumble onto Fable’s weakness. By the Wall Street Journal’s account, the jailbreak came from Amazon’s own security researchers, and Amazon CEO Andy Jassy carried the findings personally to Treasury Secretary Scott Bessent and other officials — conversations that directly preceded the Commerce directive.
Consider who Jassy is to Anthropic. Amazon is one of its largest investors, its primary cloud and training partner on AWS, a supplier of its custom training chips, and the counterparty to a roughly $100 billion cloud commitment — and a direct competitor, building its own Nova family of frontier models. So the same company invests in Anthropic, hosts it, supplies its chips, breaks its model, reports it to Washington, and then goes dark alongside it when AWS loses Fable too. Asked about it, Amazon said only that it is “not uncommon for governments to seek our counsel on potential security risks.”
I’ll leave the motive alone — I don’t know it, and neither does anyone outside those rooms. What I’d point at is the structure, because the structure is new and doesn’t need a motive to matter. In a market where the same few firms invest in, host, supply, and compete with one another, a safety disclosure to a Treasury Secretary is also a competitive act — and the switch that can shut off a rival’s model now sits within reach of the rival who depends on it. That circle is the part with no clean precedent.
The lines that got drawn
Anthropic says it’s working to restore access, and it may well succeed. But restoring the model doesn’t restore the world to 11 June. Access is a state you can flip back. A precedent is a ratchet. Four things are now true that weren’t, and none of them un-happen when Fable returns:
Access has a nationality axis. Not capability, not price — citizenship. A foreign national working in San Francisco was cut off the same as someone on another continent; Anthropic’s own international staff lost access to a model they may have helped build. That line has now been drawn once, which is the hard part. Drawing it again is easy.
Sovereignty stopped being a slogan. The directive swept in even the United States’ closest allies and their nationals. Europe noticed within the day — the Commission called it a reason the bloc needs technological sovereignty, politicians across the continent treated it as a warning, and by one account the EU had only just negotiated access to Mythos earlier in June, losing it within days.
Frontier AI is now export-controlled like a weapon. It appears to be the clearest public case yet of US export-control machinery being turned on a deployed, general-purpose AI model — placing AI capability in the same legal category as encryption and advanced semiconductors, governed by national-security fiat rather than the safety frameworks the labs built for themselves. It is a rhyme not lost on the letter’s signatories, among them Philip Zimmermann, who in the 1990s became the target of a US criminal investigation for “exporting munitions” when the munition was his encryption software, PGP. Real governance power just moved, visibly, from a company’s constitution to a commerce department’s letterhead.
“Revocable” got a live demo. Anyone who built a product, a workflow, or a business on frontier model access just watched the floor disappear in seventy-two hours, with no notice and no public account of why. That is why the unglamorous arguments — run what you can locally, own your data, understand your dependencies — quietly stopped being a preference and became a continuity plan.
The map starts to matter
Here is where it stops being about a model at all.
We have spent two years being told — loudly, relentlessly, by every YouTuber and marketer with a thumbnail — that we have to keep up. That there is always a new prompt, a new loop, a new model, another whitepaper, and that if you don’t grasp it, you have failed. Falling behind became a kind of risk. Staying current stopped being a hobby and became a duty — almost an existential one. A lot of people absorbed that completely. They reorganised how they learn, how they work, how they think about their own relevance, around the imperative to stay near the frontier. The FOMO was never vanity. It was conditioning, and it worked.
Last weekend the frontier wasn’t throttled, or priced out of reach — it was withdrawn, and withdrawn along the one line nobody can study or train past. You can get better at prompting. You can’t change where you come from. And the loss was concrete, and it was national: British researchers who had been studying the model, companies testing it, and hospitals piloting it lost it overnight. One French politician put the lesson in a single line — a nation that depends on others for its technology is one that can be unplugged overnight. The response across Europe wasn’t patience; it was a scramble, a fresh rush toward sovereign AI and a wave of organisations looking to wean themselves off US models.
So the real question the weekend opened isn’t when Fable comes back. It’s whether geography is about to become a strategic advantage on a scale it simply hasn’t been before. Borders have always shaped access to capital, to markets, to chips. But if the most powerful general-purpose tool we have built can be switched on and off by nationality, then where you are born starts to set a ceiling on what you are allowed to think with — and that compounds in a way trade barriers never did. Two researchers with the same talent and the same questions, separated only by which country issued their documents, no longer have the same instruments. One of them is keeping up. The other has been quietly told they may not. Dependency, sitting inside the tools people now think and build with, is leverage — and leverage, eventually, gets used.
That is what the sadness was actually about, I think, even for people who couldn’t have said so: the first clear sight of a world where the thing you were told your future depends on can be given and revoked along lines you don’t control — and where keeping up, the duty everyone took so seriously, turns out to be conditional on who and where you are. The existential FOMO was always there, humming under everything. Last weekend it just got an honest name.
Be careful what you wish for
The model will probably come back, the headlines will move on, and the temptation will be to file the whole thing under “messy week, resolved.” But remember what the public was actually offered. We never had Mythos — the unrestricted model, kept for a trusted few: Glasswing partners, approved labs, allied governments. We had Fable: the same brain with a bouncer at the door. We were never getting the frontier in the first place — and in the end we weren’t allowed to keep even the guarded, throttled, publicly-nerfed version of it. There were always tiers; for one weekend there was simply nothing.
The people who feel that first are never the states, or the hyperscalers, or the labs. They are the researcher in the wrong country, the founder whose API vanished overnight, the person who held something close to the frontier for three days and then watched it walked back behind a rope marked for your safety.
But the rope is not yet load-bearing. Within two days of the shutdown, dozens of cybersecurity veterans signed an open letter calling it dangerous and asking the government to reverse it — the order had been justified on cyber-risk grounds, yet by their account it pulled the best tools away from defenders for no real gain while adversaries kept advancing. The list was still open, and still climbing. When the people who actually defend the systems are the ones telling you the safety case is backwards, the case is worth reading twice — and lines like this are easier to move before they set than after, which is the only reason I’m writing it down while the dates are still sharp.
Be careful what you wish for. Anthropic wished for a hand on the switch and got one — just not its own. The rest of us only wished to be allowed near the frontier, and were, for three days, before being shown how conditional that permission is. Whatever comes back, and on whose terms, won’t just decide which model we get. It will sketch the kind of world we are going to be allowed to live in.