The Grounding

N°14 · The Export Control On June 10, Anthropic's CEO argued in public that the government should be able to ground an unsafe model the way the FAA grounds an unsafe aircraft. On June 12, the government grounded his. This is the chapter where the book's central thesis — that institutional architecture is a safety mechanism — meets the one mechanism the architecture could not design around: the state. It is also the moment the trillion-dollar thread stops being about valuation and starts being about sovereignty.

Context For a deep-dive timeline and discourse analysis of the 72-hour window between the Fable 5 launch and the government’s export control directive, see my research reporting on Substack: Anthropic’s Mythos Gambit Just Hit.

IThree Days

The launch was, by the company's own benchmarks and by independent ones, a peak. On June 9, 2026, Anthropic released two models at once: Mythos 5, the unrestricted frontier system held back for vetted partners under Project Glasswing, and Fable 5, the public-facing version — Mythos with guardrails, the cyber and bio capabilities walled off behind real-time classifiers. Fable shipped as the most capable model available to the general public, a claim third-party trackers confirmed. After three years of holding its most powerful work behind enterprise gates, Anthropic had finally put a Mythos-class system in front of hundreds of millions of people. The launch post conceded the obvious: a model this capable carries risk, which is why the safeguards existed.

Seventy-two hours later, both models were dark.

The arc between those two points is short enough to narrate in a single breath, which is part of why it landed as a shock. Tuesday: the most powerful public model in the world. Friday evening: a landing page reading that Fable 5 was temporarily unavailable, API calls returning errors, live sessions ending mid-stream, Claude Code and Claude.ai quietly defaulting new work to Opus 4.8. Nothing else in the lineup moved. Only the two Mythos-class models, and only them, went down — everywhere, for everyone.

The previous chapter ended on June 1 with a confidential S-1 and the argument that an option is not an obligation. Eleven days later the company received an instrument with no optionality in it at all.

IIThe Instrument

The thing that grounded Fable 5 was not a safety recall, a court injunction, or a content-moderation order. It was an export control.

At 5:21pm Eastern on Friday, June 12, Anthropic received a letter from Commerce Secretary Howard Lutnick, written with the Commerce Department's Bureau of Industry and Security. Citing national security authorities, it placed Mythos 5 and Fable 5 under export controls and instructed Anthropic to suspend all access to both by any foreign national — not merely persons abroad, but any non-citizen anywhere, expressly including Anthropic's own foreign-national employees. Per the letter, a license would now be required for the export, re-export, or domestic transfer of the models. Anthropic says the letter itself named no specific technical concern.

The mechanism matters more than the rhetoric, and it is worth stating precisely because it is what made the action so total. A modern frontier model is served from inference clusters to a global, pseudonymous user base. There is no reliable way to assert, in real time and at scale, the citizenship of every person sending a prompt — and the directive reached non-citizens inside the United States as readily as those outside it. A rule that cannot be enforced selectively can only be enforced absolutely. So the company did the only thing compliance allowed: it turned the models off for everybody, then said so.

This is the structural fact the book has to hold onto. The government did not order Fable 5 shut down. It ordered a category of person excluded from it — and the architecture of how the model is delivered converted a partial restriction into a complete one. The blast radius was a property of the system, not the order.

It was, as far as the record shows, the first time a leading AI company had taken a publicly deployed model offline because the federal government told it to.

IIITwo Accounts of the Same Friday

Within a day the event had two official narratives, and they did not agree about the most important fact: whether anyone had refused anything.

Anthropic's account treated the directive as a disproportionate response to a misunderstanding. In its statement, the company said the government appeared to be acting on a claimed jailbreak of Fable 5 — and that, as Anthropic understood it, the technique amounted to asking the model to read a specific codebase and fix the software flaws it found. The vulnerabilities surfaced this way were, the company said, minor and previously known, and discoverable by other public models without any bypass at all; it pointed to a competitor's system by name. No universal jailbreak — the kind that broadly unlocks the cyber capabilities the guardrails exist to contain — had been found by any of the testers who had spent thousands of hours on Fable before launch, including teams from the US government itself and the UK's AI Safety Institute. The company's posture was defense-in-depth, not perfection: make jailbreaks narrow or expensive, monitor aggressively, and retain thirty days of data to study them. On that reading, recalling a model deployed to hundreds of millions over a narrow, non-universal finding was a standard that, applied evenly, would freeze every frontier release in the industry.

We believe this is a misunderstanding.

— Anthropic, Statement on the US government directive (June 12, 2026)

The administration's account, which arrived through Axios reporting and then directly from David Sacks — co-chair of the President's Council of Advisors on Science and Technology and the administration's former AI czar — described something closer to a refusal than a misunderstanding. By this telling, a highly credible partner trusted by both Anthropic and the government, while testing Fable, produced a working jailbreak of the guardrail layer that separates the consumer product from Mythos's full cyber capabilities. The administration, Sacks wrote, asked Anthropic to remediate the flaw or pull the model. It did neither. The export control followed — issued, in his framing, reluctantly, with the door left open: patch the jailbreak, lift the control, return Fable to general release.

The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.

— David Sacks, PCAST co-chair, on X (June 13, 2026)

Sacks pressed on the one nerve most exposed by the company's framing: that minimizing a jailbreak which allegedly unlocked a cyber capability sits uneasily against a brand built, for five years, on treating safety as the non-negotiable priority. The accusation was not that Anthropic was unsafe. It was that, this once, Anthropic had chosen to keep a consumer product live over closing a safety gap — the precise inversion of its own founding story.

There is a third thread, and it is the most uncomfortable of the three. Reporting in the Wall Street Journal and Semafor traced the original flag not to a regulator but to Amazon — to CEO Andy Jassy, who is said to have told Treasury Secretary Scott Bessent and other officials that Amazon's own researchers had used Fable 5 to obtain information usable in cyberattacks. Amazon is Anthropic's largest investor, the source of a hundred-billion-dollar cloud commitment and much of the compute the company runs on. If the reporting holds, the company that bankrolls the infrastructure is the same one that told the government the product riding on it was dangerous. Amazon's confirmation was characteristically narrow: that governments do sometimes seek its counsel on security risks, and — separately — that its own cloud platform was among those knocked offline by the shutdown.

The book does not need to adjudicate which account is true. It needs to notice that, as of this writing, both are operative, both are sourced to named officials, and the gap between them is the whole story. One side says misunderstanding; the other says refusal. The same Friday supports both.

IVThe Argument for the Cage

Here the chapter becomes, in the manner this book keeps returning to, a specimen of its own subject.

On June 10 — one day after Fable's launch, two days before the directive — Amodei published Policy on the AI Exponential. Its central proposal was that governments should hold standing legal authority to block or reverse the deployment of frontier models that fail independent safety testing, with the explicit analogy that a regulator should be able to ground a dangerous AI system the way the FAA grounds an unsafe aircraft. The essay was an argument for a cage, written by the person with the most to lose from one, on the theory that a well-designed cage is what keeps the worst outcomes off the table.

Forty-eight hours later the government reached for that authority — or something wearing its clothes — and used it on him.

The irony is real, but the book should resist letting it collapse into mere irony, because the company's own response refuses to. Anthropic did not argue that the government has no business grounding models. It argued the opposite: that the state should be able to block unsafe deployments — through a process that is transparent, fair, clear, and grounded in technical fact — and that this action failed precisely those tests. The letter named no specific concern. The evidence, the company says, was conveyed verbally. The instrument chosen was an export control, not a safety statute, which made the mechanism a question of who may use the model rather than whether the model is safe — and routed the whole dispute through national-security authority that comes with far less obligation to show its work.

That distinction is the load-bearing one. Amodei did not ask for a black box that could disappear a commercial product on a verbal say-so. He asked for an FAA — an institution that grounds planes in public, on findings it has to defend. What he got, on the company's account, was the grounding without the institution. The argument for legitimate authority and the experience of its illegitimate use arrived in the same week, from the same direction, aimed at the same company.

This is what the book has been building toward without knowing it. The thesis across thirteen chapters has been that Anthropic's safety is structural — encoded in its charter, its governance, its talent, its research culture, the deliberate architecture of the institution. Every chapter has tested that wager against a different load: the OpenAI schism, the commercial wedge, the Nobel horizon, the trillion-dollar valuation, the confidential S-1. The structure held each time, because each load was one the architecture had been designed to bear.

The export control is the first load it was not. No clause in a long-term-benefit trust, no composition of a research team, no constitutional training method places a model beyond the reach of a Commerce Department letter. The institution Anthropic built to make itself safe against itself and against the market has no corresponding organ for the state. Sovereignty is the one input the architecture cannot internalize — and the grounding is the moment that gap became visible.

VThe Unsettled Ledger

A few facts sit at the edges of the record, vivid and thinly sourced, and the book should carry them as exactly that.

The directive's reach into Anthropic's own non-citizen staff produced the period's defining image: researchers locked out of their employer's most capable model by their passports. The press attached names to the abstraction — at least one prominent non-citizen scientist reported barred from the systems — and the symbol did more work than any benchmark. Around the same hours, the Defense Secretary posted that the department had expelled Anthropic from its building months earlier and that events were vindicating the choice; reporting also surfaced a claim that a foreign group had at some point accessed the model. Each of these is real reporting and none is yet bedrock. Mark them, in print, as reported and contested.

The market told its own version. The confidential S-1 of June 1 had let the secondary markets price Anthropic as an inevitability; the synthetic instruments that track its implied private valuation moved lower on the shutdown. The chapter before this one argued that an IPO had been priced before it was filed. This one adds the variable that pricing had not contained: that a model can be the most valuable thing a company owns on Tuesday and a regulated munition by Friday, and that sovereign risk, once introduced, does not leave the cap table.

As of June 14, Fable 5 and Mythos 5 remain offline for essentially everyone. Anthropic calls it a misunderstanding it is working to resolve. The administration says the ball is in Anthropic's court. Both descriptions can be true at once, which is the condition the book has documented from its first page: a company watched by everyone, understood by almost no one, arguing — now to the government itself — that the careful version of the story is the accurate one.

The grounding is not the end of the wager. It is the first time the wager was tested by a player who was never sitting at Anthropic's table, and never agreed to its rules.