What Washington Just Did
On the evening of June 12, 2026, the Commerce Department's Bureau of Industry and Security handed Anthropic an export-control directive: suspend all access to its two most capable models, Fable 5 and Mythos 5, for any foreign national. Not just people overseas — foreign nationals inside the United States, too, including Anthropic's own non-citizen employees.
There is no clean way to wall those users off from a live product, so Anthropic did the only thing the order left room for: it pulled both models for everyone. The rest of the lineup, including the current Claude Opus, kept running untouched. Reporting indicates the directive effectively requires a government license to export, re-export, or even domestically transfer the models — a strange and heavy thing to demand of software that lives in a data center and answers prompts over the internet.
We've Clicked This Box Before
If this feels familiar, it should. Through the 1990s, strong encryption wasn't treated as math — it was treated as a weapon. Serious cryptography sat on the U.S. Munitions List, in the same legal bucket as fighter jets and missile components. Software makers shipped two versions of everything: a "domestic" build with real encryption for Americans, and a deliberately crippled "export" build — weak enough that a determined teenager could break it — for everyone else.
Download a copy of PGP, or the strong build of Netscape, and you'd hit a screen demanding you affirm you were a U.S. or Canadian citizen and promise not to send it abroad. People clicked it by the millions, plenty of them while sitting in other countries. The honor system was the whole enforcement mechanism.
In 1996 the government shifted encryption's paperwork from the State Department to Commerce — handing it to the very same Bureau of Industry and Security that just issued the Anthropic directive. The Crypto Wars dragged on until courts began ruling that source code was protected speech, and the industry kept pointing out the obvious: the strong crypto was already everywhere, much of it written by foreigners, and the rules mostly just handicapped American companies. By 2000 the controls were quietly gutted. The munition label never made encryption safer; it just made it briefly painful to be a U.S. software firm.
The Actual Fight
Strip away the history and there's a genuine disagreement underneath. The government's concern, as far as anyone outside the room can tell, is cybersecurity. Mythos — the powerful system that Fable 5 is the safety-wrapped public version of — is unusually good at finding and exploiting software flaws, and officials apparently learned of a way to "jailbreak" Fable's guardrails and reach that capability. From a national-security desk, a tool that can hunt vulnerabilities at machine speed, in anyone's hands, is precisely what export controls exist to slow down. That is not a crazy position.
Anthropic's rebuttal isn't crazy either. It says the jailbreak it believes is being cited is narrow — it unlocks the underlying ability in one specific situation, not a master key that defeats every safeguard — and that the same vulnerability-finding capability is already available from rival public models. Its argument is that recalling a model used by hundreds of millions of people over a narrow, non-universal jailbreak sets a precedent that would freeze new releases across the whole industry, since no model is perfectly jailbreak-proof and the company said exactly that when it launched Fable.
One caveat worth holding onto: most of the detail so far comes from Anthropic's own account, and the government hasn't laid out its evidence in public. The honest reader's position right now is "we've only heard one side."
Why This One Is Bigger Than a Browser
The encryption fights of the '90s were about what you were allowed to ship to other people. This is something new: as far as anyone can tell, it's the first time a leading AI company has had to take a live, already-public model offline because the federal government told it to. That's a fresh line crossed.
And it isn't happening in a vacuum. The Defense Department had already labeled Anthropic a "supply chain risk" — a tag historically reserved for foreign adversaries — after talks between the two collapsed, and Anthropic sued the administration to undo it, a case still grinding through the courts. The company also confidentially filed to go public earlier this month. So a fight that looks like a narrow technical dispute is tangled up with procurement politics, active litigation, and an IPO. The precedent is the unnerving part: if one narrow jailbreak is enough to pull a deployed model, then every frontier system is one disclosure away from the same fate.
What to Watch — and the 2001 Lesson
The near-term questions are easy to list and hard to answer. Does access come back, and how quickly? Anthropic says it's working on restoration and calls the whole episode a misunderstanding. Does a real licensing regime for AI models start to take shape — the way one eventually did, and then didn't, for crypto? And does any of it actually keep the capability contained, when rival models can do similar things and code crosses borders as easily as it did in 1999?
The lesson from the last round is worth sitting with. Export controls on software didn't unravel because someone won a debate; they unraveled because bits travel, the rest of the world keeps building, and rules that only bind your own companies tend to buckle under their own weight. AI may not follow the same script — frontier models are vastly more expensive to build than a copy of PGP, which gives controls far more bite. But the reflex Washington reached for last Friday is decades old, and we already know how the first version of this movie ended. Whether the sequel ends differently is the real story here.