On Friday afternoon, the US government did something extraordinary: It issued an export control directive to suspend all access to Anthropic’s advanced Claude AI models, Fable 5 and Mythos 5, by any foreign national. This includes Anthropic’s foreign national employees.
There are three things to consider…
- Price the political risk into the AI trade: A national-security asset with a sovereign off switch should not carry the multiple of a clean secular growth story, and that gap will close.
- Value optionality: The firms that come through this best never bet their operations on a single closed model, from a single vendor, in a single jurisdiction.
- Watch the hedge, not the headline: Track how fast capital and governments outside the US start funding their own stacks because that flow is where the real realignment and money will be.
Ultimately, the engine under all this was built in a Cold War defense budget, and on Friday, the original owner reminded everyone whose engine it is. The mistake would be to read that as strength. You reach for the off switch when the durable forms of control, the physical ones, are starting to slip out of your hands.
THE SITUATION
Unless you follow this industry closely, you have probably never heard of Fable and Mythos. Mythos is not the chatbot your kids use to dodge their homework. It is unnervingly good at finding hidden security flaws in the world’s software, at machine speed. Anthropic says Mythos turned up thousands of serious, previously unknown vulnerabilities across every major operating system and web browser, including one that had gone unnoticed for 27 years. And it did this almost entirely on its own.
Anthropic was so unnerved by its Frankensteinian creation that it opted to never sell Mythos access openly. Back in April, Anthropic locked Mythos inside a vetted club of about 40 defenders it called Project Glasswing, with the White House reportedly curating the guest list and rejecting a plan to let in 70 more organizations.
The export control directive suggests Washington feels Project Glasswing is not up to the task. The White House didn’t technically outlaw Mythos and Fable, but Anthropic had no practical way to wall foreign users out, so it complied by switching both models off for everyone. Anthropic described the order as a “misunderstanding” and said it is working to restore access—perhaps thinking the blackout forces Washington’s hand. That seems like a misreading, as the US government is the one flexing its might here, and the two sides have been butting heads for months.
Back in February, Anthropic refused a Pentagon demand to strip out its own guardrails and let Claude be used for any “lawful” purpose, including fully autonomous weapons and the mass domestic surveillance of Americans. Washington retaliated by ordering federal agencies to stop using Anthropic’s models and branding the company a national-security, supply-chain risk—a fight the two sides are still grinding through the courts.
Against that history, the export control order looks less like a bolt from the blue and more like the next round of a brawl that started earlier this year.
The “Off Switch”
Geopolitical analysis is good at two things: identifying what a state must do to survive (i.e., its imperatives) and what it cannot do no matter how badly it wants to (i.e., its constraints). It is far weaker at a third thing: predicting what leaders will do. Leaders act on how they read their imperatives and are perfectly capable of charging into a constraint they refuse to see. Their individual decisions are often irrelevant. Constraints bind them, and the forces geopolitics analyze operate on timelines of years and decades, not the ephemeral sentiment of even the most authoritarian ruler.
So, after the US Commerce Department handed Anthropic an export control order, the question isn’t the instinctive “can it do that?” It can. The question is whether the order serves an American imperative or runs headfirst into a political and technological constraint. My tentative answer is that it does both… but that the constraint is mightier than the imperative.
Though the tech overlords think they are self-made men, the American technology industry grew out of a Cold War defense budget rather than a Palo Alto garage. There is none of the latter without the former. In 1958, the largest buyer of the integrated circuit was the US government, which purchased most of the early production to guide missiles toward the Soviet Union more precisely. Everything we now call “AI” sits downstream of that decision, and it shouldn’t come as a surprise that Washington owns an “off switch”—or rather, thinks it owns an off switch.
If the US government, or any government for that matter, thinks a technology can challenge the sovereignty of the state, the state will react. Cryptocurrencies are an example of this challenge, but AI is an order of magnitude greater.
The Imperative
After Anthropic previewed Mythos, a civil-liberties nuisance became an existential threat to state sovereignty overnight. Project Glasswing was supposed to help solve the problem, and Anthropic’s more general Fable was supposed to bring some of Mythos’s abilities to market with certain safeguards to prevent misuse from nefarious actors. But the US government claims that someone demonstrated a “jailbreak” that unlocks those capabilities.
Thus, Commerce reached for export control law, the same body of law that governs shipping a fighter jet or an advanced chip. Washington is treating access by a non-citizen, even one sitting at a desk in California on Anthropic’s payroll, as a “deemed export,” the legal fiction that handing controlled technology to a foreign national is the same as shipping it to their home country. It is a doctrine built for blueprints and centrifuge designs, now stretched over a chatbot login.
The US government imperative is real and pressing. A state’s first job is to keep a decisive advantage out of the hands of those who would turn it against it. The capability here is genuine. The median gap between a software flaw disclosure and exploit has collapsed from over 700 days in 2018 to roughly four hours in 2024.
The thing that keeps me up is not the superintelligence of the science fiction novels; it is the democratization of serious capability to actors who feel no constraint at all. By that measure, a government that gets nervous about an autonomous vulnerability-finder is not being hysterical; it is protecting its survival.
The Constraint
The US government cannot control this. To the extent it can, it is through physical things. You can choke AI by controlling chips, the fab that manufactures the most cutting edge of them (TSMC), the lone Dutch company that builds the machines inside those fabs (ASML), and the energy and water the data centers drink. Those inputs are physical, scarce, and slow to copy, which is precisely what makes them controllable. They are also replicable. Huawei, left for dead by the first Trump administration’s chip controls, is now wringing competitive performance from its own chips without US technology, as we discussed two weeks back.
A trained model is not a physical input. It is software, i.e., a file. Pull the American one and the capability seeps out anyway, because frontier ability no longer comes from any single model. It is distilled and recombined from many sources: open-weight Chinese systems (like DeepSeek), synthetic data, reinforcement learning, and the steady harvesting of several models at once. Denying China one American model raises the cost at the margin and not much more. (A copy of Mythos already appears to have leaked to a Discord group, and OpenAI has reportedly shipped its own answer.)
This is the chip-control story repeating, and the attempt will backfire the way most export controls backfire: without an all-out domestic build-out you mostly teach your rival to make the thing himself. Controlling software is the brittle, leaky version of a strategy that only ever worked because the controlled things were physical. And the physical lever no longer points one way: China holds its own chokepoints and dominates rare earths and critical minerals, and the whole democratic-AI edifice still rests on a single contested island, Taiwan.
A More Consequential Sequel?
There is a rich irony in Anthropic playing the wounded party here. Of all the frontier labs, Anthropic has been the loudest champion of a stronger government role in AI regulation. CEO Dario Amodei has spent the past year arguing for tighter chip and equipment controls on China and for a “democratic” AI bloc that denies frontier capability to authoritarian states. It asked for a world in which the state gets to decide which models are too dangerous to ship. It now has one and is discovering that the same logic does not stop politely at its own door, or at its own foreign engineers.
The real worry is that, yes, you can build a company in the US, land of the free, home of the brave… and have Washington pull the plug. Not because it is legal but because Washington senses a threat. We have watched this exact movie before, in another domain. In the 1990s, Washington classified strong encryption as a munition under its arms-export rules and put programmer Phil Zimmermann under federal criminal investigation for letting his PGP encryption software spread overseas after he released it in 1991. The government dropped the case in 1996 without ever filing charges because, by then, the code had spread worldwide and the courts were beginning to treat software as protected speech.
You cannot, it turns out, unpublish math. Perhaps we are watching a far more consequential sequel of that story play out.
The Winners and Losers
The government wins the narrow point: In any contest between the most valuable companies in history and the state, the state wins because power rests with whoever commands force, not servers, and a firm that obeys a 5:21 pm order within hours and answers by press release has just proved it.
The deeper winner is whoever builds the alternative, which mostly means Beijing and the open-weight ecosystem no off switch can reach.
The losers are the hundreds of millions of users and businesses that lost a tool wired into live operations overnight; Anthropic’s own foreign engineers locked out of the product they build; and the comfortable notion that American AI is a borderless utility, which just grew an asterisk every foreign buyer will now read closely.
There is a diplomatic cost too. This lands weeks before the first serious US-China dialogue on AI governance and before Beijing uses next month’s World AI Conference in Shanghai to launch a “World AI Cooperation Organization” and cast itself as the champion of open AI for the non-Western world. Chinese negotiators will now arrive holding a fresh, concrete example of Washington reaching into a private company to decide who may use a model, and they will press their standing argument: that American “AI safety” is strategic denial wearing a lab coat. Whether or not that is fair, it makes the already-slim odds of a shared US-China safety framework slimmer.
MAP/CHART OF THE WEEK
North Korea’s economy is apparently booming (credit WSJ):
BLIND SPOT
I was in New York City on Wednesday to speak at three different events, one for the shipping industry and two more focused on investment-related issues. I was hoping/expecting to glean a trenchant insight to share from my conversations there, but instead what I was struck by the most was the level of focus on the Iran-US MOU to end the war and open the Strait of Hormuz.
The US government is tired of the war and ready to wrap it up. Iran is ready to end hostilities as well. The country has suffered significant damage, and the longer the Strait stays closed, the greater the incentive for countries to find alternate sources of supply or for Saudi Arabia/the UAE to fast-track pipelines to get around the Strait. The only actor in the conflict that wants the war to continue is Israel—but as powerful as Israel is, it does not have the ability to dictate US action. The question there is whether Trump will brandish a stick to keep Israel in line (like withholding military aid) or offer the Netanyahu government a carrot.
I am not trying to say the issue is unimportant, and perhaps the length of the conflict and the muted reaction by oil markets despite the pervasive doom and gloom from analysts is making me blasé. (I’d note that Chevron’s CEO this week was singing a much different tune than two weeks ago when he suggested energy shortages were nigh in North America, instead noting that the US would be fine till Labor Day.)
But this is the blind spot, where I want to call out the myopic focus on today’s headlines. The world is shifting from unipolarity to multipolarity. Global energy markets are experiencing huge changes. The AI story is evolving on a weekly basis at this point. Huge advances in biotech, drone warfare, and robotics are rewriting what is possible. It is ironically comforting to seek refuge in the chaos of the day. But by doing so, we miss the chance to take a step back and ask ourselves, “What does the world look like in 2028?”
Read the full article here