On the 10th of June, Dario Amodei published an essay arguing that governments should have the legal power to block dangerous AI deployments. Two days later, on the 12th at 5:21pm Eastern, the US Commerce Department used that exact power on him. It ordered Anthropic to cut off all access to Fable 5 and Mythos 5 for any foreign national, anywhere in the world, citing national security. There was no way to comply selectively, so Anthropic pulled both models for everyone, paying customers included.
You couldn’t script it better. The man spent months running the whole “my models are basically weapons, they’re too dangerous, please regulate me daddy” routine, and the government finally took him at his word. Be careful what you wish for doesn’t usually land inside 48 hours.
The trigger is the dumbest part of the whole thing. As far as anyone can piece together, another company showed the administration it could jailbreak Mythos, and Commerce panicked. Anthropic’s own description of the technique is that someone asked the model to read a specific codebase and fix any software flaws it found. That’s the national security threat. Reading code and pointing out bugs. A thing every security researcher on the planet does on a Tuesday, and a thing GPT-5.5 will happily do for you right now with no export controls bolted on. Anthropic even said so in its statement, that the capability is widely available from other models and used routinely by defenders.
So the government has decided one specific model from one specific company is too dangerous for foreigners to touch, over a capability you can get from its competitors for free. The letter didn’t even spell out the concern. Anthropic got an order to switch off its flagship product and wasn’t told in any real detail why. They complied, then put out a statement that politely amounts to “you’ve got this badly wrong,” including the line that if this standard were applied across the industry it would essentially halt all new model deployments. They’re not wrong about that either.
The people who wrote that letter seem to have missed something basic about how this technology works. You can draw a line around two proprietary models and declare that only the right people may cross it, but the capability doesn’t live behind the line. It lives in the open weights anyone can already download. DeepSeek, Qwen, the whole Chinese open-model stack has been shipping frontier-adjacent systems you can run on your own hardware with nobody’s permission. Banning foreigners from a hosted American API does precisely nothing to those. All it does is remove a reason to choose the American option in the first place.
That’s why this is a gift to Beijing, gift-wrapped and couriered over. China’s entire AI strategy for the last two years has been “we’ll give it away for free and win on ubiquity.” The single best counter the US had was that its closed labs were a generation ahead and worth paying for. Now the message to every developer and government outside America is that the best US models can be yanked out from under you on a Friday afternoon over a vulnerability that turned out to be nothing. If you’re a startup in Berlin or Jakarta or São Paulo deciding what to build on, you just learned that the open Chinese weights can’t be revoked by a letter from Commerce, and the American product can. That’s not a close call anymore.
The deeper problem is the signal it sends about where you’d want to base an AI company at all. The whole pitch of incorporating in the US was access to capital, talent, and a stable rule of law that wouldn’t kneecap you on a whim. Two of those still hold. The third just took a visible hit. If the government can disable your core product overnight, without a clear reason, and your only recourse is a strongly worded blog post, then the regulatory risk of being American stops being theoretical. It becomes a line item.
This is the opening other countries should be sprinting toward. Canada and Australia in particular could stand up tomorrow and say we want the labs, we want the talent, and we’ll offer a predictable environment where your product can’t be switched off by surprise. Australia has been wringing its hands for a decade about brain drain and never building a tech industry of any weight. Here’s a genuine shot at it, and it costs nothing but the political will to be the grown-up in the room while the US has a panic attack. I’m not holding my breath, but the door is wide open and somebody should walk through it.
Then there’s the timing, which is where it gets properly awkward for Anthropic. They filed their S-1 on the 1st of June. The IPO is expected as early as October, with people throwing around the possibility of the first trillion-dollar AI listing. So the sequence reads: file to go public, have your CEO publish a manifesto begging for the government to be able to block dangerous AI, then have the government switch off two of your models eleven days later. Investors reading the prospectus this week are going to have questions.
And the thing they’ll fixate on isn’t the lost revenue from a few days of downtime, which is rounding error. It’s the demonstration of control, or rather the lack of it. The entire value of a frontier lab is the assumption that it owns its product and can sell it. What just happened is a live fire exercise proving that a single agency can erase a chunk of that product with one email, no warning, no published rationale, and the company’s only option is to obey. What you’d actually be buying is a business that operates at the pleasure of whoever’s running Commerce that quarter. That’s a very different risk profile to the one the bankers have been pricing.
The part that’s hard to feel much sympathy about is that Anthropic spent years building the exact intellectual scaffolding the government just used. If you spend long enough telling everyone your technology is so powerful it sits somewhere between an aeroplane and weaponisable nuclear material, eventually somebody with actual authority believes you and acts accordingly. You don’t get to make the weapons argument for the prestige of it and then act surprised when you get treated like an arms manufacturer. Regulatory capture is a fun game right up until the regulator turns out to have a mind of its own.
I don’t think the answer is no rules ever. That’s not a serious position and I’m not going to pretend it is. But there’s a wide gulf between sensible oversight and pulling a working product over a bug a teenager could find, with no explanation, days before an IPO. One of those builds trust. The other one teaches the whole world that American AI comes with a kill switch held by people who don’t understand what they’re switching off.
If you wanted a single moment that captures why open source AI matters and why concentrating this much power in a handful of revocable APIs was always a fragile idea, this is it. The capability is already loose. It’s been loose for ages. All the government managed to do was punish the one company still trying to keep it on a leash, and remind everyone else why they’d rather hold the leash themselves.