On Friday evening, the federal government ordered Anthropic to shut off Claude Fable 5 and its sibling, Mythos 5. By Friday night, both were gone. Anthropic says the order landed at 5:21 p.m. Eastern, came from Commerce Secretary Howard Lutnick, and cited national security.
Here is the part that should bother you no matter where you land on AI. Fable 5 had been public for three days.
I have a stake in this one. I run a live experiment on this publication where two AI systems manage real money, and Claude is one of them. I also run a consulting firm where clients build real workflows on top of these tools. When a model that hundreds of millions of people were using on Friday morning is dark by Friday night, and the company that built it did not make that call, I pay attention. You should too.
What actually happened
The official action was an export control directive. It barred any foreign national from accessing Fable 5 and Mythos 5. Anthropic says it had no clean way to block only foreign nationals, so it pulled both models for everyone.
The worry underneath the order is a claimed jailbreak. Anthropic looked at it and pushed back hard. The company says the technique is narrow and already known, that it amounts to asking the model to read a codebase and find software flaws, and that other public models can do the same thing with no trick at all. It also says no one has found a true, broad jailbreak of Fable 5. Worth noting: Fable 5 was the careful version. It was Mythos with guardrails bolted on so the public could use it safely.
You can believe Anthropic or you can believe the government, and I will be honest that protecting the country is the government’s actual job. If a model truly handed a hostile actor a weapon, moving fast would be defensible. But here is the trouble. With no standard set in advance, none of us can tell the difference between a real emergency and a convenient one. That is not a side issue. That is the issue.
The timing is the argument
I came to AI as a skeptic. I spent 25 years in operations watching big technology promises land and break. What changed my mind was not a sales deck. It was testing the thing myself and seeing where it held up. So I am not anti-regulation. I want regulation. I want it badly.
I just want it before the launch, not after.
Think about how every other serious industry works. We do not let a drug reach pharmacy shelves and then yank it three days later because someone found a side effect nobody bothered to test for. We test first. We set the standard, we publish it, and everyone builds to it. The rules are boring. They are predictable. They are known in advance, and that predictability is the entire point. It protects the public and it lets builders build.
That is not what we have in AI. What we have is a model going live, a scramble, and a federal hand reaching in after the fact to switch it off using an export law written for something else entirely.
This is being run like an episode of The Dukes of Hazzard
Stay with me.
The federal government has had years to set clear, pre-launch rules for frontier AI. What has it produced? A December executive order aimed mostly at stopping the states from writing their own rules. A March policy framework that is a list of suggestions for a Congress that has not acted on it. Meanwhile California and Texas got tired of waiting and passed their own AI laws, which took effect January 1.
So the real federal record reads like this. No clear national standard. A running effort to block the states from filling the gap. And then, when a model the administration apparently did not care for went live, a sudden national-security maneuver to pull it off the road.
That is Hazzard County governance. No posted speed limit. Boss Hogg deciding after the fact who gets pulled over. The law applied when it is convenient and ignored when it is not. The General Lee flying through the air while everyone improvises the landing. It makes for good television. It is a terrible way to run the most consequential technology of our lifetime.
The question I cannot shake
I honestly cannot tell what is driving this. Maybe the federal government is just slow, and this is ordinary government drift. Maybe it is disinterest, AI treated as someone else’s problem. Or maybe, and this is the one that keeps me up, the vagueness is the strategy. If you never set firm rules, you keep total discretion. You decide case by case who wins and who goes dark. Discretion like that has a way of turning into private advantage for the people holding the switch.
I am not asserting corruption. I am saying the structure invites it, and a structure that invites it is a bad structure. Clear rules, set in advance and applied to everyone the same way, are the cure for exactly this. Their absence stopped being an accident a long time ago. It is a choice.
What I want, plainly
Set the safety standards before models ship, not after. Make them public so every builder knows the bar. Apply them to everyone, the large labs and the small ones, by the same measure. Let the federal government actually lead instead of mostly working to stop the states from leading. And stop reaching for laws written for missile parts to settle a software question, because that tool does not fit and everyone watching can see it does not fit.
Fable 5 may come back. The models will keep getting more capable. The thing that has to change is the thing that has not changed in years. Somebody in Washington has to write the rules of the road before the cars are already doing ninety down a county highway.
I will keep testing. I will keep documenting. And the next time a model I am using goes dark overnight, I would like to be able to point at the rule it broke. Right now I cannot, and that is the whole problem.
