Anthropic launched Fable on June 9. Three days later, on Friday, June 12, the US government told the company to shut it down for a big chunk of its users.
By the early hours of June 13, the model was dark.
Not just Fable, either. Anthropic also pulled Mythos 5, its top-shelf cybersecurity model, the one normally walled off to a small set of Project Glasswing partners. Every other model the company runs stayed online. These two did not.
The order was specific in a way that's genuinely hard to enforce: no foreign nationals, full stop. Inside the US or outside it. Customers or not. Even Anthropic's own employees who aren't American citizens lost their keys. The government's stated reason was national security, as Engadget reported, though officials didn't bother spelling out what the actual threat was.
Anthropic has a theory. It thinks Washington caught wind of a way to jailbreak Fable 5.
What the government heard
Here's what Anthropic says it knows, which isn't much. Officials handed over verbal evidence pointing to a single narrow jailbreak, the kind tailored to one model rather than a universal skeleton key. Some unnamed party flagged it. That tip, apparently, was enough.
The company promised more detail within a day, and it drew a line while it was at it. A theoretical jailbreak, in its view, shouldn't be grounds for yanking a commercial product off the shelf.
That's a notable position for Anthropic to take. This is the company that has spent years arguing someone should be watching the AI industry closely. It's been one of the louder voices asking for guardrails. So watching it object to a guardrail landing on its own model is a bit like the homeowner who lobbied for a fire code grumbling when the inspector shows up unannounced.
Anthropic's actual gripe, to be fair, is about process, not principle. The government should be able to block dangerous deployments, the company wrote, but only through rules that are clear, fair, and grounded in technical facts. This action skipped all of that.
Why a Pokémon score matters here
Fable was meant to be the friendly version of Mythos, bringing a lot of that cybersecurity muscle to a wider crowd. Anthropic says its capabilities go beyond anything it had shipped before, and it reached for an oddly charming benchmark to prove it: the model beat Pokémon FireRed in testing. Claude, the company's older system, couldn't even finish Pokémon Red, the game FireRed remakes.
Silly as that sounds, beating a game start to finish is a real test of long-horizon planning. A model that can grind through hours of a structured world without losing the thread can grind through other long, structured tasks too.
Some of which are a lot less wholesome than catching a Charmander.
That's the rub with a powerful cybersecurity model. The same skills that defend a network can map one for attack. Anthropic insists it bolted on heavy safeguards specifically to keep Fable from being pointed at offensive work, and says those restrictions were aggressive enough that plenty of users complained about getting blocked from harmless requests.
The company is also blunt about the ceiling on all this. No provider can promise a model is jailbreak-proof. Every system has a crack somewhere if you push hard enough. Anthropic's strategy was to make the cracks either too narrow to be useful or too expensive to find, then watch the traffic closely enough to slam the door on anything that slipped through.
The national security angle reframes the whole thing. A standard jailbreak story is a company patching a bug. This is the federal government reaching into a private product and deciding, on its own timeline, who gets to use it based on citizenship.
What's worth watching now is whether this is a one-off or a template. If a verbal tip about a single narrow exploit can pull a freshly launched model for an entire class of users, every company shipping frontier systems is now living under a kill switch it doesn't hold.
Anthropic asked for oversight. It may have just learned what oversight feels like when you're the one being overseen.