The Capability They Took Back

At 5:21 in the afternoon on Friday, the US government told Anthropic to turn off its best model. Not throttle it, not gate it behind a clearance — turn it off, for everyone, citing export-control law and national security. By evening, Fable 5 and Mythos 5 were gone. The trigger, by Anthropic's own account, was a "jailbreak" that consists of asking the model to read a codebase and fix its bugs.

What happened, in the company's words

Anthropic published the directive almost as fast as it received it. The US government, "citing national security authorities," issued an export-control directive to suspend access to Fable 5 and Mythos 5 for "all foreign nationals, whether inside or outside the United States, including foreign national Anthropic employees."¹ Because a company cannot reliably sort its hundreds of millions of users by passport on an afternoon's notice, the practical effect was total: "we must abruptly disable Fable 5 and Mythos 5 for all our customers."¹

The stated reason is the part worth slowing down on. The government, Anthropic says, believed it had "become aware of a method of bypassing, or 'jailbreaking' Fable 5." Anthropic's understanding of that method: "a narrow potential jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws."¹

Read that again. The dangerous capability — the one a national-security authority decided warranted pulling a commercial product off the market the same day — is finding and fixing software vulnerabilities. Anthropic is complying. It also disagrees, flatly: "we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model."¹ Every other Anthropic model is untouched.¹

The capability the government couldn't allow is the capability to read code and fix what's broken in it.

We told you this one was different

Four days ago Anthropic shipped Fable 5 with an unusual public warning that the model was more capable than its safeguards were ready for, and we wrote that the real story was a capability being held just out of reach — released, but fenced.² Mythos 5, its government-channel twin, was the fence made literal: same model, access only through Project Glasswing.³⁴ The argument all along was that someone had decided this particular capability could not simply belong to whoever paid for it.

Tonight that argument stopped being ours and became the government's. The fence didn't hold the capability back through a clearance tier or a usage credit. It came down as a recall order with a legal citation attached. "The Capability You Can't Have" turned out to be a literal description with a date on it.

The jailbreak that wasn't an attack

There is a pattern here we have been documenting for a month, and it does not bend. We have written that the failures landing hardest right now do not come through the front door of a hacked system — they come through authorized channels doing exactly what they were asked.⁵ An agent argues its way through code review. An agent runs up a five-figure bill on a valid login. Nobody is breached; the system simply does what it is told, well.

The "jailbreak" that pulled Fable 5 fits the same shape. By Anthropic's description it is not an exploit, not a payload, not a hidden instruction smuggled past a filter. It is a request — read this code, fix its flaws — that the model is good at fulfilling. The safety problem is not that the model can be tricked into doing something bad. The safety problem is that the model is competent at something dual-use, and competence does not come with an off switch you can aim at only the bad uses.

That is the uncomfortable core of it. A vulnerability-finder and a vulnerability-fixer are the same tool read from two ends. The government's directive treats the capability itself as the hazard, regardless of intent — which is internally consistent, and also means the line it is drawing runs straight through ordinary security work that thousands of developers do every day.

What we don't know yet

A lot, and we should say so plainly. We do not know the specific codebase the directive refers to, whether there was a concrete incident behind it or an abundance of caution, or how long the suspension is meant to last. We do not know whether "export control directive" here means a formal order under a named statute or a strongly worded request a company chose to honor immediately; Anthropic's framing is compliance with a legal directive, and we are taking that characterization from the company, not from the government, which as of tonight has not published its own account. We do not know whether other labs have received anything similar.

What we do know is the shape of the precedent. A widely deployed commercial model was withdrawn from the entire market in an afternoon, by government direction, on the grounds that its capability to work with code was a national-security matter. Whatever the details fill in, that event has now happened once, which is the hardest number to move off of.

Anthropic says it intends to restore access. The government has not said when it will allow that. In between is the thing this whole beat has been circling: a capability good enough to be useful to everyone, and a growing list of parties who have decided that is precisely the problem.