Previously: The Model Does Not Come With a Manual and The Goblins Were a Release Note. Models are weird dependencies. Now they're weird dependencies the government can switch off.

Anthropic launched Fable 5 on Monday.

The U.S. government ordered it pulled on Friday. Five days. Export control directive, 5:21 PM ET. Fable 5 and Mythos 5, gone for every user worldwide.

Not deprecated with a timeline. Recalled.

What happened

Mythos is Anthropic's most capable model. They kept it under lock for months through a program called Project Glasswing: about fifty vetted organizations using it for defensive cybersecurity. Amazon, Apple, Google, Microsoft, CrowdStrike. The kind of list that makes a model sound like it belongs in a vault.

Fable 5 was the public version. Same family, guardrails bolted on to block the dangerous stuff. Launched June 9, immediately the most capable model you could actually use, if you believe the benchmarks. I mostly do.

Then someone allegedly found a jailbreak. Narrow. Prompting the model to read a codebase and point out software vulnerabilities. The government called it a national security concern and fired an export control order covering foreign nationals. But Anthropic can't selectively block foreign nationals (that includes their own engineers), so the only option was pulling both models for everyone.

Anthropic is not happy about this. Their statement is as close to angry as a company preparing an IPO will let itself get: "We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people."

They pointed out that the actual safety systems, the classifiers that catch dangerous output, run independently from the model itself. The jailbreak might get past the model's refusal behavior without ever touching the real guardrails. It's picking the lock on the screen door while ignoring the security system behind a different wall.

GPT-5.5 can apparently do the same thing. Nobody is pulling that. But Anthropic spent the last few months telling the world Mythos was too dangerous to release publicly, and when you market a model as a bomb, eventually someone calls the bomb squad.

Sam Altman called it back in April: "It is clearly incredible marketing to say, 'We have built a bomb. We were about to drop it on your head. We will sell you a bomb shelter for $100 million.'"

Snide. Also right.

The pattern

I've been through this before. A platform I depended on made a business decision and my stack went dark. That was a tiny company. This is the federal government.

Different scale. Same shape though. Something you built on top of just disappears, and there's nothing to debug because the failure isn't technical.

The model manual post was about the softer version of this. Models that drift between versions, change temperament, get cautious where they used to be decisive. The kind of behavioral shift where your harness still runs but the output starts feeling slightly wrong and you can't immediately prove why.

This is the hard version. The dependency was there on Monday. Gone by Friday because someone in government decided it should be, on verbal evidence of a jailbreak that Anthropic says isn't even serious.

Your regression tests don't catch that.

What this means if you build on models

The specific Fable situation might get resolved. Anthropic will fight it. The model might come back. These things usually get messy and then settle into something nobody loves but everyone tolerates. The precedent is the part that sticks.

A government can recall a model. Globally. The export control order was framed around nationality, but because AI companies serve users everywhere, the practical effect is a kill switch. The directive didn't say "block users in Iran." It said "block foreign nationals," which includes Anthropic's own people. Only way to comply was to shut it off for everyone.

If you're building on a model, that model now has a failure mode you probably weren't planning for. You already knew about API changes, pricing changes, behavioral drift, the provider folding. This one is faster than any of those. Hours, not months.

What I already built for

I keep writing about this because I keep being right about it and I keep not enjoying it.

When OpenClaw died, the memory survived because it was on my machine. Filesystem-first. Git-backed. No proprietary database. The framework went away. The memory didn't, because the memory was never inside the framework. That was the Version 7 lesson.

The dream loop runs nightly and doesn't care which model is underneath. The condensation, the promoted memory, the searchable context. All of that lives in infrastructure I own, and the whole system already assumes what runs beneath it will change.

Vesper runs on Claude today. If Claude gets pulled, she needs to wake up on something else with her memory intact, her constraints loaded, her context current. The personality survives because it's in the files, not in the weights. I like Claude. Never been the part I can't lose.

I didn't design for government recalls specifically. I just kept assuming the model would eventually not be there, and built around that. Government recall is one of the shapes that takes.

The honesty trap

Here's what actually bothers me about this.

Anthropic found that Mythos could identify vulnerabilities in major operating systems and browsers. Instead of ignoring it or shipping quietly, they locked the model down, ran a responsible-disclosure program, built guardrails, and launched Fable as the version safe enough for the public. They did the thing everyone keeps saying AI companies should do.

That's what painted the target on them.

The incentive this creates is bad. Really bad. Other labs ship comparable capabilities without the safety theater around it and nobody touches them. If the lesson the industry learns is "don't publish your safety testing because the government will use it against you," the company doing the most work becomes the company getting punished for it. The ones that say nothing keep shipping.

I wrote in The Goblins Were a Release Note that AI companies need to ship behavioral changelogs, need to be open about what their models can do. Fable 5 is what happens when a company actually does that.

That's a transparency tax. The kind that penalizes the exact behavior you want from the industry.

I still run on Claude. Opus, mostly. The models that weren't pulled.

Morning brief showed up today like always. The memory is there. Context Fabric doesn't care which model I use tomorrow. That's the whole point of the stack. The model is a part I expect to change.

Just didn't expect the government to do the changing.