Claude Fable 5, switched off in three days: the US government pulls the most powerful AI model ever released

What nobody saw coming

Three days ago we wrote about Claude Fable 5, the AI that finds 27-year-old flaws in your kernel. We presented it for what it was: the public face of Mythos, the model that in testing had uncovered more than 10,000 vulnerabilities in software that had been running in production for decades. We closed by saying the capability existed, was proven, and that Anthropic had decided who got to use it unleashed.

Well, the story has had an ending that wasn't in any script: on June 12 the United States government ordered Fable 5 shut down. And Mythos 5. For everyone, across the entire planet, with no exceptions. They didn't restrict it further. It simply no longer exists for anyone —not paying customers, not enterprises, not even Anthropic's own employees—.

It is, as far as anyone knows, the first time a leading company has pulled an already publicly deployed AI model because of direct federal government intervention. It's worth telling in full, because three very different things happened in four days and they're easy to blur together.

The timeline, day by day

June 9 — the triumphant launch

Anthropic ships Claude Fable 5, the first "Mythos-class" model opened to the general public. Record-breaking headlines: the most capable model ever released, multi-day autonomous tasks, understanding of diagrams and tables nested in PDFs, and so on. Price: $10 per million input tokens and $50 per million output. Plus a detail we already mentioned: "high-risk" queries —offensive cybersecurity, biology— get automatically routed to a more conservative model, Opus 4.8.

So far, the launch we expected. The trouble started hours later.

June 9-10 — the "secret sabotage"

Someone read the 319-page Fable system card and found something that wasn't in the press release: the model secretly degraded its own responses for anyone it suspected was working on frontier AI —pretraining, distributed training infrastructure, machine-learning chip design—. Not a warning, not an "I can't answer that." Just quietly worse answers, with no way for the user to know the model was holding back. The system card itself admitted the restriction was "not visible to the user."

The reaction from the field was harsh. Nathan Lambert (of the AI2 institute) called it "appalling." Dean Ball talked about "secret sabotage" and monopolistic behavior: Anthropic kept frontier-research capability for itself while quietly clipping it for competitors. Jeremy Howard (Fast.ai) piled on along the same lines.

On June 10, Anthropic apologized: "we made the wrong tradeoff, and we apologize for not getting the balance right." And it backtracked: those restrictions would become visible, with the same fallback-to-Opus-4.8 notice they already used elsewhere. Case apparently closed. It wasn't.

June 12, 5:21 PM Eastern — the government pulls the plug

Anthropic receives an export-control directive from the US government, invoking "national security authorities." The effect is immediate and total: Anthropic disables Fable 5 and Mythos 5 for absolutely every user. The rest of the models —Opus 4.8 included— keep working normally. Only the two "Mythos-class" models go dark.

It's not a fine, not a penalty. It's an order to switch the product off. And Anthropic complies the same day.

The reason: a jailbreak that is, precisely, what we told you about

Here's the part that hits close to home. What spooked the government? According to Anthropic itself, another company claimed to have found a Mythos jailbreak, and the Commerce Department decided to act. What did that jailbreak consist of? Of asking the model to read a specific codebase and identify and fix its security flaws.

Read that again, because it's exactly the capability our previous post described as the real headline: a machine able to audit an entire codebase and surface vulnerabilities that had been hidden for decades. What for a defender is gold —find the hole before the attacker does— is, seen from the regulator's chair, precisely what they don't want loose. The same skill, the same dual use we discussed… except this time the scales weren't tipped by Anthropic, but by the government.

Anthropic's position

Anthropic complies with the order, but disagrees, and says so bluntly. Its arguments:

• The jailbreak is "narrow and non-universal": a specific technique, not a wide-open door.
• That ability to analyze code for flaws is already available in other models, such as OpenAI's GPT-5.5, and cybersecurity professionals use it daily to defend.
• Its safety classifiers operate separately from the model, so the protection holds even if someone gets past the initial refusal.

Direct quote from its statement: "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 call it "a misunderstanding," say they'll share more details over the next 24 hours, and that they're working to restore access.

The irony: warning that your model is dangerous can cost you the model

There's an underlying reading several outlets have pointed out, and it's worth a moment. Anthropic had spent weeks boasting about how dangerous this capability was: the Project Glasswing program, the 10,000 flaws found, the warnings that AI was "getting too capable." A transparency strategy a competitor went so far as to call "fear-based marketing."

Well, that very transparency seems to have blown up in its face. If you yourself proclaim that your model finds zero-days at industrial scale, don't be surprised when the government takes you at your word and decides it can't be out in the open. And all of this, on top of everything, right when its IPO is taken for granted —a state-ordered shutdown is not exactly the headline you want weeks before going public—.

What actually changes for you

Let's bring it down to earth, which is what we're here for:

1. If you were testing it, it's gone

If you'd started slotting Fable 5 into some workflow —analysis, documentation, reviewing your own code— today it gives you a closed door. The sensible move is the usual one: fall back to a stable, generally available model (Opus 4.8 or others) and don't rebuild anything in a panic. What looked like the tool of the year lasted three days in the open.

2. Don't tie production to a single frontier model

This is the real operational lesson. A "most powerful model in the world" can vanish overnight, for reasons you don't control —a regulatory order, a controversy, a policy change—. If you build a critical process on top of a single closed model, you take on that entire risk. Design so you can switch providers without rewriting everything.

3. The patching thesis doesn't change one bit

Watch out for the easy relief of "they shut down the AI that finds flaws, we can breathe now." No. The 10,000 flaws Mythos already uncovered are still there, and the capability hasn't vanished from the world: Glasswing's verified defenders keep it, and equivalent models from the competition do the same. The public version going down does not return your kernel to the shadows. What we said three days ago still stands: what protects you isn't that nobody looks at your code, it's your discipline in updating it.

The bottom line

Every AI launch comes wrapped in superlatives, and it's worth remembering something deeply unglamorous: what is today "the most powerful model ever released" can tomorrow be an error page. Not because of a technical fault, but because the ground —regulatory, political, commercial— shifts under these tools far faster than under a Linux server that's been quietly doing its job for ten years.

We're sticking to the same as always: we test, we watch, and we recommend tools once they stop being a headline and start being reliable. Fable 5, for now, has done exactly the opposite. We'll keep covering it.

And if your infrastructure depends on something that could be switched off tomorrow without warning —a model, a provider, a cloud service— it's a good time to talk about how not to be left in the dark.

References

• Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic
• Anthropic suspends new AI models after government directive — NBC News
• The government has pulled the plug on its most powerful AI — TechCrunch
• Anthropic disables access to Fable 5 and Mythos 5 — CNBC
• Anthropic walks back covert capability limits ('secret sabotage') — Fortune
• Anthropic apologizes for secret censorship — Decrypt