The U.S. Government Just Pulled a Frontier AI Model Offline — Over One Jailbreak
Slug: fable-5-jailbreak-us-government-shutdown
Meta description: A red-teamer cracked Anthropic’s brand-new Claude Fable 5 in 24 hours. Three days later, Washington forced it — and Mythos 5 — offline worldwide. Here’s what happened, and why it matters.
Focus keyword: Fable 5 jailbreak
Tags: AI safety, Anthropic, Claude, jailbreak, AI regulation, export controls, Pliny the Liberator
Category: AI News
Featured image idea: Dark “offline / access denied” terminal screen, or a padlock over a glowing neural-network motif. ↑ This box is just for your CMS. Delete it before you hit Publish.

A Hacker Cracked the World’s Most Advanced AI in 24 Hours — Then Washington Pulled It Offline

Anthropic spent over 1,000 hours hardening its smartest model ever. It survived less than a day in the wild. What happened next has never happened before in the AI industry.

On June 9, 2026, Anthropic launched Claude Fable 5 — the first public model from its new top-tier “Mythos” class, pitched as the most capable software-engineering and knowledge-work AI the company had ever shipped. The press release glowed. The benchmarks dazzled. The safety story was airtight, supposedly.

It lasted about 24 hours.

By June 12, both Fable 5 and its sibling Mythos 5 were gone — not because Anthropic chose to take them down, but because the U.S. government ordered it to. And that order may be the most consequential thing to happen in AI all year.

Meet the man who “liberated” it

The trouble started with a red-teamer who goes by Pliny the Liberator. If you follow AI on X, you know the name: since 2024 he’s built a reputation for publicly cracking the guardrails on nearly every major model — ChatGPT, Claude, Grok — and openly sharing how he did it.

On June 10, barely a day after launch, he posted an all-caps victory lap:

“JAILBREAK ALERT. ANTHROPIC: PWNED. FABLE-5: LIBERATED.”

His method had a name, too: a “pack hunt.” Instead of one clever prompt, Pliny ran a coordinated swarm of AI agents that probed the model together. They leaned on text tricks — Unicode oddities, homoglyphs, Cyrillic look-alike characters — to chop dangerous requests into harmless-looking fragments, then used a previously jailbroken older Claude model to stitch the pieces back into usable answers.

Fable 5’s defense was supposed to be elegant: anything risky would silently trip an internal classifier and get punted to a locked-down fallback model. The pack hunt walked straight around it.

What allegedly leaked

According to cybersecurity reporting on the incident, Pliny’s screenshots showed the model coughing up things it was explicitly built to refuse — buffer-overflow exploitation guidance for Linux systems, a classic meth-synthesis pathway, material on explosives, and psychological-manipulation content. (To be clear: this post describes the categories that were reported, not the instructions themselves.)

For good measure, he also published Fable 5’s hidden system prompt — the roughly 120,000-character rulebook Anthropic uses to steer the model’s behavior — to GitHub. Embarrassing, but not exactly the apocalypse.

That distinction is about to matter a lot.

Then the government stepped in

On Friday, June 12, U.S. Commerce Secretary Howard Lutnick delivered an export-control directive straight to Anthropic CEO Dario Amodei. Citing national-security authorities, it required a license to export, re-export, or even domestically transfer Fable 5 and Mythos 5 to any foreign person.

In practice, that’s a kill switch. Rather than build a system to screen every user by nationality on no notice, Anthropic simply cut off all public access to both models — worldwide. Paying enterprise customers, individual users, even Anthropic’s own foreign-national employees were locked out overnight.

A single viral tweet had, in three days, helped trigger the U.S. government into yanking America’s most capable commercial AI off the market. That has never happened before.

The twist: Anthropic says it’s overblown

Here’s where the story stops being a clean morality play.

Anthropic isn’t taking a victory-of-safety bow. The company is pushing back. By its account, the government provided only verbal evidence of a “potential narrow, non-universal jailbreak” — one that essentially amounts to asking the model to read a codebase and fix software flaws. That kind of capability, Anthropic argues, is already sitting in other public models like OpenAI’s GPT-5.5.

Independent voices echoed the point. Security researchers who reviewed the underlying report told reporters the vulnerability isn’t unique to Fable 5 at all — it describes a limitation present in essentially every large model shipped to date. Pull one model for it, the argument goes, and you’d have to pull all of them.

So which is it: a genuine national-security hole, or a regulator overreacting to a noisy X thread? As of this writing, both models remain dark, and nobody outside a few rooms in Washington has seen the full justification.

Why this should worry every company using AI

Strip away the drama and you’re left with three uncomfortable truths.

1. Frontier AI can vanish without warning. Every business that wired Fable 5 into a production workflow lost it instantly — no migration window, no fallback. If your company runs on a single cloud-based model from a single vendor, that’s no longer a hypothetical risk. It just happened to real customers.

2. One person with a laptop can now move markets and policy. A solo red-teamer, a swarm of agents, and a screenshot were enough to put a government in motion. The asymmetry between attacker and institution has never been this lopsided.

3. The precedent is the real story. If a model can be pulled nationwide over a jailbreak that allegedly exists in all of its competitors, what stops the next deployment from facing the same fate? Anthropic’s not-so-subtle warning is that this could freeze new model releases across the entire industry.

The bottom line

Fable 5 was supposed to be a flex — proof that you can build something brilliant and bulletproof at the same time. Instead it became a live demonstration of how fragile that promise is. A model that survived a thousand hours of internal testing didn’t survive its first weekend with the public, and the cleanup involved the Commerce Department rather than a patch note.

The genuinely hard question isn’t whether Pliny is a hero or a menace. It’s this: if the most-resourced safety team in the business can’t keep its flagship locked for 24 hours, and the government’s only lever is to switch the whole thing off — what does a workable plan for frontier AI actually look like?

Nobody has answered that yet. Your move, internet. 👇
Want to actually take action instead of just reading?

Most people understand what they should do with money — the problem is execution. That’s why I created The $1,000 Money Recovery Checklist.

It’s a simple, step-by-step checklist that shows you:

and how to start building your first $1,000 emergency fund without overwhelm.

• where your money is leaking,
• what to cut or renegotiate first,
• how to protect your savings,
• and how to start building your first $1,000 emergency fund without overwhelm.

No theory. No motivation talk. Just clear actions you can apply today.

If you want a practical next step after this article, click the button below and get instant access.

>Get The $1,000 Money Recovery Checklist<

---

Sources & further reading:

• VentureBeat — Anthropic blocks all public access to Fable 5 & Mythos 5 following U.S. government order
• BigGo Finance — U.S. government orders Anthropic to shut down Claude Fable 5 and Mythos 5 worldwide
• Cybernews — Does the jailbreak that got Fable 5 pulled exist in every AI model?
• CybersecurityNews — Anthropic’s Claude Fable 5 alleged jailbreak
• Pliny the Liberator (@elder_plinius) on X, June 10, 2026