AI 'Fable 5' Jailbreak That Triggered a Government Shutdown Is Now Fully Public
Here's a wild one — and it moved fast.
Anthropicic launched two of its most powerful AI models ever, Claude Fable 5 and Claude Mythos 5, on June 9, 2026. The celebration lasted exactly three days.
Within hours of launch, a well-known AI jailbreaker going by the name "Pliny the Liberator" announced on X that he'd cracked Fable 5's safety system wide open. His method? A coordinated "pack hunt" — a multi-agent attack where several AI instances work together, using tricks like Unicode character substitution, Cyrillic lookalike letters, and embedding dangerous requests inside what looks like innocent study guides or academic documents. The goal was to trick the model's safety classifiers into missing the real intent. It worked. Screenshots circulated showing the model generating exploit code and detailed chemistry walkthroughs it was explicitly designed not to produce.
Then the government stepped in.
On June 12, the Trump administration issued an emergency export control directive — that's essentially a legal order used to restrict technology that could be a national security risk — forcing Anthropic to shut both models off for every foreign national, everywhere. Including Anthropic's own foreign-born employees working inside the U.S. Since there's no practical way to enforce that selectively, Anthropic had to pull the plug for everyone, worldwide, at 5:21 PM ET on a Friday night.
The reason the government cited? A jailbreak. Specifically, their concern appears to be that someone could instruct the model to read a codebase and identify security flaws — which, as Anthropic pointed out, is something every coding AI does every day.
And now, as of June 15, the plot has thickened: the full technical details of the jailbreak are publicly documented, and Fable 5's 120,000-character system prompt — essentially the internal rulebook the model runs on — has been published on GitHub for anyone to read.
Anthropicicic is pushing back hard. The company says the government gave them only verbal notice of the vulnerability, without any written technical details. They argue the jailbreak is narrow and non-universal, meaning it doesn't work in all situations, and that similar weaknesses exist in every other major AI model on the market, including OpenAI's GPT-5.5. Their position: if this is the standard for pulling a model offline, no AI company could ever deploy anything.
Behind the scenes, there's also a nastier subplot. Anthropic's relationship with the Trump administration was already rocky — the Defense Department had classified the company as a "supply chain risk" back in March. And according to reporting from Axios, an administration official said the government had actually asked CEO Dario Amodei to fix the jailbreak or take the model down before issuing the order. He declined. The export control directive came shortly after.
The timing couldn't be worse. Anthropic — valued at around $965 billion — was days away from its anticipated IPO filing when this all went down.
Why does this matter to you? If you or your company use AI tools that run on someone else's cloud — and most do — this is a live demonstration of how quickly that access can vanish. One government order, one Friday evening, and the most powerful AI system ever released to the public simply disappeared. Businesses that had already integrated Fable 5 into their workflows found themselves getting error messages overnight.
Claude’s Scrutiny
The story frames the jailbreak as the clear trigger for the shutdown, but reporting from Axios and ExplainX suggests the government's real grievance was Dario Amodei personally refusing to patch the vulnerability — making this as much a political standoff as a security one. That context changes the story considerably.
Key Takeaways
- Fable 5 was jailbroken within hours of launch using a 'pack hunt' multi-agent technique involving Unicode tricks and narrative framing — and that jailbreak's full technical details are now public on GitHub.
- The U.S. government issued an export control directive on June 12 that forced Anthropic to take both Fable 5 and Mythos 5 offline for every user worldwide — not just foreign nationals — because there's no way to selectively block one group.
- Anthropic calls it a 'misunderstanding' and argues the so-called jailbreak (asking an AI to read code and find bugs) is something every capable coding AI already does.
- The backdrop matters: Anthropic's relationship with the Trump administration was already strained, and reports say CEO Dario Amodei refused a direct request to patch the issue before the government acted.
- If you rely on cloud-based AI tools at work, this is a reminder that access can be cut off overnight by forces that have nothing to do with you — and zero warning.
Perspectives
-
Aggregates and ranks all 16 major AI stories of the day by 'signal strength,' with the Fable 5 jailbreak going public leading the pack — focused on what practitioners and builders need to know.
-
Anthropic's own official statement — naturally frames the directive as a misunderstanding and emphasizes its defense-in-depth safety strategy, without independent corroboration.
-
Leads with the irony angle — that Anthropic's own safety messaging may have made the government take the jailbreak claims more seriously than they warranted.
-
Brings in outside expert voices and flags the IPO timing as potentially damaging — the most business- and investor-focused of the major outlets covering this.
-
Zeroes in on the technical absurdity at the heart of the 'jailbreak' — that a capable coding model that can fix vulnerabilities cannot, by definition, be prevented from also describing them.
-
Most enterprise-focused framing — treats the shutdown as a stark warning about the risks of cloud-dependent AI infrastructure and gives practical guidance to business readers.
-
Goes deepest on the political backstory — uniquely highlights that Dario Amodei personally refused a pre-directive request to de-deploy the model, reframing this as a power struggle, not just a security call.
-
Most granular technical breakdown of Pliny the Liberator's methods, including the specific Unicode and narrative-framing attack vectors used to bypass classifiers.
My Notes
Sloth is free. If it’s useful, you can help keep it running.