Anthropic’s Fable 5 Got Killed by the US Government at 5:21 PM — Over a “Jailbreak” That Was Just Asking It to Fix Bugs

Photo by Planet Volumes on Unsplash

A national-security directive pulled a model used by hundreds of millions offline overnight. The scary part isn’t the model — it’s that the kill switch worked on the first try.

At 5:21 PM Eastern on June 12, Anthropic received a letter from the US government. By the time most of America sat down for dinner, two of the most capable AI models on the planet — Claude Fable 5 and Mythos 5 — were gone. Not throttled. Not geofenced. Gone, for every customer on Earth.

The official reason was national security. The actual trigger, by Anthropic’s own account, was a demonstration of someone getting Fable 5 to read a codebase and fix the software flaws in it.

Read that twice. The capability that got a frontier model recalled was debugging.

What actually happened

The order was an export-control directive, citing national-security authorities. On paper, it only restricts access by “any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.”

But there is no real-time switch that cleanly separates foreign nationals from everyone else mid-request. So the “narrow” order had one blunt outcome: Anthropic had to disable Fable 5 and Mythos 5 for everyone, worldwide, to stay compliant. Every other Claude model still works.

Reporting indicates the directive came from Commerce Secretary Howard Lutnick, in a June 1 letter to CEO Dario Amodei, after another company claimed it had found a way to jailbreak Mythos.

We believe this is a misunderstanding and are working to restore access as soon as possible. — Anthropic

The “jailbreak” was asking the model to do its job

Here is where it gets thin. By Anthropic’s account, the government provided only verbal evidence of the jailbreak, and the technique “essentially consists of asking the model to read a specific codebase and fix any software flaws.”

Anthropic says the demonstration surfaced only a handful of previously known, minor vulnerabilities — the kind other public models, including OpenAI’s GPT-5.5, find every day without any bypass at all. That is not a weapon. That is what every security engineer now uses an LLM for.

And Fable 5 was not some unguarded release. Before launch, Anthropic ran a bug bounty and external red-teaming for over 1,000 hours and found no universal jailbreak. It blocks high-risk cyber, bio, and chem requests outright, and even forced a controversial 30-day data-retention policy on all Fable traffic specifically to hunt novel attacks.

The real story: the government found a kill switch

Strip away the jailbreak framing and here is what is left: the US government reached into a private company and switched off a globally deployed commercial product overnight, on thin, unwritten evidence.

Anthropic’s own rebuttal is the most important line in the whole episode:

If a narrow jailbreak justifies recalling a model deployed to hundreds of millions of people, it would essentially halt all new model deployments for all frontier model providers.

They are right. Every model has non-universal jailbreaks. If that is the bar, the bar can be cleared against anyone, anytime. What changed on June 12 is not that AI got more dangerous. It is that the mechanism to pull a model now exists, it has been used, and it worked on the first pull.

The legal scaffolding showed up two weeks earlier. On June 2, Trump signed an executive order creating a category of “covered frontier models” and a pre-release review window, with the NSA director alone empowered to make the designation. The hammer that fell on Fable 5 did not come from nowhere.

Why Anthropic — and why now

This did not land on a random lab. It landed on the one company that already told the Pentagon no.

Rewind to March. After negotiations broke down, Defense Secretary Pete Hegseth designated Anthropic a “supply chain risk” — the first time a US company has been publicly tagged under a procurement statute built to keep foreign sabotage out of military systems. The trigger: Amodei refused to let Claude be used for autonomous weapons or for surveillance of American citizens.

Anthropic sued. A judge openly questioned the government’s case — “that seems a pretty low bar” — and the company won at one level before losing its appeal in April. That fight is still live. And it is all happening while Anthropic is mid-IPO, having filed confidentially just weeks ago.

So the company that drew a public ethical red line with the military just had its flagship models yanked by that same government, over a debugging demo, while it is trying to go public. You do not need a conspiracy theory to notice that “national security” makes excellent leverage.

My take

I think Anthropic is mostly right here — and also walked straight into it.

They are right that this is a dangerous precedent. Due process for software should not be a verbal phone call and a same-day shutoff. If “someone might jailbreak it” is sufficient cause, no frontier model is safe from a political off-switch, and every lab now has to price in the regulatory mood of whoever happens to run the Commerce Department.

But Anthropic also handed over the script. They spent months warning that AI is becoming too dangerous to develop unchecked, pleaded for an industry-wide brake pedal, and then — days later — shipped a Mythos-class model to the general public at $50 per million output tokens. If you spend a year telling Washington your technology is a national-security-grade weapon, you cannot act shocked when Washington finally treats it like one.

The uncomfortable lesson for everyone building on these APIs: your access to a model is now partly a political variable. It can vanish at 5:21 PM on a Friday, and “we are working to restore it” is the only SLA that matters.

Fable 5 will probably come back. The kill switch is not going anywhere.

If this saved you a 5:21 PM scramble, you can buy me a coffee — it keeps the daily posts coming.

[embed]Abhishek Agarwal Full-stack mobile developer - native Android (Kotlin), iOS (Swift), and KMM. Clean architecture and performance aren't…buymeacoffee.com

I write daily on AI, Claude, and the agentic dev economy. Subscribe to my Substack for the deeper cuts:

[embed]noob/programmer | Abhishek Agarwal | Substack Contrarian takes on AI tools, shipped code, and the startup stack - from a mobile engineer who reads the diff before…noobprogrammer16.substack.com