The 90-minute ultimatum that took the world's most powerful AI offline
On Friday 12 June, Anthropic had 90 minutes to fix a security flaw in Claude Fable 5 or take it offline. By 10pm it was gone. It came back this morning. The 19 days in between are worth understanding.
Right, so Fable 5 is back. Well, back for us in the UK as of this morning anyway.
If you missed what happened over the last few weeks, here's the short version: Anthropic launched what is genuinely the most capable AI model ever released to the public, and within three days the US government gave them 90 minutes to either fix a security flaw or take it down. By 10pm on that Friday it was gone. Not just in the US. Everywhere. All of it.
It came back this morning, 19 days later, with a new security patch and what I imagine was a very tense couple of weeks for everyone at Anthropic HQ.
The story of what actually happened is, honestly, a bit mad.
How this kicked off
Fable 5 launched on 9 June and the benchmarks were properly impressive. Stripe said it compressed months of engineering work into days. It's the first model to break 90% on Anthropic's hardest reasoning benchmark. By all accounts, a genuine step up from anything that came before it.
Within two days, Amazon's research team had found a jailbreak. The technique, and this is the bit that made a lot of security people roll their eyes, was essentially: ask it to read a codebase and spot the flaws. That's it. That's the whole technique. Fable 5 is so good at code analysis that when you point it at certain types of code, it can surface information useful for an attack. Not ideal, but also not exactly a secret decoder ring.
Amazon CEO Andy Jassy happened to be on a call with White House officials about something completely unrelated when he brought it up. And then things moved very quickly in the way they do when governments get involved.
By Friday afternoon, Commerce Secretary Howard Lutnick had sent Anthropic a letter: fix the jailbreak or pull both Fable 5 and Mythos 5 from the market. The deadline was 90 minutes. The letter arrived after 5pm. By 10pm, both models were offline globally.
The Amazon thing
I keep coming back to this because you genuinely couldn't write it.
Amazon is Anthropic's largest investor. Amazon Web Services is the cloud infrastructure that runs Anthropic's models. And it was Amazon's CEO who made the phone call that started this whole chain of events.
Investor. Cloud host. The one who triggered the shutdown.
While Fable 5 was offline for 19 days, AWS Bedrock was still running. Competing models were still available. Enterprise customers who needed to keep working went and found alternatives, some of them on the same infrastructure. I'm not saying anything. I'm just saying.
Anthropic's side of it
To be fair to Anthropic, they pushed back pretty hard. Their public statement made three points: the jailbreak technique was narrow and specific, not some fundamental weakness in the model; the same technique worked on GPT-5.5; and if "no jailbreaks ever" is the bar before you can ship, no model from any lab would ever see the light of day, because that's simply not how any of this works.
White House AI adviser David Sacks posted that Dario Amodei had simply "refused" to fix it. Anthropic's account was that they got the letter at 5:21pm with no specific technical details, had 90 minutes, and were expected to assess and patch a complex security issue before most people had finished their commute home. The two versions don't quite line up, which probably tells you something, though I'll leave you to work out what.
Why this matters beyond the drama
Here's the part that should probably end up in your risk planning, if you're the sort of person who has risk planning.
For 19 days, anything businesses had built on top of Fable 5 or Mythos 5 just stopped. No warning, no migration window, no SLA that covered it. A Baker McKenzie partner came out and said companies should start treating "AI model unavailability" as a category in their business continuity planning. Which, after this, seems entirely reasonable.
A French politician put it bluntly: "A nation that depends on others for its technology is a nation that can be unplugged overnight." Swap "nation" for "business" and that's a conversation a lot of IT leaders are going to be having.
Nearly 80 cybersecurity professionals signed a statement saying the whole thing was counterproductive, that it had taken the best models away from defenders while doing very little about actual risk. Meanwhile DeepSeek raised $7.4 billion during the blackout and Chinese AI labs quietly had a very good three weeks. So that worked out well.
So where are we
Fable 5 and Mythos 5 are back as of this morning. Anthropic agreed to a new classifier that blocks the jailbreak in over 99% of cases, plus pre-release government access to frontier models going forward and some joint research arrangements. Independent researchers who've looked at the new safeguards are calling them "extraordinarily strong," which is reassuring.
Is Fable 5 worth using now it's back? By all accounts, absolutely yes.
Was the 19-day shutdown the right call? I genuinely don't know. But I do know that "your AI infrastructure can disappear with 90 minutes' notice on a Friday afternoon" is something worth having a plan for.
Just maybe don't make that plan on a Friday.
