Artificial intelligence (AI)

The World’s Most Powerful AI Lasted Three Days.

Then the Government Switched It Off.

No outage. No hack. Just a letter that went out on a Friday evening, and the best AI on the planet went dark for every single person on Earth. The reason why should unsettle you more than the shutdown itself.

You can read this here for free :)

The World’s Most Powerful AI Lasted Three Days.

Picture launching the best thing your company has ever made, watching the whole world fall in love with it, and then having a government letter quietly switch it off three days later.

That is exactly what just happened. Last Tuesday, Anthropic released its most powerful public model yet, a system called Fable 5 that topped almost every benchmark in existence. It was, by a clear margin, the most capable AI the public had ever been handed. Then, at 5:21 in the evening on a Friday, a letter arrived from the US government, and Fable 5 went dark for every user on the planet. Not a regional block. Not a technical outage. Gone, everywhere, all at once, including for the very engineers who had just built it.

So what made this model so terrifyingly dangerous that the government had to step in within seventy two hours of launch? The honest answer, once you dig in, is probably “not much at all.” And that is the part of this story that should genuinely worry you. Let me walk you through what really happened.

First, what the model actually was

To understand the shutdown, you need to understand the model, and it is simpler than the headlines make it sound.

Anthropic builds a special top tier of AI that it calls its Mythos class, a clear notch above its normal lineup. The newest one came in two flavours. The first, called Mythos 5, is the raw, unfiltered model, handed out only to a small group of carefully vetted cyber defence partners. The second, called Fable 5, is that exact same powerful brain, but with safety filters bolted on top, and it was the one sold to everybody.

Those filters were the clever part. When you asked Fable something sensitive, say about cybersecurity or biology, it would quietly hand your question off to a weaker, safer model instead of answering with its full power. Same brilliant brain, but wearing padded gloves. Fable, the careful gloved version, is the one that got switched off.

The World’s Most Powerful AI Lasted Three Days.

The thing that triggered it

The shutdown was set off by what people in this world call a jailbreak, which simply means someone found a way to trick the model past its own safety filters.

Here is the messy part. According to reporting from outlets like Axios and CNBC, it was not a government lab that found this. It was an outside company that claimed it had found a way around Fable’s guardrails and flagged it. The Commerce Department acted on that claim. And by Anthropic’s own account, the letter that arrived never even spelled out clearly what the specific concern actually was.

So Anthropic now finds itself in a strange position, obeying the order and openly disagreeing with it at the same time. And the criticism is not only coming from inside the company. One former government official described the move as cartoonish, pointing out the bizarre contradiction that the same government is happy to sell advanced AI chips to China while slamming the door on Britain and every other close ally getting access to the best American models.

The one detail the whole argument turns on

Strip away all the scary national security language and this entire story comes down to a single, surprisingly simple distinction.

When it comes to a security weakness in software, an AI can do one of two very different things. It can find the weakness, which just means reading through code and noticing a weak spot, the digital equivalent of pointing at a door and saying that lock looks flimsy. Or it can exploit the weakness, which means actually breaking through the door and doing real damage. Those are worlds apart.

The first one, simply spotting flaws, is something every coding assistant on the market does all day long for millions of people. The second one, actually attacking, is the genuinely dangerous skill, and it is the exact thing Fable’s filters were built to block.

The World’s Most Powerful AI Lasted Three Days.

Here is the kicker. The jailbreak that got this model pulled was the harmless kind. By Anthropic’s own description, it amounted to asking the model to read some code and point out its flaws, which is precisely what people use ordinary tools like ChatGPT for every single day. Anthropic even pointed out that the very same ability sits inside its main rival’s model, the one that security teams everywhere quietly rely on to defend their own systems.

So the blunt question underneath all of this is uncomfortable. Did a government just switch off a model used by hundreds of millions of people over the completely harmless half of a scary sounding skill?

My honest take

This is an own goal, and the painful truth is that Anthropic teed it up itself.

To be fair, the company’s safety work here was not fake at all. It stress tested this model for thousands of hours alongside the US government and others, nobody managed to find a master key that broke it wide open, and it even took on real costs and privacy tradeoffs so it could catch genuine attacks. And the underlying fear, that a frontier AI might one day hand a bad actor a working cyber weapon, is a completely reasonable thing to lose sleep over.

But that fear is not what got this model pulled. What got it pulled was a model reading code and pointing out the weak spots, which is closer to a spellchecker than to a weapon, and which competing tools do for ordinary people all day. Anthropic admitted as much itself when it held up its rival to prove the capability is everywhere. And that admission is the entire problem. You do not get to spend a year marketing how powerful and dangerous your model is, and then act shocked and call it a harmless commodity the very week that reputation finally costs you something.

The result is that nobody actually wins. Anthropic’s own foreign employees are locked out of the tool they built. Allied governments lost a system they were genuinely using to help defend things like their power grids. And the bad actors the controls were supposedly aimed at were quietly working on completely different models the entire time.

The World’s Most Powerful AI Lasted Three Days.

The part of all this that will outlast the news cycle is two simple words: export control. Access to an AI model is now something you can be granted based on your nationality and stripped of by a single letter, the way a government would treat a physical missile component. Every AI lab watching this unfold just learned a quiet, chilling lesson. Keep your very best model somewhere a regulator cannot reach it, and hand the public a watered down version with a friendlier name on the box.

Fable 5 might even be back online by the time you read this. But that precedent, the idea that the most powerful tools of this era can simply be switched off by decree, is not going anywhere. And that is the thing we should all be paying attention to, long after this particular model fades from the headlines.

So I want to put the question to you directly, because I keep going back and forth on it myself. Should any government be able to reach in and switch off a piece of software that millions of people are using, even when the danger is this blurry? Tell me where you land in the comments, because I genuinely cannot decide.

Give Me A Follow

If this helped you make sense of a story that most coverage buried under jargon, please take a second to clap and leave a comment below. On Medium, claps and comments are the signal that tells the algorithm a piece is worth showing to more people, so even a quick tap helps this reach the next person trying to understand what just happened. Thank you for reading.