The US Government Just Shut Down the Most Powerful AI | EP 1

00:00:00:10 - 00:00:07:03

On Friday afternoon at 521 eastern, an American AI company got an email from the United States government.

00:00:07:03 - 00:00:12:17

90 minutes later, the most powerful AI model ever released to the public was gone.

00:00:12:18 - 00:00:15:01

Switched off for everyone.

00:00:15:03 - 00:00:21:23

Not because it broke the law, not because it hurt anyone, but because the government decided it was a national security weapon. And

00:00:22:03 - 00:00:27:00

one of the biggest tech companies on earth may have been the one that pulled the trigger.

00:00:27:02 - 00:00:39:16

This is the story of mythos fable, a jailbreak and a snitch. And by the end of this video, you're going to understand why every AI company in America is watching their back right now.

00:00:39:22 - 00:00:51:07

Welcome back to Super Confident AI, the channel where we take the biggest, most confusing stories in artificial intelligence and break them down so they actually make sense. No PhD required.

00:00:51:07 - 00:01:03:03

I'm going to walk you through exactly what happened. Who did what. Why it matters for you as a regular person who uses these tools, and what it means for the entire AI industry going forward.

00:01:03:06 - 00:01:15:04

Because this is not a small story. This might be the moment the rules of the game changed permanently. Let's get into it. So let's set the stage. The company is

00:01:15:06 - 00:01:17:11

anthropic. The makers of Claude,

00:01:17:15 - 00:01:27:02

one of the top AI labs in the world. And honestly, the lab with the strongest reputation for being the safety first good guys in Washington.

00:01:27:04 - 00:01:32:08

On June 9th, they launched two new models, fable five and mythos five.

00:01:32:13 - 00:01:38:02

Now, here's the key thing to understand. These two models share the same underlying brain.

00:01:38:05 - 00:01:40:08

The difference is the guardrails.

00:01:40:09 - 00:01:52:00

Mythos is the raw, ultra capable version with fewer restrictions. Fable is the same power, but wrapped in safety classifiers that block the really dangerous stuff.

00:01:52:06 - 00:02:01:21

Think cybersecurity exploits in biology and for the first time ever, anthropic put a model this powerful into the hands of the general public through fable

00:02:02:00 - 00:02:04:06

that lasted three days.

00:02:04:06 - 00:02:18:22

On Friday, June 12th, the Commerce Secretary sent a letter directly to anthropic CEO. The message. Mythos five and fable five are now subject to export controls under National Security Authority.

00:02:19:00 - 00:02:23:05

So what's an export control? In plain English,

00:02:23:05 - 00:02:36:08

it's a rule the government uses to stop sensitive technology from leaving the country or reaching the wrong hands. It's the same category of tool used for missile parts, nuclear materials, and advanced military chips.

00:02:36:08 - 00:03:07:23

And for the first time, they pointed that weapon at an AI model itself. Here's the kicker the order didn't just ban foreign countries. It banned access for any foreign national, meaning any non-U.S. citizen, including people physically inside the United States and even anthropic own non-citizen employees. The rule was so broad that anthropic said they had no choice but to shut the models off for absolutely everyone while they figured out how to comply.

00:03:08:01 - 00:03:11:17

Reportedly, they were given about 90 minutes to do so.

00:03:11:19 - 00:03:19:11

Okay, so why did this happen? This is where the story goes from interesting policy news to. Are you kidding me?

00:03:19:13 - 00:03:35:00

According to reporting from the Wall Street Journal in fortune, the trigger was a jailbreak. Let me explain that word real quick, because it's central to everything. A jailbreak is when someone finds a clever prompt that tricks an AI into bypassing its own safety rules,

00:03:35:05 - 00:03:39:13

basically talking it into doing something it's actually not supposed to do.

00:03:39:15 - 00:03:46:02

So who found this jailbreak and reported it to the government? Here's where it gets crazy. It was Amazon

00:03:46:05 - 00:03:51:22

and you need to understand why that is wild. Amazon through AWS is anthropic

00:03:52:04 - 00:03:57:01

single biggest cloud partner. They invested billions of dollars into Anthropic

00:03:57:04 - 00:04:02:23

Fable five, literally launched on Amazon's own platform the same day it went public.

00:04:03:01 - 00:04:07:05

These two companies are business partners at the deepest level in

00:04:07:08 - 00:04:08:03

the security world.

00:04:08:03 - 00:04:29:20

There's a standard, respectful way to handle this. It's called coordinated disclosure. If you find a flaw in someone's system, you quietly tell them first, give them time to fix it, and then you go public. It's a basic professional courtesy. It's how the whole industry is supposed to work. Amazon, according to reporting, did not do that. Instead

00:04:29:22 - 00:04:31:09

of bringing the floor to their partner.

00:04:31:09 - 00:04:39:22

Anthropic. They reportedly took it straight to the Commerce Department. The CEO of Amazon himself reportedly raised it with officials.

00:04:40:00 - 00:04:43:04

So what was this terrifying, nation threatening jailbreak?

00:04:43:04 - 00:04:52:06

Anthropic says the entire thing was asking the model to read a specific piece of code and point out the software flaws in it. That's it.

00:04:52:09 - 00:05:24:22

And anthropic argues that this exact capability is already freely available in other public models, including OpenAI's GPT 5.5, which notably the government did not touch. Now, in fairness, Amazon hasn't fully explained its side and the government insists the threat is real and the lockdown is temporary while they harden their systems. But you can see why people are calling this what it looks like a business partner finding a flaw and running to the government with it instead of picking up the phone.

00:05:24:22 - 00:05:54:00

Let's bring this home, because so far this sounds like a fight between giant corporations, but this hits regular people directly. Remember, fable five was the first time a model this powerful was available to the everyday public. Not researchers, not the government, not big enterprises behind a paywall. Hundreds of millions of people had access. And then, with basically no warning, it was yanked away from all of them all at once.

00:05:54:02 - 00:06:10:21

Think about what that establishes. A tool you were using on a Friday afternoon can be switched off by Friday evening, not by the company that made it, but by a government order. You're no longer just a customer. You're a variable in a national security equation.

00:06:11:02 - 00:06:19:16

the foreign national piece should make every immigrant, every international student, every visa holder, every green card holder in the US.

00:06:19:17 - 00:06:21:04

Pay attention.

00:06:21:05 - 00:06:37:06

The order treated foreign nationals inside the United States the same as a foreign government overseas. If you're not a citizen, this rule said you're a risk, even if you're sitting in an apartment in Ohio using a chatbot to help you with your homework.

00:06:37:08 - 00:06:43:16

Here's the silver lining. And it's important. This was targeted anthropic. Other models,

00:06:43:16 - 00:06:48:08

including their flagship Claude Opus, kept running just fine.

00:06:48:10 - 00:06:54:08

So this wasn't AI is banned. It was a precision strike on two specific models.

00:06:54:09 - 00:07:11:18

Now let's zoom all the way out, because this is the part that actually keeps AI executives up at night. Number one, a brand new precedent for years. Export controls in tech were about hardware. The physical chips, the machines, the stuff you put in a crate.

00:07:11:23 - 00:07:20:06

This is the first time the government has aimed at that same legal weapon, at a model, at software, at weights and math.

00:07:20:07 - 00:07:38:11

The door is now open and it doesn't close. Number two, the chilling effect. Anthropic has been the company most willing to cooperate with the government. They even have a formal partnership to test their models with a federal AI standards body before release.

00:07:38:11 - 00:07:40:14

And this is the things they get.

00:07:40:19 - 00:07:59:07

So what's the lesson every other lab just learned? Maybe don't be so transparent. Maybe don't tell the government what your models can do. One policy expert warned this could push the entire industry toward less openness, not more. The exact opposite of what you'd want for safety.

00:07:59:09 - 00:08:03:05

Number three competitive distortion in thorax.

00:08:03:06 - 00:08:20:19

Core argument is brutal and simple. If this standard were applied fairly across the whole industry, it would halt new model releases for everybody. But it wasn't applied across the industry, it was only applied to them. Their competitors comparable model stayed online and untouched. So

00:08:20:19 - 00:08:23:17

So now we have to ask an uncomfortable question.

00:08:23:19 - 00:08:29:13

Is this about national security? And is it about who has friends in the right places?

00:08:29:15 - 00:08:41:07

One analyst who briefly worked inside the administration, said he couldn't tell whether this was targeted lawfare against one company or just extreme security paranoia, and called the whole thing using

00:08:41:07 - 00:08:43:16

his words cartoonish.

00:08:43:18 - 00:08:51:23

Another expert called it a step toward balkanization of technology, the internet and AI splitting into walled off national fortresses.

00:08:52:02 - 00:08:57:17

And underneath all of this is the real tension that's going to define the next decade.

00:08:57:21 - 00:09:00:03

Innovation versus control.

00:09:00:05 - 00:09:29:04

The same administration has been loudly pro AI, pro speed and anti. And then it turned around and used one of the heaviest tools in the entire federal arsenal to flip the off switch on the most advanced AI in the world. You cannot have it both ways forever. At some point the government has to decide what it actually wants, and the rest of the industry is going to build around whatever that answer turns out to be.

00:09:29:06 - 00:09:43:04

So here's my super confident take. This wasn't really about one jailbreak that asked an AI to read some code. This was a stress test of a question we've all been avoiding. Who actually controls frontier AI?

00:09:43:06 - 00:09:53:07

Is it the labs who build it? The investors who fund it? The cloud giants who host it, or the government that can switch it off in less than 90 minutes

00:09:53:09 - 00:09:56:03

for one Friday night in June, we got our answer.

00:09:56:03 - 00:10:17:17

And it wasn't the people who built the thing. Now, in throwback says this is a massive misunderstanding and they're working to bring the models back. The government says it's temporary. Maybe by the time you're watching this. Mythos and fable are already back online, but the precedent doesn't reboot. The door that open on Friday stays open.

00:10:17:22 - 00:10:19:23

So I want to hear from you.

00:10:20:01 - 00:10:29:22

Drop a comment. Was this a legitimate national security move or was a company that builds the future? Just punished for telling the government the truth?

00:10:30:00 - 00:10:37:12

And if a business partner found a flaw in your product and ran to the government instead of telling you, would you ever trust them again?

00:10:37:14 - 00:10:51:02

If this broke down a confusing story and made it make sense, you know what to do. Like subscribe. Hit the bell. Because in this industry, the next plot twist is usually about a week away. I'm Alex Smith.

00:10:51:04 - 00:10:56:07

This is super confident AI. Stay curious, stay confident, and I'll see you in the next one.