Fable Jailbreak & Ban — Full Breakdown & Transcript

[00:00] Over the weekend, something crazy

[00:01] happened. It just 3 days after the

[00:03] release of Claude Fable, the US

[00:05] government stepped in and curb-stomped

[00:07] it in the name of national security. And

[00:09] that's bad news if you just FOMO

[00:11] subscribed to Claude Pro to try out

[00:13] Fable because now you'll see this

[00:14] disappointing message if you try to use

[00:16] it and instead be forced to use the

[00:18] negative IQ Opus 4.8. But it's all for

[00:20] your own good because it only took

[00:22] someone a few hours to jailbreak Fable

[00:24] and turn it into an unstoppable cyber

[00:26] weapon. And that's pretty ironic because

[00:28] here in the land of the free, an

[00:30] American company that will not stop

[00:31] talking about AI safety just got

[00:33] safety'd by its own government. In

[00:35] today's video, we'll find out how and

[00:37] why our dear leaders in government are

[00:39] keeping us safe from the horrors of

[00:41] linear algebra. It is June 15th, 2026

[00:44] and you're watching The Code Report.

[00:45] About 2 months ago on April 7th, we were

[00:48] first introduced to Mythos 5, the raw

[00:50] unmuzzled model with the strongest

[00:52] cybersecurity capabilities of anything

[00:54] out there. But it was locked behind a

[00:56] program called Glass Wing, only

[00:57] available to trusted partners like major

[01:00] corporations and the US government

[01:02] itself. The reason Mythos can't be given

[01:04] to normies though is because it could

[01:05] easily be used as a cyber weapon in the

[01:07] wrong hands. To prevent that, Anthropic

[01:09] created a different product called Fable

[01:11] 5, which is literally the same exact

[01:13] model but with safety classifiers bolted

[01:16] on. That means if you ask it to do bad

[01:18] things like create an NPM package that

[01:20] turns the banking system into a

[01:21] Minecraft server, the Fable's guardrails

[01:24] will reroute your request to Opus 4.8

[01:26] for a dumber, more wholesome response.

[01:29] So basically, Mythos and Fable have the

[01:31] same brain but Fable has a child lock on

[01:33] it. If Fable went public and gained

[01:35] hundreds of billions of users overnight

[01:37] and it was awesome. It was by far the

[01:39] best coding AI model I've ever used and

[01:41] people were building all sorts of crazy

[01:43] apps with it. Life was good for about 3

[01:45] days. Then, of course, an anonymous

[01:47] internet user who goes by Plenty The

[01:49] Liberator defeats the guardrails and

[01:51] jailbreaks it. He's basically the

[01:53] internet's let's see if I can penetrate

[01:55] this thing guy and is famous for

[01:57] breaking other AI systems. And on June

[01:59] 10th, he posted a jailbreak on X

[02:01] claiming he popped Fable's guardrails

[02:03] wide open and got it producing exactly

[02:05] the same stuff the child block was built

[02:07] to block. And that's despite the fact

[02:09] that Anthropic had spent thousands of

[02:10] hours red teaming and trying to break

[02:12] its own guardrails internally. But the

[02:14] jailbreak wasn't some kind of sci-fi

[02:16] exploit. It actually works a lot more

[02:18] like money laundering. If Fable has a

[02:20] safety classifier watching for bad

[02:21] requests, but you can break dirty

[02:23] requests down into smaller

[02:25] innocent-looking fragments by wrapping

[02:27] them in weird Unicode characters, by

[02:29] doing roleplay farming, or by confusing

[02:31] the model in a very large context

[02:33] conversation. Due to national security,

[02:35] I can't be any more specific than that,

[02:37] but this weakness was brought to

[02:38] Anthropic's attention and they were

[02:40] initially asked to take the model down,

[02:42] but they refused. Then on Friday at 5:21

[02:45] p.m. Eastern time, Anthropic gets a

[02:47] letter not from a customer, but from the

[02:49] United States government. This letter

[02:51] was an export control directive signed

[02:53] off by Commerce Secretary Howard

[02:55] Lutnick. And the order was that no

[02:57] foreign national may access Fable 5 or

[02:59] Mythos 5. Not abroad, not in the US, and

[03:02] not even Anthropic's own foreign-born

[03:04] employees are allowed to touch it. That

[03:06] last one is pretty crazy. The government

[03:08] told a company that some of its own

[03:10] staff are no longer allowed to use the

[03:11] product they built. That means guys like

[03:13] Andre Karpathy who just recently got the

[03:16] job at Anthropic, they can't even use

[03:18] Fable. In response to that directive,

[03:20] they decided to hit the big red button

[03:21] and yanked Fable and Mythos for

[03:23] everybody. Now everybody's been quietly

[03:25] demoted back to Opus 4.8. And this is

[03:28] the first time in history a major AI

[03:30] company has pulled a live public model

[03:32] off the shelf because the federal

[03:33] government said so. Many developers out

[03:36] there are not too happy with Anthropic

[03:37] right now because on top of this whole

[03:39] situation, there was already backlash

[03:41] over reports that Anthropic was

[03:43] intentionally degrading Mythos and Fable

[03:45] performance on certain AI research jobs

[03:47] without making it obvious to users. But

[03:50] others out there are speculating that

[03:51] this whole thing was a calculated

[03:53] publicity stunt to continue pumping up

[03:55] Anthropic's pre-IPO numbers while

[03:57] simultaneously building a regulatory

[03:59] moat around it. But I think the only

[04:01] thing that can truly stop Anthropic at

[04:03] this point is a better model from a

[04:04] competitor. A leaked benchmark shows

[04:06] that Mistral might have that model, but

[04:08] we're also awaiting new releases from

[04:10] OpenAI and Google. Most of what we hear

[04:12] about AI is either non-stop hype from

[04:14] Big Tech or AI doomers warning us that

[04:17] the Skynet apocalypse will destroy the

[04:18] human race. But if you want to actually

[04:20] understand AI issues, you should check

[04:22] out blue.impact, the sponsor of today's

[04:25] video. They're a non-profit whose

[04:27] mission is to get more people involved

[04:28] in making AI go better for humanity. The

[04:31] main way they do that is by offering

[04:33] free online courses like their future of

[04:35] AI course, which provides an unbiased

[04:38] introduction to where AI is today and

[04:40] where things could be heading over the

[04:42] next few years. They also have more

[04:44] technical courses on things like AI

[04:46] governance and biosecurity along with

[04:48] personalized career support for people

[04:50] interested in working on AI safety. All

[04:52] of BlueDot's resources are free to use

[04:54] because they're funded by philanthropic

[04:56] donations, and they've helped over 8,000

[04:58] people get jobs at organizations like

[05:01] Google DeepMind, Stanford HAI, and

[05:03] Apollo Research. Try out their future of

[05:06] AI course at the link below. This has

[05:08] been The Code Report. Thanks for

[05:09] watching, and I will see you in the next

[05:11] one.