[0:00] Over the weekend, something crazy
[0:01] happened. It just 3 days after the
[0:03] release of Claude Fable, the US
[0:05] government stepped in and curb-stomped
[0:07] it in the name of national security. And
[0:09] that's bad news if you just FOMO
[0:11] subscribed to Claude Pro to try out
[0:13] Fable because now you'll see this
[0:14] disappointing message if you try to use
[0:16] it and instead be forced to use the
[0:18] negative IQ Opus 4.8. But it's all for
[0:20] your own good because it only took
[0:22] someone a few hours to jailbreak Fable
[0:24] and turn it into an unstoppable cyber
[0:26] weapon. And that's pretty ironic because
[0:28] here in the land of the free, an
[0:30] American company that will not stop
[0:31] talking about AI safety just got
[0:33] safety'd by its own government. In
[0:35] today's video, we'll find out how and
[0:37] why our dear leaders in government are
[0:39] keeping us safe from the horrors of
[0:41] linear algebra. It is June 15th, 2026
[0:44] and you're watching The Code Report.
[0:45] About 2 months ago on April 7th, we were
[0:48] first introduced to Mythos 5, the raw
[0:50] unmuzzled model with the strongest
[0:52] cybersecurity capabilities of anything
[0:54] out there. But it was locked behind a
[0:56] program called Glass Wing, only
[0:57] available to trusted partners like major
[1:00] corporations and the US government
[1:02] itself. The reason Mythos can't be given
[1:04] to normies though is because it could
[1:05] easily be used as a cyber weapon in the
[1:07] wrong hands. To prevent that, Anthropic
[1:09] created a different product called Fable
[1:11] 5, which is literally the same exact
[1:13] model but with safety classifiers bolted
[1:16] on. That means if you ask it to do bad
[1:18] things like create an NPM package that
[1:20] turns the banking system into a
[1:21] Minecraft server, the Fable's guardrails
[1:24] will reroute your request to Opus 4.8
[1:26] for a dumber, more wholesome response.
[1:29] So basically, Mythos and Fable have the
[1:31] same brain but Fable has a child lock on
[1:33] it. If Fable went public and gained
[1:35] hundreds of billions of users overnight
[1:37] and it was awesome. It was by far the
[1:39] best coding AI model I've ever used and
[1:41] people were building all sorts of crazy
[1:43] apps with it. Life was good for about 3
[1:45] days. Then, of course, an anonymous
[1:47] internet user who goes by Plenty The
[1:49] Liberator defeats the guardrails and
[1:51] jailbreaks it. He's basically the
[1:53] internet's let's see if I can penetrate
[1:55] this thing guy and is famous for
[1:57] breaking other AI systems. And on June
[1:59] 10th, he posted a jailbreak on X
[2:01] claiming he popped Fable's guardrails
[2:03] wide open and got it producing exactly
[2:05] the same stuff the child block was built
[2:07] to block. And that's despite the fact
[2:09] that Anthropic had spent thousands of
[2:10] hours red teaming and trying to break
[2:12] its own guardrails internally. But the
[2:14] jailbreak wasn't some kind of sci-fi
[2:16] exploit. It actually works a lot more
[2:18] like money laundering. If Fable has a
[2:20] safety classifier watching for bad
[2:21] requests, but you can break dirty
[2:23] requests down into smaller
[2:25] innocent-looking fragments by wrapping
[2:27] them in weird Unicode characters, by
[2:29] doing roleplay farming, or by confusing
[2:31] the model in a very large context
[2:33] conversation. Due to national security,
[2:35] I can't be any more specific than that,
[2:37] but this weakness was brought to
[2:38] Anthropic's attention and they were
[2:40] initially asked to take the model down,
[2:42] but they refused. Then on Friday at 5:21
[2:45] p.m. Eastern time, Anthropic gets a
[2:47] letter not from a customer, but from the
[2:49] United States government. This letter
[2:51] was an export control directive signed
[2:53] off by Commerce Secretary Howard
[2:55] Lutnick. And the order was that no
[2:57] foreign national may access Fable 5 or
[2:59] Mythos 5. Not abroad, not in the US, and
[3:02] not even Anthropic's own foreign-born
[3:04] employees are allowed to touch it. That
[3:06] last one is pretty crazy. The government
[3:08] told a company that some of its own
[3:10] staff are no longer allowed to use the
[3:11] product they built. That means guys like
[3:13] Andre Karpathy who just recently got the
[3:16] job at Anthropic, they can't even use
[3:18] Fable. In response to that directive,
[3:20] they decided to hit the big red button
[3:21] and yanked Fable and Mythos for
[3:23] everybody. Now everybody's been quietly
[3:25] demoted back to Opus 4.8. And this is
[3:28] the first time in history a major AI
[3:30] company has pulled a live public model
[3:32] off the shelf because the federal
[3:33] government said so. Many developers out
[3:36] there are not too happy with Anthropic
[3:37] right now because on top of this whole
[3:39] situation, there was already backlash
[3:41] over reports that Anthropic was
[3:43] intentionally degrading Mythos and Fable
[3:45] performance on certain AI research jobs
[3:47] without making it obvious to users. But
[3:50] others out there are speculating that
[3:51] this whole thing was a calculated
[3:53] publicity stunt to continue pumping up
[3:55] Anthropic's pre-IPO numbers while
[3:57] simultaneously building a regulatory
[3:59] moat around it. But I think the only
[4:01] thing that can truly stop Anthropic at
[4:03] this point is a better model from a
[4:04] competitor. A leaked benchmark shows
[4:06] that Mistral might have that model, but
[4:08] we're also awaiting new releases from
[4:10] OpenAI and Google. Most of what we hear
[4:12] about AI is either non-stop hype from
[4:14] Big Tech or AI doomers warning us that
[4:17] the Skynet apocalypse will destroy the
[4:18] human race. But if you want to actually
[4:20] understand AI issues, you should check
[4:22] out blue.impact, the sponsor of today's
[4:25] video. They're a non-profit whose
[4:27] mission is to get more people involved
[4:28] in making AI go better for humanity. The
[4:31] main way they do that is by offering
[4:33] free online courses like their future of
[4:35] AI course, which provides an unbiased
[4:38] introduction to where AI is today and
[4:40] where things could be heading over the
[4:42] next few years. They also have more
[4:44] technical courses on things like AI
[4:46] governance and biosecurity along with
[4:48] personalized career support for people
[4:50] interested in working on AI safety. All
[4:52] of BlueDot's resources are free to use
[4:54] because they're funded by philanthropic
[4:56] donations, and they've helped over 8,000
[4:58] people get jobs at organizations like
[5:01] Google DeepMind, Stanford HAI, and
[5:03] Apollo Research. Try out their future of
[5:06] AI course at the link below. This has
[5:08] been The Code Report. Thanks for
[5:09] watching, and I will see you in the next
[5:11] one.