CBS News97%
Why Anthropic is saying its new AI model, Mythos, is too dangerous to release 50%
4/10/2026, 10:24:53 PM
BS Summary: This video contains 29 faulty reasoning types, including Appeal to Authority, Hasty Generalization, and Slippery Slope, with Appeal to Emotion as the most egregious example at 21.3% saturation with 190 hits. Analysis detected 1,904 faulty-reasoning hits from 893 analyzed words, generating a BS Score of 50% and a BS Rank of 50% (8,526 of 16,813 videos). This video is better (less manipulative) than 50.70% of the video peer group.
Welcome back to TechOut.
Thanks for hanging out with us.
Anthropic, which is a heavy hitter in the artificial intelligence world, has announced that it is teaming up with tech giants to, in its words, quote, "secure the world's most critical software."
What are they securing it from, you might wonder?
Anthropic's own AI model called Mythos.
Mythos has already found vulnerabilities in, quote, "every major operating system and web browser that it's interacted with."
And for that reason, fearing that bad actors, if they are thrown out in the public domain, could exploit Mythos's capabilities, Anthropic says it is not releasing it just yet.
A heck of a story.
New York Times reporter Mike Isaac joins us now to explain.
So, Mike, walk us through the basics.
What Mythos is, what it's been allowed to encounter in the software world, and what are the implications?
So, yeah, Mythos is the latest, what's called, model from Anthropic, a Silicon Valley startup, one of the most, I would say, buzzy AI startups out in San Francisco right now.
And with each of these models, they're getting a bit more sophisticated in what they're able to do.
You know, it's it's a type of research called reasoning, which allows them to, you know, think a little bit further ahead than each last one.
But Mythos is kind of different because this is very specifically focused on digging into operating systems for everything from desktops to smartphones to, uh, you know, different systems inside of the government, and finding any possible vulnerabilities inside them.
And I was reading about this as as they had released it, and they had made notes that, you know, some software that's been around for decades, uh, Mythos found bugs in it, vulnerabilities that were critical in just a few hours of of digging through it.
So, it's really, I think they I think it's probably smart of them to not give it to anyone if this can immediately find security, uh, flaws in the world's most important and powerful software.
So, Anthropic is not an altruistic entity.
It is doing this on a cost-saving basis for now, but when this is perfected, as I understand it, Mike, it will charge a pretty high rack rate.
I mean, you could also say this is one of their most effective marketing campaigns ever, right?
You know, I
>> You might also say that.
Yes.
>> [laughter]
>> I I called around to some of the big tech companies.
Right now, they're not releasing it.
They're only allowing like the big tech companies, some of the banks, uh, to use it.
And I called around and I said, you know, like, is this real?
Is this hype?
You know, Anthropic is filled with a bunch of people who are often saying this is very dangerous that AI is going to harm us.
Uh, and they kind of landed somewhere in the middle.
They said, look, this is actually finding real security issues that we need to take seriously.
You know, Anthropic is, uh, often seen as like the Chicken Little of Silicon Valley, but this is actually important stuff in how the infrastructure of the world works.
Uh, two people not named Chicken Little, the chairman of the Federal Reserve, Jerome Powell, and the Treasury Secretary of the United States, Scott Bessent, met, we are told, we have confirmed at CBS News, with a large number of very large banks to discuss Mythos and vulnerabilities because there is a sense, Mike, if I have this correct, that if vulnerabilities are exploited, lots of global finance could be at some level compromised.
I think that's certainly right.
And this is [clears throat] one of the things, too, that a lot of these very important financial systems and the stuff that, you know, our bank accounts essentially are running on or our 401(k)s, uh, are running on software that's been around for a very long time, increasingly large and complicated to support, you know, millions, if not billions, of customers.
And that means there's just like a lot of holes in it.
That's millions of lines of code that now, with these AI models, you don't have to have like a security researcher like me, I'm not a researcher, but like an engineer, looking through it.
You sick your little robots, uh, a crowd of robots out into the world to dig through it, and in hours you can find crazy flaws.
So, I think there's a real, uh, sense of urgency, and like there's real legitimate, you know, concerns by the big banks.
And I'm I'm pretty impressed, I would say, that the government is moving as quickly as they are on this sort of stuff, you know, considering they may not have done so, you know, just a few years ago.
Mike, very quickly, if AI were smart enough, and this is the central question to its future, how smart will it get, could it not only detect vulnerabilities, but create them in themselves within software programs?
That's right.
I think that's the sort of threat.
And all of the testing of all of these robots that you see out here is done basically in like internet clean rooms.
Analysis
Hover over highlighted words in the article to view the associated bias or fallacy analysis.