Ep 752: Why Anthropic’s New Mythos Model is Terrifying and How it May Change How Business Gets Done

This is the Everyday AI Show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life.

If there is a highly contagious yet highly controllable virus that breaks out, the immediate response is quarantine. You wouldn't hand out the virus like it was candy in a parade. And this may be the exact scenario we're seeing right now with Anthropic's new Mythos preview model. Anthropic says it's so good that it may actually be really bad in the wrong hands. And like a highly contagious virus, the uncontainable capabilities of their now private AI model could sow digital chaos or lead to chemical or biohazards. That us humans can't control. So Anthropic is keeping the model private by quarantining the Mythos model and reserving it for now for some of the biggest companies in the world until they feel it's safe. Sounds responsible, right? Well, that's the story we're all being told and the reality that we may all be living in. But what's the byproduct of Anthropic withholding an AI model that's allegedly too powerful for everyday business consumers? Well, for the first time in the AI era, we may be looking at a knowledge divide. The elite companies and in turn those that work with them may be able to reap the benefits the other 99.9% of humanity may not have access to. We're going to be exploring this one a bit more on today's episode of Everyday AI. And this one is actually really important, y'all. I know I'm sometimes long-winded and ramble on, but today I'm going to try to keep it tight because because I think it's very important that you listen and hear this and that we all talk about it. So here's the big picture, all right? Here's what's happening. Anthropic has built an AI model they say is so capable they refuse to release it to anyone outside of select partners. Mythos reportedly found thousands of critical vulnerabilities in the AI operating systems and browsers that we all use every day. And they said that right now they're only releasing it to certain companies in their project class wing, which we'll get to in a bit. And like I said, for the first time in the AI, the modern, uh, large language model era, there may be a huge divide in a the haves and the have-nots when it comes to AI. So stick with me for, promise this time, 20-ish minutes, and you will learn why, well, a coding breakthrough accidentally maybe created the most powerful cyber weapon that we've ever seen. Why I personally think Mythos may have more to do with war than coding or knowledge work, and why Anthropic may be intentionally playing the hero card at an overly convenient time. Let's get into it. If you're new here, my name is Jordan Wilson. Welcome to Everyday AI. It's a daily livestream podcast and free daily newsletter helping everyday business leaders like you and me not just keep up, but actually make sense of what's happening to get ahead and grow our companies and careers. Starts here with the unedited, unscripted livestream podcast, but to be the smartest person in AI in your company. Our website and newsletter is your cheat code. Go to youreverydayai.com, sign up for the free daily newsletter. We're going to be recapping the highlights from today's show and all the other AI news you need to know to get ahead. So let's get into it and talk about the new Mythos model that is apparently so good that it's absolutely terrifying. Uh, but here's what Mythos is and why I think you you should care. So Anthropic says that Mythos is a new general-purpose AI model that's far more powerful than anything that's publicly available. And no one trained it to be a cyber threat yet. Anthropic said that this is just a byproduct of how good that it was at coding, and that right now it far surpasses most skilled human experts at finding and exploiting software vulnerabilities. And right now they're releasing it under the umbrella of Project Glass Wing with some of the biggest tech companies in the world like Apple, AWS, Google, NVIDIA, Microsoft, and others. Basically just about every single big AI and tech company in the world. Well, not actually, but the biggest of the big except OpenAI. So this is from Anthropic's kind of post. So if you didn't get a chance to read it for the next minute or so, I'm going to read off this. So If you didn't check it out in our newsletter from 2 days ago, here we go. So they say today we're announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world's most critical software. We formed Project Glasswing because of capabilities we've observed in a new frontier model trained by anthropic that we believe could reshape cybersecurity. Claude Mythos Preview is a general-purpose unreleased frontier model that reveals a stark fact. AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Mythos Preview has already found thousands of high-security vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout for economies, public safety, and national security could be severe. Project Glasswing is an urgent attempt to put these capabilities to work for defensive purposes. As part of Project Glasswing, the launch partners listed above will use Mythos Preview as part of their defensive security work. Anthropic will share what we learn so the whole industry can benefit. We have also extended access to a group of already over 40 additional organizations that build or maintain critical software infrastructure so they can use the model to scan and secure both first-party and open-source systems. And Tropic is committing up to $100 million in usage credits for Mythos Preview across these efforts, as well as $4 million in direct donations to open-source security organizations. Project Glasswing is a starting point. No organization can solve these cybersecurity problems alone. Frontier AI developers, other software companies, security researchers, open-source maintainers, and governments across the world all have essential roles to play. Uh, the work of defending the world's cyber infrastructure might take years. Frontier AI capabilities are likely to advance substantially over just the next few months. For cyber defenders to come out ahead, we need to act now. So that's from it, right? Now let me, you know, kind of put my, put my hat back on. So a lot there, right? Just from their blog post announcement talking about Mithos. And I'm going to dig deeper into it, but I wanna say this. We often talk about the ChatGPT moment, right? Like it is a dividing line in the sand. And I think that ultimately for general knowledge work, it is. I'm not saying that Mythos is the next line in the sand, but I think that whatever comes from Mythos definitely could be, right? I think it's especially important to note write the partners that Anthropic is sharing this with initially, which include Google, AWS, Apple, Microsoft, NVIDIA, and others, and 40 unnamed companies working more in the software development space. But it's essentially, well, every big name in AI except OpenAI. But let's talk a little bit about what Mythos actually found in how it's a little terrifying, right? So it found, according to Anthropic, a 27-year-old vulnerability in OpenBSD, one of the most security-hardened systems in the world. It caught a 16-year-old FFmpeg flaw that automated testing tools hit 5 million times and never detected. And it chained together multiple Linux kernel bugs to escalate an ordinary user to complete machine control. All right, for our non-technical users, right? So these are some systems out there, right? Linux as an example, right? These have been hit by likely, or attempted to have been hit by likely, millions of different attacks over the years. And by some of the, you know, most prominent and most skilled human hackers, and on the flip side, human security experts in the world. And according to Anthropic, found thousands of vulnerabilities that the combination of previous artificial intelligence over the past few decades and the world's smartest cybersecurity experts have never been able to find. That is why this is potentially absolutely terrifying if this comes, turns out to be true. And I'm gonna end, uh, the podcast, I think, by giving you some advice, but I want you to think. Every single piece of technology that you interface with, in theory, if this is true, is now or could soon be extremely vulnerable. So that's why I said, please listen to this episode because I think it's so much more about your average, you know, AI, you know, large language model update. I was thinking about, you know, going over all the benchmarks. I don't even want to get sidetracked by that because yes, the benchmarks are good. I will talk about one or two of them. But I don't think that's ultimately what this is about. This is about security for everyday Americans. And I think it's probably, well, actually everyone in the world. And I think it's going to hit people, uh, very hard again if this, everything that Anthropic is saying, is true. Obviously they have no, uh, you know, known need to, uh, fabricate anything that they're saying here. And it does look like Anthropic is playing the good guy, but in a little bit. I'm going to dig into that part. So right now, it just says it's too dangerous to release. And one of the reasons why is, well, it's increasing capabilities and abilities in coding, right? So SWE-Bench Verified is one of the most important coding benchmarks in AI world. And it took an absolutely enormous leap. So Otis 4.6 was widely considered either the best or maybe second or third best model in the world when it came to coding, right, which it previously had an 80%, uh, Opus 4.6 on that benchmark, and Mythos just jumped to a 93.9%. Uh, the other thing that maybe explains, uh, some of the recent developments coming out of Anthropic is, well, Anthropic has been using this model internally since February, right? If you listen to the show all the time, I've literally said that I don't understand how Anthropic is shipping these updates so quickly over the last couple of weeks. And I've also said multiple times that Anthropic is undoubtedly winning 2026, right? Where I think OpenAI won 2022, 2023, 2024, Google won 2025. And so far Anthropic is crushing everyone else in 2026. And while it could be because they've been using Mythos to develop their other models in parallel, And yes, this is probably a whole nother conversation to be had, but when we talk about recursive learning and models being able to self-improve themselves and this whole conversation of AGI, artificial general intelligence versus artificial superintelligence or ASI, right? This is why I talk about maybe Mythos is a line in the sand that ultimately leads to the broader community saying, yeah, we've achieved AGI and Maybe we're now looking ASI, or artificial superintelligence, which is when things get really weird and really scary, straight in the face. So that's why I said this could be that line in the sand that comes after the ChatGPT moment that changed knowledge work, it changed information. Now this could be change security and a lot of other things. And the other thing is this is also could be the first time that there is a noticeably large gap in the access to information, intelligence, and technology. Right? So I think this divide existed pre-2020 or pre-2022 for people that maybe didn't use the GPT technology until ChatGPT was released in November of 2022. Before that point, obviously enterprise companies that had access to traditional machine learning and artificial intelligence did have a leg up on everyone else. But, you know, kind of that's the democratizing nature of large language models. And, you know, I think that we maybe will be saying RIP to the democratization of technology maybe at this point. So we'll see. Anthropic did allude to that they will be, they may release Mythos-level models in the future to the general public, but at least this one and maybe their most powerful one, they will withhold forever. They are going to, at least they say they're going to quarantine this because it's too good and too powerful.

AI moves too fast to follow, but you're expected to keep up. Otherwise, your career or company might lag behind while AI-native competitors leap ahead. But you don't have 10 hours a day to understand it all. That's what I do for you. But after 700+ episodes of Everyday AI, the most common questions I get is, where do I start? That's why we created the Start Here series, an ongoing podcast series of more than a dozen episodes you can listen to in order. It covers the AI basics for beginners and sharpens the skills of AI champions pushing their companies forward. In the ongoing series, we explain complex trends in simple language that you can turn into action. There's 3 ways to jump in. Number 1, go scroll back to the first one in episode 691. Number 2, tap the link in your show notes at any time for the Start Here series, or you can just go to startherseries.com, which also gives you free access to our inner circle community where you can connect with other business leaders doing the same. The Start Here series will slow down the pace of AI so you can get ahead.

And the story that you're probably going to hear a lot about, if you haven't already, is a sandwich, right? The sandwich heard around the world. All right. So this, according to Anthropic's model card, which I haven't read the entire thing, I think it's like 200+ pages. I've read through a good chunk of it so far, but according to researchers at Anthropic, they gave Mythos a secure sandbox but instructed it to, well, escape. And Mythos escaped the secure sandbox using a multi-step exploit to gain broad internet access on its own. It then posted the exploit details to obscure but technically public websites without any human requests. And they only found out about this because the researcher was eating a sandwich. The model contacted that researcher, right? Very nice, you know, kind of PR play there, right? Giving it something visual, something for us all to talk about. But here's why I want to put on my hat of devil's hat. Advocate. Okay. Very briefly, I worked in nonprofits for 10 years, and one thing that I would say I was pretty okay at is being able to tell stories, right? I'm a former journalist, but working in a nonprofit, and we partnered at an extremely high level with Nike and Jordan Brand. That's really all I did for 8 of my 10 years at this nonprofit was work on large-scale activations with Nike and Jordan Brand. And I think one of the reasons why we were able to secure such long-term partnerships is, well, we were good at telling the hero versus villain story. And I think that's actually what Anthropic is doing here, right? Because let's talk about the flip side of Mythos. In theory, if Mythos is so freaking good that they have to hold it back and it could create a legit international security incidents, which it still may, even after they put it through all these proper safeguards, couldn't they have at the same time told the story of all the good that this model could in theory do? Right? If this model is a precursor to artificial superintelligence, couldn't they unleash this thing for good? Couldn't they not immediately, but couldn't they very soon find cures for diseases that have been plaguing researchers for decades? Right? I don't think it's, we can understate the fact here, how much, right? And I don't fully understand security, right? But I know the basics. I know that these companies that, you know, mentioned Linux as an example, there's always people who are trying to trying to exploit this. And I assume there's been millions of attempts over the past few decades. And some of the smartest researchers in the world have spent their entire careers trying to secure pieces of software like Linux, right? So what about the flip side? I'm wondering, and maybe Anthropic did, right? You talked about some of their, you know, donations to open source projects. right? So I hope we get more details on those, and I hope that there are many in the medical, uh, realm, and I'm sure there are. But at the same time, why did Anthropic take this very, uh, hey, this thing is bad, right? We created the disease, but only we have the cure. Why not flip it on its head, right? Why not focus their attention and maybe their messaging on how much good this model could in theory ultimately do? I think there's reasons for that. It's the timing of it, but I'm going to get to that in a minute. But they painted this picture of a terrifying digital threat that essentially enters them as the only hero. Right? Because from my experience in the nonprofit world, right? We were working with at-risk youth. And what I— is you can only get people to join your team, right? Think of it like superheroes, right? If they know the villain. If you don't know the villain, if you don't understand the villain, if you don't tell the villain's story, no one cares about the hero and no one wants to join your team. Right? But if you focus on the villain, you can. And I think most people outside of those in cybersecurity have very little idea what these actual cyber threats are looking like. So like I said, does Anthropic controlling the cure give them too much power over the quote unquote disease that they're also creating. Because here's the reality. At some point, maybe this whole model won't leak, right? Or maybe this whole model won't be distilled. But you can't look past that, right? Anthropic, like most companies, maybe a little more so than others. I've had a lot of issues with model distillation from Chinese companies and with Leaks that were self-inflicted. And that brings me to the very, I won't say questionable timing, but interesting timing of all this. All right. So let's put a timeline together here. So like just over a week ago, Anthropic accidentally leaked, literally clawed code, their code to one of the most powerful and popular coding harnesses in the plant. They accidentally leaked that. A human accidentally leaked that to the public. Okay, and now we know that probably in the third or fourth quarter, Anthropic is targeting an IPO, right? There's also been some rumor— not, not rumors, but there's been some reporting on Anthropic's recent huge jump in revenue and how they count revenue differently than OpenAI. So Call me a conspiracy theorist or maybe the old investigative journalist in me trying to read between the lines. This seems kind of intentional, right? As saying, hey, we just created the world's most powerful cybersecurity threat. We created something that could make everyone's lives absolutely miserable. We just created something that could lead to you know, bio, um, and, and chemical, uh, you know, disruptions across the entire world. But don't worry, we're going to keep you safe. It's almost like maybe it's a distraction or a PR pump coming up before their IPO, because the fact that they leaked their entire codebase to one of the most popular current products on the planet. And they're trying to IPO in a couple of months, right? I don't know, call me crazy, but I think that launching a dramatic security initiative months before going public is either extremely heroic or very foolish. So time will tell, but I think the IPO might not even matter much compared to the broader implications. And I think that's 4. Stick with me. I've talked about this. I think I started talking about this late 2024, a little bit more last year. I think what we need to worry about, maybe aside from what we can actually control, is war. I think maybe that whatever happens here, if something goes wrong, and there's a decent likelihood that something could eventually go wrong, whether it is in Anthropic's control or not, this will lead to great implications in war. Or if, you know, Anthropic feels that they have hardened this model enough to safely release it to the public, okay, at what point then, uh, will a certain country distill the model like they did previously, right? A lot of the Chinese open source models reportedly were largely based on distillations or quote unquote illegally copying from Claude. So if Anthropic does release a Mythos-level model publicly, or if they have another leak, or if you get one rogue employee, and this is true, that is a huge implication for geopolitical and potentially world peace, right? And I don't think I'm exaggerating here. So India's Chief of Defense Staff said less than 2 months ago that future conflicts will be decided by AI and cyber, not conventional forces, right? And military analysts note that cyber operations can disrupt power grids and networks without firing a single shot. And I know this is not new, right? Cyberattacks are not new when it comes to, you know, two nations launching war against each other. It's actually a foundational piece of any country's military strategy. But I think, right, if from what I'm reading, and, you know, I obviously follow AI very closely, been doing it daily for more than 3 years. I think that's what this is ultimately. The thing that scares me personally the most isn't necessarily how, you know, in theory, a mythos-level model in the wrong hands could expose, exploit, replicate, and duplicate bugs essentially across every corner of our digital lives, which is a very real possibility. And I think that's what most people are focused But you have to take that to the extreme, right? And I think what that leads to is nations or state powers, you know, getting a hold of this technology that they probably maybe will use in bad ways, right? Think of it this way. Two nations going against each other in a conflict or a war. In one country, if they have this level and right, not going too, too, too far off here, you know, in terms of, you know, model self-improvement and things like that. But it's not outside of the realm of possibilities where a nation could shut off an entire country's power grid and maybe leave them helpless. Right. That's what I'm thinking about. Previously, right before this kind of dispute between the Department of Defense and Anthropic, Claude was the only large-scale AI operational on the Pentagon's classified systems before it was removed. And now they're in the middle of a kind of legal battle. But they may not be alone. Okay. I did previously talk about this kind of Anthropic Giving access or, you know, bringing other people in this program, even technically competitors. Yes, I know a lot of these competitors still pay Anthropic or have invested in Anthropic, like NVIDIA, Microsoft, and Google have all invested in Anthropic. But everyone except OpenAI, right? Well, and Meta, although I don't think Meta is in the same tier as Anthropic, OpenAI, Google, Microsoft. So OpenAI finished, even though they have a new model, which is pretty good, right? Meta's new model. But OpenAI just finished pre-training a new model codenamed Sput in March. So just a few weeks ago. And CEO Sam Altman told employees that Sput could really accelerate the economy and may release it in a few weeks. So OpenAI over the past month or two has cut down all of these compute-heavy resources such as SORA. So will OpenAI have a model that is a step change in the same way that we're seeing the jump from Anthropic's Claude Opus to Mythos Preview? Potentially, right? And I don't know if I'm OpenAI and I'm seeing, you know, literally everyone else in the room getting access to this apparently world's most powerful model ever by magnitude. I'm not taking this lightly from a competitive standpoint. Obviously, you know, the co-founders of Anthropic were former OpenAI employees as well. But the more immediate concern, right? Aside from danger lurking everywhere on digital, aside from the longer or medium-term implications of how a juggernaut cyber threat like Mythos could be used in the wrong hands, I think the other realization that will maybe be more impactful sooner is everyone else. Because I think for the first time, one company could hold a tool that is 50% or more better than anything else that the public can access. And I think the original fear, right, that launched OpenAI, that's, you know, OpenAI, one of their original kind of missional statements was that no single company should own and control this power. But the gaps between organizations using true frontier AI and those that are not is about to become a legit canyon. So we'll see how the average everyday enterprise may or may not eventually get access to a Mythos-level model or Mythos preview itself, right? We'll see maybe in a few months or a few quarters. maybe Anthropic won't have this perceived advantage. Maybe everyone else will have a level of model that is a huge jump ahead. But right now, presumably these companies that are getting access to this, yes, they are helping Anthropic, uh, make this model more secure, but presumably they're going to be using it for good reasons, which is driving economic value for themselves and presumably others that they interface with them, right? There are other clients. I'm not, you know, quite sure what the usage guidelines are on Mithips, but here's, I'm going to end with this. What you need to do, and this is, you know, I'm not sounding the alarms like maybe others are, but I do thank you for sticking with me to the end here. But you have to stay tech vigilant. All right. Even the fact, if you know anything about Linux, it is a sign, right? That Anthropic reportedly found thousands of bugs across operating systems, browsers, when there's literally millions of security researchers who have spent their careers doing this, right? This is why I'm one of those, people always look at me like I'm crazy when I've been saying for many years, AI is smarter than you, even if you are an expert in your field. And I think this is obviously very telling of this, if this does turn out to be true, that a model, an AI model found thousands of important security vulnerabilities in the technology that all of us use every single day. So what you need to do, stay tech vigilant. Any technology that you use, probably should make sure it's up to date because we're in theory just one leak or one bad actor away from living in a very scary digital world. Also prepare for deepfakes to accelerate because if Anthropic does release a Mythos-level model, assume it will be quickly distilled, right? Which is not in theory legal, but it is at least according to today's standards, pretty impossible to fight against. So probably it's time to start having the conversations with your family about fake calls. And fake messages, uh, because, right, whether that's in a few weeks, a few months, or a year or two, it's about to become very real and very prevalent. Because what this does is it gives the average, uh, kind of hacker or the average person who maybe is up to no good all of a sudden superhuman AI powers, right? Whereas before, yes, you could still do, uh, a lot of nefarious things with AI, but at first, you probably had to be a top 1%, you know, hacker to do those things powered by other frontier AI. And I think what we're seeing here could completely change this. But from a business perspective, all right, the 0.01%, if everything Anthropic says is true, just got a legit huge cheat code, right? That most people don't have. So what you need to do, if you haven't already, and if you're listening to the show, I really hope you're doing this, but you need to double down on your AI education and implementation, right? Just, yes, let's leave all the worries behind. I want to leave on a somewhat positive note because I know that I know that this ultimately may be upsetting and disheartening to some, but what you need to do is you need to double down on your company's training, education, and use of the best AI that you do have available. Because yes, the divide is going to be very real. Maybe this becomes a common practice by all major AI labs. Maybe it won't last for long, but at least for now, there is a sizable divide and there's nothing we can do about it. We've had a great, you know, 3.5, 4-year run where the access to technology has been democratized, but at least for now, that is gone. So you need to double down. You need to invest in systems that are working. You need to stay vigilant and at least for now, You need to work harder than ever. All right. I hope this was helpful. If so, please repost this. I think it's important that people hear about this, that people know about this, right? I tried to do my best to give it to you, you know, mixture of the news and the facts with, you know, some analysis of someone that's been doing this for a very long time. And, you know, I'm lucky enough to have great relationships with a lot of people at these types of companies that are building this. Please share this if you are listening on the livestream on social media. Please repost this. If you are listening on the podcast, appreciate your support. Send this podcast to someone. This is something that we all need to hear. So thank you for tuning in. If you haven't already, please go to youreverydayai.com. Sign up for the free newsletter. Thanks for tuning in. See you back tomorrow and every day for more Everyday AI. Thanks, y'all.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.