Anthropic announced a new, non-public model that is more capable than ever, finding bugs and security vulnerabilities that existed for 27 years. It's kind of frightening...
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
As a software developer, and I guess in general (laughs) as a human, but especially as a software developer, Anthropic right now, if you want to or not. And I don't think you should try to ignore it because it is relevant. It's relevant for our future as software developers, I would say. And in this episode, I won't talk about the Claude Code leak we had last week. I won't talk about their, uh, reinforced terms regarding the usage of their subscription
offerings, Claude Max and so on, and how they're cracking down on unauthorized use of those subscriptions. They are doing that right now because, of course, their subscription offerings, just like the ones by OpenAI, are heavily subsidized and they can't make any money if everybody maxes out their subscriptions. So yeah, they're really restricting, or they're trying to restrict, subscriptions to just humans, just on their
website, or in Claude Code or the Claude desktop app, I guess. Uh, but yeah, again, this is not the focus here and I won't even focus on their impressive revenue growth, which is worth, uh, a, a short note though, because Anthropic has reached a annual recurring revenue of $30 billion, which is already impressive, but especially impressive $9 billion at the end of 2025. So, they
more than tripled their annual recurring revenue within just a few months, which is really impressive. And therefore, of course, if you wanna learn how to efficiently use Claude Code, how to get the most out of it, And, um, it's also very popular, which of course makes me happy, and you'll find a link below if you wanna join it and learn how to efficiently work with Claude Code. But as mentioned, that's not even the main topic here. Instead, I wanna talk about Project Glasswing and
their new model, Mythos, which they haven't released to the public, and they also shared why. And I think this is important to understand, and it's also important to try to look behind the scenes, behind their rationale, and what the impact of this new model and how it works and its capabilities is for us developers. So, what is Project Glasswing? What is their new model all about? Uh, below, of course, you'll find a link to this article as
article on the official Anthropic site where they announce Project Glasswing and also talk about their new model. And if I scroll down a bit, we can already see some, uh, summary benchmark statistics here, where we can see that this new model, the Mythos, uh, preview version of the model, the, the, the model name is Mythos, um, performs way better than Opus 4.6. And depending on which exact benchmark you look on, there is quite a big difference
between Opus 4.6 and this new model. Now, of course, this is, uh, on its own, not super impressive. Whenever a new model is announced, no matter by which company, it performs way better, or at least a tiny bit better, models, otherwise it wouldn't get released. And of course, there are ways of gaming some of these so I typically don't care too much about those benchmark numbers, and that wouldn't really be different for this model here, but there are interesting things about the
new Mythos model. And that's the fact that Anthropic decided to not release it to the public because, as they say, it is too good at finding and exploiting vulnerabilities in operating systems, any other software, browsers, really, in software in general. And in this article, and also in a separate article, which is also linked below, they share some details, and especially this separate article is
extremely long and gives concrete examples for vulnerabilities and potential exploits this new model found. Uh, for example, they start in this article with a very serious exploit and vulnerability that was found in, uh, OpenBSD. OpenBSD is, of course, an operating system which is popular on certain pieces of networking software, for example, and Mythos, their new model running in
Agentic Harness, like Claude Code, I guess, was able to find and exploit, a- and that's the interesting part - a vulnerability that was related to integer overflow and, uh, memory access, unexpected memory access, um, that was able to crash machines that were running OpenBSD in a reproducible way, which could, of course, be leveraged to run very damaging,
uh, denial of service attacks by over and over sending specific packets and requests to such machines here, which exploited that vulnerability to bring down those machines and potentially bring down entire corporate networks. And this vulnerability was detected in a run that cost under $50, though the overall set of runs cost under $20,000. And since you, of course, don't know in advance which run will find a vulnerability, it's that number that matters.
Still, of course, it's easy to imagine that a model capable of finding such critical vulnerabilities for such a comparatively low cost...Depending on who you are, if you're a nation, um, for example, or some, uh, serious bad actor, that may not be a lot of money to you. That, of course, is a problem because it's easy to imagine that if such a model were developed by a company, an organization
that cares a bit less about security and/or that maybe doesn't have to fear any consequences of abusing such vulnerabilities, that this could be a problem. And it seems as if we're entering a new age with AI, with these AI models, where nothing is secure and
it's easier than ever to mass deploy AI agents running models like this to scan all kinds of pieces of software and find and potentially exploit vulnerabilities. And of course, as a human on, on your own, there is no way of keeping up with it. I mean, the bug, the exploit that was found here existed for, I think they said 27 years or something like this. This shows that no human was able to
find this bug in such a long period of time, including bad actors, which of course would've had an interest attack this operating system in the past too. Now this is just one, maybe the most prominent finding, um, this new model had. They are listing way more bugs and exploits the model was find and sometimes also able to abuse, and they also shared other stories. Uh, on X, for example, like the model being able to
escape a sandbox or the AI agent running the model being able to escape a sandbox in which it was running. And that brings us back to Project Glasswing, which is an initiative created by Anthropic together with other big companies like AWS, Apple, um, Microsoft, the Linux Foundation, and others to use this model to basically patch up their software before this is
publicly released and before the public gets their hands on this model. That is the, the narrative of this article. That is the explanation of Anthropic. And I have some mixed thoughts here. Now, for one, I have no strong reason to believe that this isn't true. Clearly, Anthropic would have some reasons to not release this model outside of what they're mentioning here. For example,
um, I read that this model is a roughly 10 trillion parameters model, which is way bigger than all the frontier models we had thus far or were able to use publicly thus far, and training it is said to have cost around $10 billion. The token cost of this model, I read, is expected to be around this range, $25, $125 for, uh, input and output tokens.
And of course, that would also be reasons for not releasing this model because they can't include it in their Claude subscriptions because it's just too expensive. They would have to ramp up the to a price point where not a lot of people are willing to pay it, uh, and therefore there wouldn't really be a exposing it to the public, at least as part of Claude code. Now of course, they could still expose it through their pay-per-use cost basis, and if it's expensive, who
cares? If there are companies, people that would be willing to pay it, they could do that. And that, of course, is the part where the cybersecurity concerns might really come into play because clearly that all is very likely not made up. I mean, it's definitely not made up. The FFmpeg team, for example, which is also listed here as a, um, as a vulnerability that they were f- able to find a vulnerability in FFmpeg,
uh, the team confirmed on X that Anthropic sent, uh, sent a patch for, uh, uh, a vulnerability in the FFmpeg, uh, software, uh, program So yeah, this is clearly not made up. These concerns are valid. Cybersecurity concerns are valid, especially since, of course, if money is not the main issue, you could deploy thousands of agents running simultaneously using this or
similar models, which we may have in the future to, yeah, scan all kinds of software and exploit them. And of course, the big problem is that using this model to find vulnerabilities and patch them up is possible, but it's only possible if the owner or the maintainer of a certain piece of software can afford the model or gets access for free or anything like that.
And even if a vulnerability is patched up, we all know that not all (laughs) computers out there, not all machines, not all users have up-to-date software running on them. If you were to take a look at all the various servers running out there in the worldwide webI would guess the vast majority of them is running outdated software. I mean, on our phones, our, our laptops, we are
of- not running the latest software, the latest version of our operating system. The latest security patch may not be installed and that is true for all layers of s- of software. And in a world where it's easier than ever find security vulnerabilities, that of course becomes an even bigger issue. Because of course the good thing about this AI model is that it can also be used for proactively looking for security
vulnerabilities and patching them. So it's not just a tool for attackers. It can also make defense easier because you now have a tool that can be run simultaneously in parallel across thousands of agents to make your software secure. In theory, this can be a very useful tool for of course, again, not every company, person that may be developing crucial software may be able to afford it, may be interested in using it, and then even if it is
used to find and patch vulnerabilities, still these latest versions will not be installed everywhere. And that of course gives attackers, uh, opportunity where they know about way more vulnerabilities than before at some point because way more vulnerabilities are detected, but not every machine, every user is protected against those vulnerabilities and that is one real concern I have
about this development. Now that is the, the bigger picture which affects everybody, all companies, all humans in the end. Another question of course is what does a model like this mean for us developers? I mean, clearly this seems to be a highly capable model that was able to look for vulnerabilities on its own and exploit vulnerabilities on its own. So yeah, what is the impact for
developers? And I think here, when it comes to that, not a lot changes for now. I mean, we're already living in a world where AI agents like Claude code and the underlying models, true for Codex and so on, whatever your favorite AI agent are able to generate most of our code. You may not be using them, you may not like them. I created a separate video where I share my feelings about that and that
this sucks the joy out of the, the software development part for me, but it is the, the reality nonetheless, if you like it or not. And believe me, I don't necessarily like it, but yeah, reality. Nonetheless, what a human brings to the table or why humans still matter here and may matter more than ever is, of course, that you definitely don't want an AI agent like this go rogue and work totally on its own. Steering
such models and agents, controlling them, giving them clear tasks, limiting the scope of the work they do, all these things are more important than ever. These models can, as it seems, do way more than the vast majority of developers can do. Definitely way more than I can do. And yet, when it comes to shipping products, comes to building software used by humans,
the influence of a human is of course extremely, extremely important. What's changing, of course, is our role as software developers. We're changing from the people that are writing the code to the people that are steering the model, that are reviewing the code, that are understanding what it does, that are setting the scope. And yeah, again, I talked about this in that other video, and then that this may not necessarily be what, uh, what, what you like. It's definitely not w- why I got
into software development in the first place. But yeah, this is the impact here and the more capable these models get, the more important I think it gets to have that, that human voice in there as well, that human influence, uh, in there as well So that is, that is that changing role and, and our role in the future. But yeah, I mean, these are really interesting developments and
especially this model and its implications and that cybersecurity relevance it has makes one think what would have happened or what would happen if other actors, other nations or organizations in the world get their hands on this model or models that are similar in capability. Because of course it's only a matter of time until models with similar
capabilities are accessible by the public or certainly at least by other nations and actors. And yeah, I'm not sure if we are prepared for that new race in cybersecurity and that delay between bugs being found and being patched and people installing those patches. I think we'll enter a new era of cybersecurity and we'll
be able to adjust, I'm sure, but this, uh, definitely marks an interesting, uh, point in the history of, of model development I would say
GPT 6 Astra was released. It's a good model, I have to say. Especially for 3D work it seems to be a real game changer. But due to all the hype, you may be disappointed...