Claude Mythos, Automated Bug Hunting, and AI Eating Everything

And I think she had some great points about, you know, this is one thing for the companies whose products, you know, the companies like Microsoft and Apple and Cisco who are the Project Glass Wing participants, they know how to handle vulnerability disclosure. They've been doing it for decades. But for the defenders on the other side who are about to deal with like the the deluge of the CV ES coming their way, even for large mature security teams that know how to triage these things and prioritize them, no one in living memory has ever seen a flood of vulnerability disclosures like this. It just, it has never happened because that kind of volume of research hasn't been possible before. So we don't really know what we're about to see. Right.

Yeah, I love the this quote from Katie. She said the repo man for technical debt is coming, which was I was just like, that's a plus. But it's it's so true. And I think it's smart to kind of differentiate into several different buckets, which is that, you know, like you said, we have like these very large tech vendors and software makers and people that have The Who are at the forefront of making kind of the all the technology that we use. And that's one kind of bucket that needs to focus on these AI vulnerabilities.

But then also just kind of the everyday organizations, whether that's like large enterprises and kind of the the security teams that defend them all the way down to, you know, companies that are smaller and don't have those resources, don't have those IT teams in place. So I think there's just a really wide, you know, like a wide wave I guess of how this is going to impact different organizations. And I know you, you also came out with this Q&A or this, it was almost like a FAQ earlier this week. And, and one thing that you talked about in that was how if you know, you're someone is the only IT person in a company with 200 employees, like, you know, how is this going to impact their security posture like, And I think it's important to highlight what exactly what you said, which is, you know, probably not a lot in the foreseeable future at least. So I think there's a lot of

different kind of angles. Yeah, I think in those cases, you know, and we tend to forget, I mean, when I say we, I mean us as people who live in the security world, as a profession, you know, kind of eat and breathe it. We, you know, a lot of the people that we talk to in the community tend to work at large companies. They work at big software vendors, you know, they're on well funded research teams or or they're on well funded security teams. And they have that point of view of like, OK, I do have resources to address this.

How am I going to deploy them? But the vast majority of organizations don't have the resources of Google or even like a small regional bank. You know, if like a small town's school board or they're, you know, water department, they don't have a security person. They probably don't even have their own IT person. They have shared IT resources with their town or they have outsourced IT resources.

And if they're outsourced IT person all of a sudden has 600 CV ES that are relevant to your network, there's no way in God's good earth that they're going to be able to address those for you anytime soon. You know, even with automated patch deployment and all that kind of stuff, it's just, it's just not going to happen. So I think we're about to see some really weird things break and fall down in places that maybe we weren't expecting. I don't know exactly what it's going to look like. I don't think anyone does, you know, even folks like Gary and Katie or even the folks on the on the research side who are running these LLMS against research targets and finding these bugs, They're kind of like, hey, look what this cool thing can do.

We don't really know what the consequences of all of this disclosure is going to be or of this kind of research is going to be. You know, how is this all going to shake out? Nobody has a freaking clue. I don't think, no. I've talked to at least. Yeah, I know. And it's hard because this is something where it's just going to take time for this to develop and things to come out. But I think there's also just so many unexpected impacts that we haven't thought of either. Like when I was talking to Katie, one thing that she talked about that I couldn't even, she had so many good thoughts that I couldn't even get it into the, into the video I posted.

But she was saying because there's going to be some sort of impact also on cyber security talent as well. Because people are going to be relying so much, so much on AI that, you know, we're going to see a big gap there of, you know, the, you know, different things of that that that LLMS might start doing that people used to do and kind of how things to learn from that. And so I thought that was really interesting as well. And there's just a lot of different kind of takes on the Mythos news that I've seen all over Twitter and everywhere. So yeah, it's just.

There's a lot, yeah. The hot take cycle just went crazy in the last 10 days. After this, the the effects on cybersecurity professionals, I think is a really interesting question. There's the very small pool of elite vulnerability researchers who, you know, for the past 25 years or whatever have been doing the kind of research that now is being automated. And I don't, you know, you're talking about dozens or maybe a few, 100 people at the most that are at that level of doing that kind of bug hunting.

Those people are not going to be out of jobs. They're incredibly smart people that are, they're fine. But the automated research and vulnerability development and exploit development and patch development, I don't know how that's going to affect say people early to mid career in cybersecurity that might have been doing jobs like that. Those might be, you know, there might be like a 8 to 10 year generation of people that are just like, well, that job doesn't exist anymore. Like I don't, you know, like auto workers from 50 years ago when robots started coming into auto plants and those jobs got automated out, you know, completely different scale of people.

But yeah, I haven't run into anybody that has like a real definitive answer yet. Yeah, I know. And it's all, it's hard already because you know, this is a, this is a tough job market. Yeah. So it's just like on top of that, dealing with with some like a, a potential challenge like that is, is also difficult. It is, yeah. And the the hot take, you know, hype cycle thing about around Project Glasswing, there's there's two separate things. I think everybody was conflating Claude Mythos and Project Glasswing, and Mythos obviously is its own thing. It's this unreleased LLM that Anthropic, I guess, announced, you know, right alongside Project Glasswing. And Glasswing is the sort of partnership with all of these software makers that allowed Anthropic to run Mythos against their products and look for bugs, which it found, you know, hundreds of, we don't, we don't

know exactly how many, as we said, but you know, you can go and people can go read our stories. You can see who those companies are, but they're the ones you would expect Apple, Microsoft, Cisco, Crowdstrike, a couple of open source projects, Google, the Linux kernel, few others that I'm I'm forgetting, But you know, it's, it's the usual suspects and a lot of people were like, oh, that's, you know, this is another class division in technology that's great for those companies. But what about all of these other open source projects that could really use some help that don't have funding to go and spend with Anthropic on this kind of thing? That whole line of reasoning, which I completely sympathize with.

You know, there's, we see this all the time with something like, you know, the Axios compromise or another, whenever there's another open source project that gets compromised or a large vulnerability is found in it and everybody realizes how fragile the ecosystem is and there's just these open source projects kind of holding everything up. But that's a problem for another day. I think we're going to probably see some other AI providers trying to do something along the lines of what Anthropic is doing with this. It wouldn't.

It wouldn't. I'd be shocked if that didn't happen honestly, because it it makes a lot of sense to try and do this kind of large multi party vulnerability disclosure instead of individual researchers or companies or teams just running models against products, which, which they're all doing. I mean, that's, that's legitimate stuff, but trying to do it in a coordinate coordinated way makes a lot more sense.

It's difficult and you have to get a bunch of people to agree to things, which is never easy, but I think we'll probably see more of that. Yeah, didn't chat GP or Open AI come out with something too like Speaking of other, you know, it was something like trusted access for the next era of cyber defense. I'm just looking at it now. That was yesterday, OK, and there are some not so subtle shots at Anthropic in that very, very long blog post, which I mean, as a very petty person myself, I enjoy seeing that kind of thing, but it was it was one of those as like, you know, we're here for everybody, not the privileged few type thing like, OK, you're owned by Microsoft, let's all settle down. I mean, but they also I I did see something this morning too where I I might get this wrong.

I don't want to let me see if I can find it real quick. Oh yes, Open AI has like a cybersecurity grant program where they're giving money to some smaller companies and I think open source projects to, you know, fund research. And I saw some news this morning, they a couple of the first recipients of that were Socket, which does very cool security stuff. So it's worth going and looking at that. I don't know how much money it is, but Caliph is another company along with trail of bits that were recipients of these grants.

So Caliph for people that aren't familiar, you know, relatively new company, but is doing some of that completely insane AI assisted vulnerability research right now. You can go and just see them posting stuff every day on Twitter of like, here's another like critical bug we found. Here's this, here's this. Some of these bugs are like, you know, older than we, not older than me, but you know, 20 years old crazy things that they're finding. So there's just there's a lot happening.

Yeah, it's those bugs that Caliph has been finding. I feel like it's every single day that I've just been. Seen it is I've. Churned them out and it was like yesterday or the day before it was using codecs to hack a Samsung TV. And then there were a couple other ones like the Linux kernel, like something about, you know, a bunch of other stuff. So. As the owner. Watch come out. Oh, it's crazy to see. And those, there's a lot of talent at that company, but I own a couple Samsung TV's. It's completely unsurprising to me that they could hack them. They're that's.

True. The least surprising thing I've ever heard. Yeah, the the thing you mentioned at the at the beginning, we talked about this before we started recording. I was listening to this completely unrelated podcast this week, 3 guys that basically talk about conspiracy theories and like kind of other things, but in like, you know, light hearted way. And somehow this week, one of them brought up Claude Methos.

And these are guys that are not in the tech world whatsoever. They're, you know, essentially like sports and entertainment podcasters. And they had a pretty good handle on what Claude Mythos had, like is and what Project Glasswing is and how what Mythos had found. And but they were they were also like, is this real? Like, did this actually happen? Or is this just PR marketing BS, which is I thought a pretty funny take on it from, you know, complete outsiders looking at it from, you know, through the, the lens of like, can we really believe what these companies are telling us? What this shit like, is this actually happening? Like did they find a 27 year old exploit in open BSD? Like they talked about this on this podcast, like the open BSD bug, which I happen to know the guy that wrote the code that that bug, that bug lived in.

And he was just like, I can't believe they found that bug after 27 years. Like I committed that code like all those years ago and had no idea. But it's, it's just crazy to me to see like complete non security, non tech people looking at this and being like, is this total garbage or are are we just on the verge of something completely new? And I think the answer is like a little bit of both in some cases, but. Probably, Yeah. Yeah, that's interesting that I would say like even some people in cybersecurity are having that take as well. Like I've seen a lot of LinkedIn and Twitter, you know, hot take posts about that, so.

No doubt. Yeah, it's yeah. It's just like I said at the beginning, it's, it's taken over the almost every conversation that I have with people that aren't even meant to be about AI somehow just go down that road. You know, it's, it's crazy. Like I did a podcast yesterday that hasn't been published yet, but it was about it was with the CEO of a company that does it does a couple different things, but one of the things they do is AI assisted like GRC compliance stuff. So they'll help like this AI assisted product will help you essentially fill out their security compliance questionnaires so you can actually go and do work, which is like seems like a good use of AI to me. But yeah, there's just a million different things happening. Yeah.

Well, you know, to your point earlier about too, about how people outside of cybersecurity are, you know, latching onto this, if you look up what I mean, Mythos is like one of the biggest headlines that we're seeing related to AI right now. And it's, it's completely cybersecurity centric. But like, like, for instance, I just looked up Anthropic on Google News and like, you know, it's not like they're like releasing new models for every single use case or application across every industry. Like this is, it's just really funny that this is the area that they decided to focus in on and pivot towards. But like, yeah, outside of like I'm seeing a lot of the mythos AI news just in the general broader news. And then of course, like, I don't know, anthropic peace talks at the White House, like those types of things. But like, like, I don't know,

it's just really funny that, you know, get to be in this industry where all this change is happening. Yeah, it sure is. It was. Our industry. I know it's, it's weird where you'll like, I'll walk by while my wife is watching like the Today show or something and they'll be talking about this stuff. And it's funny because normally they're talking about, you know, Sydney Sweeney's Instagram feed or something like that, but now they're talking about Claude Mythos. And you're just like, what is happening? Reminds me of like 30 years ago when, you know, Katie Couric is on the Today show being like, what is Internet? Can you like, can you tell me what is Internet? You think, Oh no, we've come so far and not. Come.

Uh huh. Literally, yes. It's crazy. The other thing that I guess it's somewhat related to this is I think you were the first person I saw posted yesterday, was the thing where NIST said that they're only going to enrich specific vulnerability reports now that apply to government agencies. Because they literally, I don't think they said it specifically, but I think you can infer that the reason is there's just too much. Now they even the National Vulnerability Database, which you know, says it right there in the name. It's supposed to be a database of vulnerabilities. Yeah.

Doesn't have the resources to keep up with every CBE that's coming out. Yeah, that was really interesting. They said starting in, well, starting this week, they're going to prioritize CV ES on SIS Kev for enrichment and then also, yeah, software used within the federal government, CB ES found there and then CB ES for critical software as well. Yeah, which makes sense. But it's also, again, like leaving out, you know, a lot of different organizations that that need it.

Most I would say yeah. That was interesting because obviously it is AI, but then also, you know, like there's been a lot of, there's a lot of interesting context there as well. I think it's with NIST and just the nature of the NBD and I. Don't know Money. Money. Funding all that, yeah. Yeah. So I think unfortunately it's it's like on top of the funding issue, like this AI issue is just like complete chaos. It is and there was something that I forgot to mention in the in all the Project Glasswing stuff last week. Anthropic also put out the separate blog posts that was essentially here's kind of the defenders point of view of how this is all going to happen or whatever. And they had some recommendations in there. And there were two that caught my eye. The first one was it literally says this go patch everything that's in Sis's Kev immediately

go like, I don't know if you can pull it up right now, Linds, but I don't know how many bugs your insist is Kev, And I don't know if you can just find out right now, but it's thousands and many of them are years and years old. Like they added a bug this week, an Excel bug from 2009. So I don't know like who's exploiting that? Like, congratulations, but if you like, if you haven't patched that or if you're running a vulnerable version of Excel from O9A, what are you doing?

But you, you probably don't have the resources to patch it. That's why you're running it. But just the like, arrogance isn't the right word. But maybe like my, I don't know, myopic view of like, just go patch this. Oh, why didn't I think about that? Our friend Matt Johansson was like, oh, why didn't I think about just patching everything before? What a dummy. Like, OK, like, like, we were talking earlier, like, just eliminate your attack surface.

Just just don't have bugs. Totally logical. Why didn't I think about this? Unhackable. Yes, unbreakable. Yeah, don't be God and. So true. I understand the thinking behind it like first address the known vulnerabilities that you have in your environment before you worry about this giant pile of bugs were about to shovel onto your plate. Understood logical, right, But the way that it was written, it just boggled my mind. I'm sure it was written by an AI, but the the other piece of advice in there, and this is kind of like really interesting to me because I know these guys, but they and recommend organizations using the EPSS system, which is the exploit predictability scoring system.

I think I'm getting that right from that was developed by a group of guys, Mike Roitman, Ed Bellis and Jay Jacobs, three of the smartest like data security guys you will ever find. And I mean, like data is in like, you know, quants and like statistics and all that kind of stuff. And it essentially it is what it says, like they use this really complex statistical model to figure out what's the likelihood that this particular bug is going to be exploited over a given period of time. And you know, the anthropic post I think says pick a number and then go patch everything above that number or something like that.

But those guys reached out to me earlier this week and said, hey, we should do a podcast about all this that's going on. I said yes, we should. So we're going to do that next week. So that'll be fun and. That sounds really interesting because I do think that the nature of how patch management is going to play out and like the severity of things like it's going to completely change with when there's that much noise being added to the mix. And I'm just waiting and this is probably already happened and we just haven't heard about it. For somebody to turn one of these models loose on things like ICS and OT technology shit that just cannot be patched because it can't be taken offline or you can't reach it or there's not a new version because the software, you know, is 25 years old and it just works. And nobody else, no one knows how it works, but it runs, you know, an electric plant somewhere in Tennessee just like, well, OK, that's great.

There's a bug in and we can't fix that. Yeah, I'm sure that's happening. We, I don't, I haven't seen too much noise about ICS or SCADA stuff yet, but somebody will get there. Yeah, I think that's kind of what bothers me about a lot of about this news is that, you know, obviously with OT and ICS, there's so many challenges that we are nowhere close to fixing already when it comes to IT and OT convergence. And like all the problems like legacy hard legacy software on OT systems and things like that.

We're just nowhere close to that in, in the OT space, let alone the, the IT space where companies are already struggling to catch up to and patch vulnerabilities that, that have come out. You know, you mentioned it like you've been like over a decade ago. So why are we, like I, I understand the benefit of from a defense perspective of being able to unearth vulnerabilities in your own product, but like it makes me a little nervous that we're throwing out models that can find vulnerabilities to this level if they're if, if we haven't even fixed the fundamentals like and we don't even understand businesses can't even roll out patches like, you know, effectively yet.

No, be a little nervous. Yeah, it absolutely does. And there's plenty of organizations that they can't really address those basics. You know, they can't roll out MFA on vital systems. They still have identity management problems. They have a version of Excel that's exploitable from 17 years ago, things like that. And I saw John Oberhide, the one of the founders of Duo where we used to work, posted this week like we've never in this industry, like bug fixed our way to security.

Like that's not something that is scalable as a way to just solve security. Yes, you have to patch bugs, but just finding more bugs and piling them on a pile of other bugs and then trying to fix all of those is not a path to security. It's just it's really, I do feel like in some ways all of this LLM based research is a little bit of like, hey, look what I can do with this shiny toy and admiring the problem in some ways instead of being like, OK, that's all great, but how can we use this new toy to fix things or make things better? I'm sure we'll get there.

I'm, you know, people will do that. But right now it's just a lot of like vroom vroom. Look at my cool thing. Yeah. Yeah, it's like that's what boys do when they get a new, new toy as you. Yeah, exactly. Yeah. It's like if we can figure out how to use AI to like stop, you know, threat actors from coming in to like expose RDP instances or. Something that'd be great. And I will be impressed. Yes, I know. Or just. Like the most, you know, like, I don't know, that's it's a good point too, is like that vulnerabilities also are such a small part of like the broader attack vectors that threat actors use. And it's a lot of like the boring stuff really that doesn't honestly make the headlines like business e-mail compromise attacks.

There was like totally like the FB is IC3 report like every single year. That's the top category for like the top amount of losses. Yeah. Like that would be a very valuable place to focus the the efforts. But you know, I think it's like you said, vulnerabilities, shiny new tools like it's. Oh yeah, make a little more sense there. Yeah, I mean, yeah, phishing still works really well. People still click on links. People make mistakes, you know, like you see all these stories every once in a while of like car thieves cloning, you know, smart keys to steal cars.

Guess what? Go ask cops how most cars are stolen. People leave their keys in cars or that's how they get stolen. Like most of the time. You know, yes, there's lots of that cloning stuff going on too, but like most threat actors are not chaining together like 7 different vulnerabilities to exploit a 27 year old open BSD Colonel bug. They're sending you a really well crafted phishing e-mail or they're doing that crazy shit that you wrote about a couple weeks ago with the Drift Protocol. You know, compromise.

We don't have time to get rehash it here, but go read that story. That thing's bananas. Like that's a months long intelligence operation. Like obviously that's an outlier, but like they didn't need a vulnerability. They just, you know, met up with people and like did old school spy shit. Yeah, Yeah. That's a really good point. I think AI can help there when you're meeting up at a conference. No, I hope not, because I really like spy movies and TV shows, so I need those to continue. Yeah, yeah.

All right. Well, I think that's enough. AII think we did it There are there. It was other security news, but it's all on our website. You can go read it. There's there's some cool Mac malware that I wrote about. There's other things happening, but it just kind of gets overwhelmed by the the AI stuff. Yeah, yeah, definitely. It's it's all day, every day AI. It truly is. Yeah. Yeah. All right, Linds, thanks so much. Great to see you. Have a good weekend. Thanks you too.

See ya.