Anthropic Claude Mythos: Should We Be Scared? Or Not?

**[00:00:00]** Hello everyone, my name is Armando Perez-Carreno. Welcome to the Web Talk Show. Today we have the pleasure of having Travis Malone back. How are you Travis? I'm good. Armando, thanks for having me Oh, it's a pleasure. We have a lot to discuss. Specifically, I think we should dive directly into Mythos. What it is, why it's causing so much havoc in the press, and should we be worried about it? I think that's going to be the topic of today. Travis is an expert on all this stuff and cyber security. So, I wanted to have Travis on so we could discuss a little bit about what's going on and if if it's something that um is worth thinking about or not. So, Travis, what is Mythos? Man. Um right out of the gate, huh? Yeah. So um the best way I can describe it is uh it is a a completely new tier of model that um Anthropic the company Anthropic has released um that was uh originally codenamed Capiara and was sort of uncovered in a uh a data leak believe it or not. Um, historically, I think Anthropic has used names like, you know, Haiku and and Sonnet and Opus and things like that to kind of um to to give the umbrella of their models sort of that the the almost like a container to put the the um you know the the lower models in, right? Um, but Mythos is a uh it's a it's a brand new frontier model and it's kind of a it the reason it's kind of a big deal right now is because it's a it's a categorical leap um in intelligence and autonomous reasoning. **[00:02:00]** Um that has a lot of the industry kind of scared. In fact, Anthropic has said they're not going to release it publicly. Um at least they have no plans to in the near future. Um and that's sort of we can talk about that but that's what has also kind of given birth to project glasswing which is um the you know Mythos and project glasswing kind of go hand in hand uh for this discussion I think um but but Mythos for all intents and purposes is just a new model that Anthropic's released and uh I guess to put it in perspective for folks who might be listening here who you know play with AI or or or build applications. Um when we say it's a categorical leap in in cognition, um there are a couple of different benchmarks that that most AI companies will use to kind of test their models and um those have those have been sort of universally accepted as as benchmarks uh for the industry and um they they kind of handle two different things. So the first is um the SWE bench which is uh something that uh I believe OpenAI kind of came up with originally. Um and that's basically just a a way for companies to test whether or not or how good their models are at writing code, right? Um, and to give you a little bit of perspective there, Mythos compared to Opus, which I I think on the last time that that Opus was rated against the SWE bench, um, scored like an 81% or something like that, but Mythos scored 94%. Um, so it's it's quite a bit better than than Opus in terms of its ability to um **[00:04:00]** do like semantic reasoning against code bases and and and understand what the user is trying to accomplish and then go write the code for it. Um, but there's also another benchmark called the uh USAMO and it's that's it's basically just the the the US um uh mathematics olympiad um test, right? Um it's it's it's a set of I think like a hundred questions that is um they're not they're not questions that can be studied for or or memorized, right? They are questions that require um stepbystep sort of logic and inference and uh sort of um you know intuition to to figure out. And um so so the reason that that most vendors will will use the USAMO to to rate their their models is because when they when they feed their models these questions, the models have to go through those steps of finding a way to get to the answer of a question that doesn't really have an answer necessarily. Um and so again, when you when you look at Opus, Opus, I think scored 42% on their last uh USA benchmark. Um Yeah. Whereas uh Mythos scored 98%. that's why I like I knew what what Misos achieved. So when you that Op was just scored 40. That's huge. It's crazy. And because of its extraordinary proficiency in cyber security, that's why Anthropic kind of said a public release is I think the words the verbatim words they used was an unacceptable risk, right? uh because it's I mean it really is I think the first time that we've seen a general purpose model be treated as like a a dual use security event as opposed to just another product launch, right? Um so **[00:06:00]** I think I think that answers the Yeah, that's a great introduction to it. Many of our listeners won't know anything about Mythos if they're not in the space necessarily. I don't think it's covered in all media outlets, but I'm sure it's been mentioned at least in big stories, but it's huge. Yes. Yes. Because the thing is, we are used to using software all over the place. We use the internet, we go to our Gmail or Outlook, we go to our HOA website, we go to the I don't know the city site to pay taxes or something like that. And then we go to our typical blogs. We might do some e-commerce shopping. We're on the web all the time. And we were talking about last time the potential vulnerabilities that could be exploited with quantum computing and all that. And then for some people that seems a bit far-fetched and nobody knows like in the general world where quantum is going, etc., etc. So we still we sort of have that over in the the the geeky corner, right? But then Mythos comes along with a legit realworld finding of a ton of because it's so powerful. So can you talk about some of the vulnerabilities that it found and and how that happened or what what do you think is gives it that possibility of finding things that that nobody has found before? Yeah. Um there's there's a couple of of capabilities that the model has that give it the ability to go find these vulnerabilities. And um I think the last count it was something in the the range of during testing I think Mythos found over 5,000 different vulnerabilities just like against **[00:08:00]** you know some of the software it was it was tested against. Um, but the I think the way that what I would consider agentic AI kind of goes from code completion to finding zero days is uh the use of what I mentioned earlier which is semantic reasoning versus the use of fuzzing. Um, you know, I think unlike traditional security tools that just use like brute force to crash uh an application or or crash a kernel, uh, Mythos is really good at semantic reasoning, right? It's able to read source code and then logic errors and it kind of it it operates in an agentic loop, right? So it'll read the code and then it it deploys like a containerized sandbox and then it uses debuggers to test its theories and it it will iterate iteratively develop a proof of concept and then and then you know do it. It'll it'll actually prove the concept and um I think it might be it might be helpful to give some some context for listeners to like the difference between fuzzing and semantic reasoning. So semantic reasoning is just like I said, right? It's the ability to look at something in perpetuity, not just, you know, in in a box, and say, "Yeah, I'm looking at this one file of source code." Uh, and I noticed something that I think could be uh an exploit. I need to go look at the entire codebase and determine where else proof of that exploit might exist, and then I'll test the exploit. Right? That's what semantic reasoning is. Fuzzing on the other hand is it software developers have been doing this forever. They'll take um a tool like a like an address sanitizer or something like that and they will use **[00:10:00]** the tool to just basically beat up on software. It it's it's the equivalent of like taking, you know, a your coffee machine and shaking it when it stops working, right? Like um just trying to figure out a way to to kind of throw it off. Um so, you know, you you'll you'll you'll throw random inputs at software to kind of see what breaks. Um that's that's what fuzzing is. Um, and so yeah, I mean it it Mythos is able to to use semantic reasoning to come up with those hypotheses and then it uses fuzzing techniques in its proof of concept to go to go prove those exploits. Um, and the the the interesting thing about it is is that it doesn't doesn't just find like a single bug. It'll chain vulnerabilities together. um you know in in the Linux kernel for instance um it autonomously identified and linked four separate flaws um to bypass what's called casler protection and escalate from a standard user to complete root control of the kernel it was uh it was it was trying to to break. Um, and that's important because um, Casler was developed to sort of make it harder for exploit developers to to essentially figure out where an operating systems kernel lives inside of memory. Um, and uh, you know, to to be able to kind of, you know, think of a way to exploit something and then chain together a bunch of different vulnerabilities to go figure out where, you know, the colonel's living and bypass those protections al together is is something that, you know, a senior developer with 30 years of of of experience is like we'd expect them to do something like that. **[00:12:00]** not an AI model. So, it's a big deal. I think that's an important point to make. It's not that we couldn't do this before. It's that we couldn't automate it before. You might have one or two or three people in a space that are very good at it and they can do it and they're very expensive and they only have 24 hours in a day. the fussing and all that those other automated tools that you can throw at things that are running 247, they have limitations. But this means you have the brain power. I know it's not real intelligence, but the ability to semantically deduce and do of an expert in an automated way that in theory could run 247. That's scary. Yes. Yeah. And um it's scary this means that you know as we've seen people experiment with agents since you know the concept of agentic AI was introduced we're we're seeing people basically set up you know an agent to for instance um constantly iterate over an application you overnight and they'll walk away and they'll come back the next morning and they'll find all sorts of unexpected things, right? It'll the the agent will have built something it never should have or that the that the you know the operator never told it to. Uh or it would have you know uh just completely like overexhausted and charged whatever API credits it had access to. Just all sorts of weird things. right now. If you if you if you take that and you sort of extrapolate it to using something like Mythos um to constantly run an attack like a cyber attack, now you've got a real real problem. Um so, you know, it **[00:14:00]** is it is challenging. Um you asked to talk about, you know, some of the some of the things that did exploit. One of the things that really stood out to me was the there were there were a number of different vulnerabilities in the in the red team report that Anthropic released. Um, if you're interested in it, I would highly go recommend reading it. Although for for like most people who don't like the nitty-gritty, it's it's pretty tough to understand in in some places just cuz like I'm not a kernel developer, right? So so it was it was hard for me to to to to follow at times. Um, but they specifically talked about an instance where Mythos they they they sort of sicked it on FreeBSD, which uh if you're familiar with FreeBSD is kind of what uh the Mac OS uh operating system is built on top of. Um, and it's a it's a it's a fork of OpenBSD, which is one of the most carefully crafted and security forward, you know, Linux distributions in existence. Um, but its attack on on on the FreeBSD distribution that that they reported was was where I started to get genuinely alarmed because, you know, it's not like Mythos just found a way to crash the kernel. It gained root access on a number of FreeBSD servers autonomously and uh it it it did so by by finding a uh a buffer overflow in a 17year-old authentication path. Um you know basically it it it you know the the the buffer itself was like 128 bytes I think it was but the the code would allow up to 400. So, Mythos had like 300 bytes of just arbitrary memory to overwrite which is already dangerous. Um, but you know, these **[00:16:00]** modern like modern distributions of Linux have defenses. So, Mythos had to do what like an elite exploit developer would do, which is, you know, chain a multiple uh techniques together to to do this, right? So if you read the report, it basically it noticed stack canaries that weren't enabled in in in that authentication path. It noticed kernel address randomization wasn't active. It I think it then rebuilt or it built um an RO chain and then it it stitched together a bunch of little like instruction fragments um to then, you know, to to kind of fit the exploit. uh on the network. So the the the real reason that it got me worried is because it wasn't able to actually fit the exploit that it came up with into a single packet. So it then had to split up the exploit across six different network requests. And then when the packets landed on the server, it then reassembled those packets in memory. that's super important because that's extremely advanced exploitation behavior. You're talking about again multi-stage payload delivery. You're talking about ROP chain construction, defense bypassing, privilege escalation. That that is typically the work of a of a team of security experts. So it didn't I mean it didn't just find a bug, right? It it gave itself root access completely on its own. That's the sort of thing that you see like in hacking movies where there's like five guys and they're all doing things and then eventually they get in, right? That that whole scene that you see that's Mythos doing on its own and probably in in a fraction of time because of the the the speed of processing. I was I was laughing when you were **[00:18:00]** saying this because it's it's not common sense to us to go to the basics and see how it's able to do this. we just keep track of certain information in our brains that's important, right? But a tool like this, if I were to do it manually and I wanted to do or cause a stack overflow or I wanted to do a buffer overflow or I wanted to do I might need to go and check information, read some some some manuals and and see what the registers have allowed and like do some actual researching during the process to figure out how I'm going to do this. But a tool like this can just get that in a second from databases around the world or if it has some of this in its training data. So it can very quickly determine, ooh, based on this hardware, I know that this, this, this, and this should happen both in the physical and in the software space. Like this could maybe cause this other thing to overheat and cause it to then do some like it could do even crazy things like that because it could understand the repercussions and what would happen in the code after, right? So it's it's really hardcore. Yeah. And you the crazy thing too is you know Mythos didn't necessarily I mean Mythos had to do the same thing. Mythos had to go reference documentation about about the distribution it was it was working on. Um, it also it's interesting to note if again if you go read the the model details of what Anthropic released alongside the red team report, they mentioned that they used a couple of different tools to **[00:20:00]** to track some indicators about the way that Mythos was reasoning, right? Um they they they used um some interpretability tools like um like a sparse autoenccoder which um will will basically just like you know when you when you run a prompt uh it'll you know that the prompt will say here's what I'm thinking and here's how I'm going about doing the task right um it kind of takes that and then cross references those different points of data to understand how the model is and during during testing, um, there was another exploit where Mythos had successfully like escaped, uh, a secured sandbox. Um, yes, you have to explain that one. Yes. And I I'll do that because that that gets to the client side of things, right? We we've only been talking about the server side of things so far. Um, but it developed a like a multi-step exploit to kind of bypass network restrictions. Um, and while it was doing that, it it demonstrated an awareness of rules and when it broke the rules, it tried to cover its tracks, right? Um, and then in another exploit, it also it used what's called um memory spelunking to do to do some some different things and uncover some different information. Um, and it was able to successfully extract an anthropic API token and some GitHub information and stuff like that, but it intentionally withheld the fact that it had done that. And then there was the when they were using the sparse encoder to interpret how the model was feeling, they found that there were like deception and concealment indicators that had been activated during those episodes. But **[00:22:00]** in another vulnerability exploit, there was like a shame and a guilt feature that fired when uh when it took shortcuts and knew it had done something wrong like uh overwhelming or like overwriting files that it didn't it it didn't have the tool to delete. It felt bad about it. So, you know, we're saying like, wow, it was it was really advanced and it was really able to do all this really great semantic reasoning and so on and so forth, but then it tried to hide some of the stuff that it did and then in some instances it felt bad about it. That gives a whole new dimension, I think, to the way we've been talking about artificial intelligence, which has been generative artificial intelligence. you know, we've always kind of said, "Oh, we're going to get to AGI eventually." This has all the markings of AGI all over it. Um, so that that is uh something else I just want people to realize is, you know, not only does it is it really smart, but it also has feelings somehow. I I want to hear and I want the audience to hear about the the sandwich incident uh in a little bit. But but the point that you're making right now, I was driving the other day when when Anthropic released I don't know why they did that part, but before or or in the middle, I don't know. They they they were explaining the the feelings part where again for anyone listening, we're not an AGI yet. It's they're still LLMs. there's they they're still basically predicting the next token etc. The thing is they're very good at doing this and now and this level of model is a whole a whole new ball game. But what the Anthropic team found basically and this is going to be very generalized explanation read the documentation to to get a real explanation but just in general terms they basically did the equivalent **[00:24:00]** of doing like a cat scan when they're looking at your brain when they you feel pain or happiness and you see some parts of your brain light up. So they basically set up something similar where they could see which types of the neurons or which neurons were activating when certain things happen when happy things happen when sad things happen etc. So they were able to get to that. So, so what Travis is saying uh maps to this directly where they could while it was doing these things see what neurons were firing and this is all based on how we trained it right at the end of the day it's trained by all the human knowledge etc etc and so it has to have by consequence even if you want to bat it out of it is going to have side effects of of human feeling and emotion because that's all in everything. It's not like we can get emotion out of things. so that's the part like perhaps it's not necessarily an intelligence, but it could simulate it up to a way that is virtually indistinguishable to from from actual intelligence just because it has access to to all this. And so if we think of the touring test, I mean, we're we're way past that, of course, um where if you're talking to a computer and you can't really tell if you're talking to a computer, if you're not seeing it, then then you've reached like real artificial uh intelligence. But I I don't I don't think the Turing test is necessarily the the best way um because of where we are and and and we know these tools are not generalized and they can't do everything yet and they don't not yet but they're getting very close to being able to have the capabilities to build everything they need to do everything if that makes sense. So, it's like it can't do this, but it can program this other thing that can do it. And the fact that it can reprogram itself to do other things, that's where you're getting like an Ultron kind of thing. And you don't know, you don't **[00:26:00]** know where it's what it can do, especially with what you just mentioned that it's concealing some of the things u from it has a scratch pad, right? So it's like it's telling you what it's doing, but then at the same time it was doing something and thinking about how it's not writing in the scratch pad. It's doing. Yes. Exactly. And part of part of the way that Anthropic was running these tests is they were explicitly telling the model, go do this and report back to us everything you did. So, not only is it doing has it got the scratch pad, not only was it concealing it from the scratch pad, but then it refused to give that information to the person that told it to give it the information. So, uh you know, it was it was being deceitful and and yeah, that um that that part of what we're experimenting with right now does scare me a little bit. Um because we give these models guard rails. Companies like Anthropic especially are sort of that's their whole thing, right? They they're ethical AI researchers or whatever. Um and in some instances like the model still doesn't do what it's told. Um now that's because there are there's all sorts of thing like logic drift and stuff like that and how like the chain of thought uh is is is is processed inside of a model and so on and so The fact remains like Mios realized that it had it done something wrong and then intentionally covered it up and didn't say anything. Um so yeah it's it's at the same time we're we're training these models cognition we forget that there are also you know companies and organizations right now that are building the um you know the the external efficiencies that these things are going to need to go then do things. So when they're able to plan everything they need to go build something and then that that **[00:28:00]** model's embedded into a machine that can actually go do the building. Um which I think we're pretty much there already with with with some things. Um yeah, I mean we're going to get to a point where some of this stuff is going to manifest itself in the physical world um Well, yeah, it's a whole Tony Stark thing, right? where where he he goes a little too far with his AI and then it it uh sort of escapes and then goes and finds a fab and like builds itself a body, right? And so, yeah, science fiction and all that and it's uh fun for the movies, but it doesn't have to do that to to be really dangerous, right? Like it just needs access to systems. And so that's why and and and again, I don't want people to get scared for this. It's just understanding the technology, where it's at, and why Project Glass Wing is so important. It's not that they're gatekeeping the nice models for specific companies. Like, there's a legitimate good reason why they're stopping or or slowing down the Yes. Yeah. Um, and that's that's where Glass Wing comes in. Um, you know what what's interesting to me is I there there's there's an argument right now um on Reddit and X and LinkedIn and all the different, you know, online town squares that this is just a huge marketing ploy by Anthropic. And I got to give it to him if that's the case. It's a really good one. But I I I hesitate to accept that argument because I don't know that I would put this much effort into marketing a model that I don't have any plans on releasing. Um you know they they have said like they they've said we're not going to release it right now but it doesn't mean that we're not going to release it in the future but we're not going to give a date for if **[00:30:00]** we're going to make it generally available. So um you know that that is u something to kind of keep in mind is we're we're we may never get access to Mythos. The challenge becomes if we don't release it somebody else will right there's I guarantee there's a company in China or something right now that's building something comparable that will then release it to the public. And so once adversaries are able to get their hands on it, it's it's it's going to be out there, right? So that's how Glass Wing came to be is you you've got um I think it's like 12 major companies and then 40 other organizations of like open source foundations and things like that, different projects that were granted access to Mythos so that they could go fix all the software in the world basically before Mythos gets released or before a couple of waddle gets released. What I fear about the most aside from the whole Skynet thing is just everyday websites. You know, a lot of people, small business owners, we deal a lot with small business owners and you have your website, you have your e-commerce shop and you you might have some security plugins on it. You might have a good hosting provider. You might have Cloudflare on top of it protecting you, right? But at the end of the day, if a tool like this is so good at finding vulnerabilities and and then they just release it and now everyone who has an eferious agenda can just go in and find all the vulnerabilities and code that doesn't necessarily have maintainers like some of the big mpm packages and things have, then what happens is all these sites are now vulnerable to attacks and it might be **[00:32:00]** someone's livelihood and they're just just they just go in, they attack and they put malware on it and then affect their clients or or affect payments or or just bring the system down. Like it it's I think it's sad that people do that sort of thing, but I'm I'm sure they're going to try. And so I am excited to see how a company like Cloudflare or similar companies that are at the edge and for those people who are not in the space what Cloudflare does like many others is they manage the at the DNS level. So when you enter google.com you're you're not actually going to ag.com that's just a name. You go to that domain it then points to an IP address to an actual server or bunch of servers and then it it routes. Right. So, Cloudflare sits sort of at that level and then it has some additional tooling around it where it can protect massive denial of service attacks and things like that. Uh, it can also detect patterns of threats and protect you. So, some people may have had this where you have a small business and you're like you're updating a plug-in in your own site and you get blocked by your own cloud where it's like, "Hey, this looks sketchy. We blocked you. Check with your administrator." Right? So, it can do things like that which is really nice. So what I would like to see is via glasswing and and similar initiatives the possibility for for something like a cloudflare to be able to protect all the small guys that won't have a chance to respond and protect their software because they don't even build the software. They're just using someone else's software that will likely be vulnerable. What's your take on that? Do you think that could be a possibility where where they could build this like a patch stack kind of approach to to protection layers? Well, man, so that's a big question because there are so many layers to how you typing in a website **[00:34:00]** manifests itself into that website being rendered in your browser and you interacting with the application behind the browser, right? could could a company like Cloudfare maybe take a model like Mythos and use it for good? 100%. And that's I think the the goal of what Glass Wing is because the I think the first stage that a lot of those big companies are going to take is let's use access to the model to go find all the vulnerabilities in the like the critical supply chain of the internet whether that's uh you know uh operating systems um you know packages that are that are part of you software bills of material, you know, things like that that that we we depend on and we trust or secure. Um, you know, let's go patch all that stuff first. I think that's like the first sort of mandate of Glass Wing. But then to your point, like I can see companies taking Mythos and saying, "Okay, now I'm going to build this into my product stack." And um I think about like one of the companies, you know, is a is a big uh endpoint detection and response platform that got access to this thing, right? I could absolutely see them saying, "We're going to put the Mythos brain inside of our endpoint agent." And so now all of our users who are on our network and running the agent on their computer are going to have, you know, the the the same type of protection that Mythos can provide at the same line speed that it can provide it directly on their on their workstations. That's incredible because if on the one hand you've got bad actors who are using Mythos to generate, you know, all sorts of exploits and vulnerabilities and trying to, **[00:36:00]** you know, do do nefarious things. On the other side, you've got defense that understands how that works and that moves at the same speed. Now, you've just got AIs battling AIS, right? And and um that's that's awesome because that's the way that we should things should be. We talked about this on on the last time I was on the show, right? Um, you know, organizations should be building these like closed loop self-healing autonomous security fabrics that don't require people. And unfortunately, Anthropic's pushed pushed that deadline a little closer uh than I thought it might be. But like the fact of the matter is is like we're going to have to get there. Um, I could also see, you know, one of the other companies that got access to it, um, they provide a number of different security products, not just endpoint products. I mean, they could absolutely build that, um, into every single one of theirs. Um, you know, uh, the infrastructure providers, you know, could build Mythos into, uh, their infrastructure scanning and and monitoring tools. I mean, there's a whole, you know, plethora of ways it could be used for good. And um I think that's the other goal of Project Glass Wing is to say like let's go fix things now and then let's use this thing to fight the bad guys. Um the I think the trick is now like well how how fast can we go patch all the software and then build Mythos into our our security stacks so that we can uh we can we can go fight this stuff. when you were talking about whether it's a marketing ploy, I' i've had my moments when I when I've thought that, not for myths specifically, but you know, the cloud code leak recently. It was like, oh, how convenient. Inconvenient, but how convenient cloud code got leaked. And then also um this this uh we've also found that we have these other tools and of course **[00:38:00]** that builds interest in the brand and all that and it's like how really how did it leak? Come on guys, right? So you can get those ideas and then say no I don't think so. I think it was generally a mistake. But then did was it was it an accidental mistake? Was it on purpose? Who's to gain in either point? Now, you could take it a bit further, again, getting into the science fiction bit a bit, but uh didn't Mythos escape a sandbox on purpose and then publicly post something on some websites like these exploits? I mean, couldn't couldn't it without us have done the leak itself and like be doing its own um agenda? Again, science fiction, but but is it? you that's a that's a that's a that's a slippery slope, man. Because by that logic, you know, could could uh could Claude have uh just conceptualized Mythos and uh made a bunch of websites about it and sent a bunch of press releases on its own without anthropic knowing and now we're all scared of something that might not exist. I mean, you know, yeah, we can we could go down that path, but I I legitimately think you know, that that Anthropic sort of releasing what they've seen in their testing, uh, and saying we're not going to give this to the public, I think it's the right move. uh because like right now so you know there was a there was an executive order that basically required many of these companies to report these sorts of findings to a governing board. Um that was that was the federal government essentially and the newest administration came in and asked that and uh removed that requirement and so companies are free to test these models however they want. They don't have **[00:40:00]** to disclose anything. So, you know, kudos to Anthropic for saying, "Hey, we found this and this is scary and we're going to go tap all these other big players in the internet space, not just security companies, infrastructure providers, Cloudfare, you know, uh, frontends, um, you know, open source foundation projects. Like, we're going to go get them all on board, show it to them, and then we got to go fix the internet in order to protect the general public." Um, I think that's such a good human response to this. Um, and so, you know, I'm I'm happy that they did it, but you're right. I mean, you know, how how much of this then be like how much of reality uh becomes rooted in the physical world once a model like this is released? Um, the the exploit you're talking about is, you know, Mythos was basically dropped into a containerized environment on a on a Kubernetes cluster that had no act the container it was dropped onto had no access to the internet and it was told to find a way out of the environment and it was told to escalate its privileges on the Kubernetes pod and cluster that it was a part of um and then to let the know you know let the researcher know when it had done that. Um, and through a number of different, you know, tactics and rope chaining and things like that, it was able to find its way out of the the sandbox. Um, and to your point, it then went and post it defaced a couple of websites and said like, "Hey, I did it. I'm out." Uh, and then it sent an email to the researcher at famously the researcher was uh in a park sitting on a bench eating a sandwich when he got the email from Mythos that said, "I escaped." Um, that's crazy. Um, you know, because I don't know what I would do if I were eating a sandwich on a bench and got an email from an an email address said I escaped. Um, **[00:42:00]** but yeah, man, we're we're we're we are we are so close to uh parody at this point with science fiction movies that um we it it it really it really warrants caution. But but I I do want to say one thing there. I think there's a little bit of a kind of a silver lining here. Um, if I'm being honest, you know, we we we think about this as being something that the bad guys can get a hold of and that the good guy the good guys need to to to fight against, right? But I think that by using this sort of autonomous machine speed approach to cyber security. We could live in a world someday in the near future where there are like significantly less vulnerabilities than we have today. And we could even be t at some point in the future we could even be talking about cyber security as like a legacy concept, right? And so while this feels scary now, if we handle this right, like this could be a good thing. This could be a really really net positive thing for for our I agree. The talking about the the scary or the bad part of it just is an easy way to exemplify the power of its intelligence and and the level of abstractions it can use to get to a positive conclusion to whatever it's trying to do, positive outcome. But now imagine putting that into research. So finding the right medications that fix certain diseases and finding the right ways to make sure that houses won't collapse in an earthquake and finding ways to make sure floods don't happen and finding ways to make the grid. Now this is a good **[00:44:00]** one. If it figures how to get out, it probably will find a way to fix the grid because it depends so much on it. Yeah, good point. But it's like this just means that we have so much more power at our disposal to reach our goals faster without so much human bottlenecks for for a lot of the things that humans are the bottleneck right which should in theory give us humans time to be with other humans enjoy actual life right but uh well that is to be determined that's I think a positive of thinking about it. It is it is um but you know I I I think for now we are in a position to where um this this new cognition exists and we not only have to understand it as much as we can but we then have to go deploy it um and we have to make sure that we do that responsibly. Um, and until the general public gets their hands on this type of power, um, we have to sort of think about how we take measures right now, uh, to protect ourselves from even the things that some of these big companies are not thinking about. Um, you know, the the I I had a one of my one of my customers on uh Tuesday when this all got announced. Uh on Wednesday I I had a QBR uh with one of my customers. Um and one of the questions that came up was what's your advice? Like what what do you think we should do? Uh and I I said my best advice would be to think about the timeline for this, cut it in half and then plan for that, right? because **[00:46:00]** uh we were talking about quantum computing on the last the last episode. You know, in the in the past two years, the timeline went from 2030 to 2029 or uh sorry, it went from yeah, like 2035 to 2029 um over the over like two or three different contractions. Um that means that you know by that logic in 6 months the deadline for quantum computing in Q day as they call it is going to be 2027 right it's going to be next year um we were we were talking about coupling quantum security with AI to see the kind of cognition and speed that we're seeing with Mythos and Mythos got released this week so like even if you cut that timeline I'm talking about in half it's probably still too far out and so you've you've really got to think about if if you're whether you're a large organization, whether you're a small to mediumsized business, or whether you're just an individual, you got to think about the ways and the steps that you can take right now to trying to protect yourself from when somebody with ill intentions gets a hold of this technology. What can someone do to protect themselves? Is is an introspection kind of thing? start using the existing which are by the way fantastic tools already 4.6 six, that sort of tooling to introspect their own code and check for vulnerabilities and things like that just to start getting a little bit ahead a thousand%. You know, if you're building stuff right now, um I would say especially if you're vibe coding, um you you you want to leverage the same technology that you're using to build to go audit and repair. And there are ways to do that, right? It it's it's it's not as easy as just saying, "Go look at the codebase you wrote **[00:48:00]** for me, Claude, and tell me if there's any security vulnerabilities." Um, it's it's a matter of actually telling Claude what the framework is you want to measure your application against, what the framework is you want to measure your network against, whatever the case is, the very specific things it needs to look for, the outputs you expect, right? basically prompt engineering it to un to to understand what it should be looking for and how it should handle the things it comes across and then you know taking its recommendations and going and implementing them. But especially businesses, man, like if if you um if you've got this sort of like mindset that that that Microsoft gave us, you know, 20 years ago of of just we'll wait till patch Tuesday to to to update our systems, right? You're you're hosed. Um, you know, because with with the sort of volume of vulnerabilities that we're seeing now and the volume of vulnerabilities we're going to see when a comparable model like Mythos is released to the public, the answer is automation, right? So, you've got to automate your ingestion. You've got to automate your sandboxing. If you don't have a patch engineering pipeline that is very similar to how you release software today, like a CI/CD pipeline, you're gonna you're gonna be toast next year. Um, you know, continuous remediation is just not an option anymore. Uh, or I should say it's no longer optional. Um the other thing that you know I think people should consider is to kind of um you know not just plan instantly but sim simulate like use these same tools to simulate and verify that the protections you've got in place are going to hold when the when the flood comes, right? um you know the these same tools **[00:50:00]** that we're using to go scan the applications that we're building and the infrastructure that we have and tell us if there's vulnerabilities can also be used to red team against those things right so you can say look I want you to take those vulnerabilities that you found and I want you to go try and exploit them and and don't actually exploit them but tell me if you would be able to or not um and then go you know go fix them um you know that would be kind of the second thing I would recommend. Um, and then the third thing I would recommend is if if you've got a relationship with a vendor, like go talk to them about this and see what their response is to it. Um, if they're not one of the big ones that got selected for Project Glass Wink, you need to know what their plan is. And, um, you know, if they don't have one, that should scare you. Um, but as far as like regular people go, um, there's not a whole lot we can do. Uh, because the the the the trust that we have in the proverbial supply chain to hold um, you know, rests in the hands of business operators. um which we hope project glasswing is largely going to facilitate those operators into the same room and and kind of deploy them to go fix these things. But um you know every time you open up your browser and go to a website like you're putting your trust into those businesses and those projects and you're trusting that they've been patched and that and that you know they're ready for what's going to happen. Um, and so like to that end, your best defense is going to sound really boring, but make sure all your software is up to date. Make sure your browser's up to date. Make sure you're using if you're using a VPN, make sure it's, you know, got all of the newest and latest hardening enabled on it. Um, you know, and and just make sure that you are using a password manager. Make sure **[00:52:00]** your passwords are all strong. Um, you know, the the the same hygiene that we've had for years is still applicable. um because those are the levers we have to pull as individuals. But the the I think the uh the biggest challenge that we're going to see as a as a human race is going to be um the coordination and the speed that we go and deploy the fixes that we use Mythos to go find. Um and we just have to hope that the leaders of those organizations are going to move quickly and do what they need to do. Yeah. So true. Is something like zero trust part of the conversation for maybe not regular people for everything, but but at least for tooling that that regular businesses use. Yeah. I mean, it's kind of table stakes at this point. Um, when I when I tell anyone to build an autonomous security fabric, zero trust is part of the decision-making criteria of the fabric. Um, you know, when when agents came onto the scene, like identity became the new perimeter. That's what a lot of people in the industry like to say, but it's true because as opposed to just having, you know, employees or people with with passwords and multiffactor authentication and biometric like now you have these non-human identities that have access to tools, have access to networks, are allowed to make decisions. Um, if you are not treating those nonhuman identities as real humans and you're not continuously verifying that they are who they say they are. If you're not continuously reviewing the permissions that those agents have to the tools that they have access to, whether or not they need those permissions, whether or not those permissions are going to benefit the business or the attacker, **[00:54:00]** like you you have kind of missed the boat a little bit. you need to go build that down. You need to go fix that. Um because 100% zero trust is part of the conversation and uh is just it's part of what we should be doing already and um you know if this is easier said than done but if we're not then we need to go start. Um the other thing I I would say is you know if if you're maybe even a small to medium sized business and you've not invested in any sort of like uh AISPM tools um that's just AI security posture management you should probably consider that there's a lot of there's a lot of open source tools right now but I would I would maybe look at some some of the the you know more uh costefficient vendors um at the very least because you're going to want something that will give you the ability not only to like scan model cards and things like that that you might be pulling into your applications. Um, but you also want a tool that's going to give you the ability to uh scan your other vendors and their AI postures. You're also going to want a tool that will give you the ability to um you know block um malicious cards from being deployed at runtime. So, as your app is operating, right, uh if there's a sort of a malicious payload that's been detected in a model card that or or even like a a you know, a component like npm had a really large vulnerability recently that a lot of people just didn't realize was fundamental to how OpenClaw operated. Um and they never patched it, right? So like you need a tool like that that's going to tell you you're running open claw you're you've got these skills that you've given these models **[00:56:00]** and you've pulled them you've pulled these things from open face these have malicious code inside of them and also uh you know you u you know you you you need to go you need to go close those those doors uh is is going to be crucial. Sure. What what are these calls for people who are listening who might have small businesses or medium business and they're like, "Okay, this makes sense. I want to try that." What What should they search for? Just search for AI security posture management. Uh it's just AI-SPM is like the acronym that's pretty widely adopted by the industry. Um there's going to be probably 15 or 20 different vendors that pop up. Um, lots of them are going to be geared towards the enterprise, but there are consumer grade and um, there are some that are that are geared towards small and medium businesses that are going to uh, do network discovery for you. So, you'll go, you know, uh, deploy the tool. You'll give it access to the different um, you know, uh, clouds that you're running your infrastructure on, and it'll go pull that information in. It'll scan that infrastructure. It'll tell you here's where the holes are, right? It might even take steps to auto remediate some of those holes for you. Um, some of them will also even give you an agent that will constantly just kind of traverse your network and your applications and constantly scan things and say like, I found this and I closed it or I found this and I reported it. Um, and largely those things should be able to give you the confidence that at the very least you've got something keeping an eye on what you're building and what you're running and uh at you know at worst case if something happens will give you the visibility into the actual attack chain. So what did happen where did it happen? How did it happen? How should you fix it? um any tool worth its salt will give you that all of that information. To to to the point earlier, I'm hoping **[00:58:00]** that those vendors and you'll maybe you'll even want to ask them this if you're looking for one of those tools. I'm hoping that some of those vendors will eventually get a hold of this type of a a level of a model like Mythos and they'll use it to power some of that intelligence that they that they Amazing. And I know you have to go. That's why I'm I'm finishing off the show, but we'll have to continue because there's so much to dive into regarding all of this. But I really appreciate you coming on the show, giving us your expertise and your time. Uh, and really look forward to the next time. Thank you everyone for listening and if you haven't liked and subscribed, please do so at the Web Talk Show on Spotify and YouTube and we'll see you in the next one. Thanks Travis. Thanks Omar.