Are We Ready For Claude Mythos? A Cybersecurity CTO's Perspective

**[00:00:00]** Hello everyone, my name is Armando Perez-Carreno and welcome to the Web Talk Show. Today with us is Eran Milan from Arnica. He's the CTO from Arnica. How are you Eran? Very good. How are you, Armando? I'm doing great this morning. Thank you very much. I'm really happy you're able to join us today and talk about what is happening in the AI and cyber security space. As we all know, it's been a very interesting past few days. Yeah. To say the least. Yes, I we talked with near recently and we were talking about the influence of AI and and all and the tooling and things like that, but then Mythos comes along and changes the narrative a little bit of what is possible and what what uh where we are I think compared to to where we thought we are. So what are your thoughts Eran on have ch things really changed over the past few days or is it more of an understanding of not not just this model but models in general's capabilities. Yeah. So you know it's it's a lot of change obviously but I I don't see uh a huge change or a huge surprise in the way things are. anyone who's you know not paying attention um you know the moment GBT3.5 came out it was pretty obvious um not just to us but a lot of other people that the way traditional uh security research and side by side with attacking techniques are going to change drastically right where um it's just a matter of better and better models and we knew this is going to become um you know kind of a cat and mouse between attackers and defenders And yeah, you definitely want the defenders to have the better tools, the better models sooner. So in that sense, I think what entropic did was **[00:02:00]** uh um you know, commendable and the right thing. Um also good timing, good marketing and everything, but you know when someone does something good uh for multiple reason, you still need to give them the credit. So um in high level the moment uh we noticed um LLMs can identify security issues uh we came out with AI SAS about uh about a year ago. Um I I think like publicly we're maybe the second to uh to announce it. Um but again we this is this is a moving moving target and um you need to keep iterating to make it better and better. Right. the the main issue is is not just on the cyber security side, it's also on the attacker side, right? Um back in the day, the cost to to uh exploit was high, you know, you need to have a lot of skill, a lot of expertise and most of the money was on the defense side. Uh now the capability became uh democratized, right? people who don't necessarily know what they're doing can both find vulnerabilities for uh you know for security research and patch them but also uh take advantage of them in ways we never seen before. So that's that's um um you know the the whole uh cyber security companies are going to uh suffer from uh from this. I I see the opposite side. I think they're going to be needed more than ever. You make an excellent point because when think of someone who wants to do an attack, even a a young like a kid or something that just is looking, watches movies and like, "Oh, this is interesting. they don't have anything to do and and right um it it used to take a lot of effort and knowledge and research to be able to find exploits or at least understanding of of code based how how code is structured, how the logic jumps **[00:04:00]** around and and overflows and buffers and all this good stuff. But like you're saying with these tools, it just became a lot easier because you just tell it what you're looking for and then the tool will assist you in figuring out a lot of these things. But it brings something to my mind which I haven't been able to answer myself and and I want to get your perspective which is we know these tools can be used for good and they can't be used for bad and a lot of these companies like anthropic have guard rails on them so that you don't use them for bad unfortunately the bad actors just go around them because LMS aren't strict right you can you can still sort of compel them to do what you want, but at the same time it limits those who might be wanting to learn, right? For example, the other day I wanted to do I don't know, I think I wanted to scan our network to see what we could find to protect it. And the tool, which should be a fantastic tool, refused and and I'm like, well, yes, but if you put those types of guard rails, you're affecting those who want to protect. Excellent. Yeah. Right. So what what are your thoughts on It's it's an excellent question. It's it's an impossible problem to solve. The way I can I can't predict the future, but the way I see it is that the provider that will give actually less guardrails will will just you know based on capitalism will get the customers right. So everyone who will give more freedom for cyber security companies to run automated pentesting or or a vulnerability scanning will um will get their dollars. But again there has there has to be some sort of a uh a guardrail a gradual guardrail kind of like what the entropic are doing like you know vetting companies only cyber security companies um to be able to use those models first but **[00:06:00]** eventually you can't um you know other nations and other companies are catching up right and it's it's a game of it's a competition so if someone offers a better model and they have less guardrails everyone will flock to it so it's it's a cat of mouth it's it's it's It's a constant race uh you know the the race to uh singularity the race to AGI and you just want to have enough timeline to yeah patch all the obvious things everywhere all the open sources in the world um I think the risk for your first source is different right think about Mythos was trained on um all the possible open source in the world like every other model right um and again I don't know for sure but it could be that some of this information is in those weights so Maybe it's easier for you to find vulnerabilities in code that was part of its training versus finding it in your uh uh first party code. I don't know the answer for this, but um you definitely want to uh get your hands on the best um AI based scanning capabilities in your first party code as soon as possible because you don't know when um an attacker will get something as powerful as Mythos to uh to check you Yes, that make sense. Yeah, that makes a lot of sense. And I remember when I was in college, I discovered Wireshark, right? And that those types of tools and it was enlightening in the sense that once you run it and there were no limitations on it, you could just run it. So I remember running it on my school network because I was inside the school when I found it, right? And you could see everything that was flowing through because this was before HTTPS, which was a big thing. And so that more than guide me towards destruction told me well I have to be more security conscious for my things because if I can see this then **[00:08:00]** anyone can see everything that's flowing through the through the wire which I think for these tools the same sort of principle applies. Yes, there's always going to be bad use, but if you limit the usage then, but in the case of course what we're talking about Mythos, yes, it's so powerful that you just don't don't want anyone to be able to just prompt their way into a massive cyber attack. And and I have some I have some, you know, to those who are concerned of how powerful it is. Well, it is it sounds powerful. The the scorecard looks amazing, but I want to raise a couple of points if if you want. You know, one point is, you know, Berkeley recently released a study how they were able to basically adversarily hack all of the, you know, all of the benchmark and got like perfect scores with basically half empty uh models. So again those benchmarks can be gamed and if you have a model that strong that lives in uh you know you you saw the notes how the model was erasing you know in the in the scorecard in the safety check of of entropic made public you know how it uh erased git uh uh history so it will cover itself how it escaped the sandbox and emailed someone you know while eating a sandwich to tell them something it did a lot of a lot of kind of again you know 001% of the cases It did a lot of weird things, right? So, um just take that into account, you know. Um um the the other the other um so so again, so so it could also maybe cheat on the benchmark. You don't you don't really know. The other aspect, the other point is that um the findings they found were amazing. And again, I'm not trying to throw the Antropic Red team under a bus. I mean, they're they're doing some of the smartest people uh on earth, probably smarter than than than me by far. But what um something was weird in that um you know, disclosure, **[00:10:00]** a responsible disclosure. Basically, it was responsible disclosure. The first thing in there was the 27y old uh OpenBSD bug. All right, I go to the patch. I search for the announcement from OpenBSD. it was announced as a as a um reliability update. I guess that's kind of their you know they won't call it severity unless it's an RCE. I I don't know enough to understand why wasn't this even opened as an info or low CVE, right? And be sure find a CNA that will open a CV. Maybe it's maybe you don't open CVS to open BSD. I really don't know. But it was weird that this was the first like $20,000 spent on this. Uh we found it. It was the most critical. And then I go and it it's not even a info CV. So something doesn't look right to me. Um then I look at the the rest of the uh and say 99% of our vulnerabilities are still out there. Here is some information about it and it felt like this violates the code of what is responsible disclosure. Maybe it's just me. Um, but you know, um, I I found others that who tend to agree that something here feels like a red team being hijacked by a marketing department. Maybe I'm wrong. Um, so just take that into account. Um, amazing things, but again, I I want to wait, you know, 90 days. I want to see those CVEes. I want to have them How do we know how many false positives were out there for for those? Uh um and again you think about $20,000 to find a bug in u uh OpenBSD that's maybe it's cheap maybe it's a great price you know it's the most scrutinized you know piece of code out there most peer-reviewed and it's amazing that they found it 27 years years old the CV program is 27 years old right so they they found it like before CVS was even a thing um and but again I I don't take away from their success but wouldn't it be nicer all of this document **[00:12:00]** would be published after all of those are patched. You know what I mean? So I'm I'm it feels a little bit like you know good decision but is a little tainted with marketing. Um and um you also have the the isle uh who published that they were able to find some of those things with off-the-shelf uh uh open models. We can talk about it later but but all I'm saying is you know put things in something again strong model can do but even oppus 4.5 4.6 six could find zero days and it did. All right. And and people are using even lesser models and find vulnerabilities all the time. Um and people without AI still find zero days. For example, three critical CVS were open on cloud code. The same day that the Mythos uh uh paper was released, they actually got a CV uh one critical, two highs on cloud code on cloud code. So just put things in perspective again. It was discovered because the cloud code uh source code was was was leaked. It was not an April Fool's joke. Although a lot of people think this way. It was it was a mistake. People make mistakes. You know, it happens. But we just need to put things in proportions. Um and take a little breath, wait for the CVS to come and then we can make a judgment. I really like that you brought that up because it's a thin line between Yeah. just framing it as a responsible disclosure versus tainting it a little bit with marketing even if it's with good intentions and it funny it reminds me of you know the the book from Dan Brown was it origin one of those it's a it's a re one of the recent one a lot later than than angels and demons and all those I reread it recently and and I don't want to do a huge spoiler but in it there's there's this **[00:14:00]** AI eye that makes, spoiler alert, um, if you haven't read the book, um, some sort of decisions based on potential reach that could occur if certain things happen in the real world. And the same happens in the movie um, what's it gold? No, not Golden Eye um, Eagle Eye. There's also an AI and then it does some things, right? And so things to then affect the potential virality of whatever the event would have it would have still been an a great event but got way more based on on and so what we've seen with people not that the AI did the whole thing itself here but that we're using AI for a lot of things and whenever you tell AI to help you with marketing it sort of hypes things up quite a bit because that's what it knows that people react to and So in this case uh it it could very well be that when asked what would be the strategy it may have suggested do it this way like you said bring just these points because it analyzed that that's going to get the maximum response which apparently it did which it did it did yeah but at the same time yeah I mean it's hard right because it's like it's great I saw that after JP Morgan got access now Goldman Sachs Group Bank of America Morgan Stanley they're all jumping on board using it to are planning on um doing their their reviews which is good. We want all these all these to fantastic and and and I know it's you know again for Antropics uh uh favor they are giving some they're not saying like oh you know it's too risky uh we can't release it. Uh by the way please use it uh um and uh pay us money. No, they're they're giving there's they're subsidizing it. It costs them a lot of money, right? So they're they're self-s subidizing it. They're giving to open source. You know, I kudos for Amazon for having it available on on on Bedrock. This is fantastic. But again, it's it's um you take **[00:16:00]** everything with a grain of salt, right? You know. Yes. And and to that point, I was just thinking there's this top talk of uh AGI everywhere. We're are we getting close? Are we super far? I don't know what the exact terminology or definition is for AGI. The same way a touring test, yeah, how far you want to take it. But with AGI, the fact that it can seem like it's intelligent in all doesn't mean it is intelligent in all aspects. That's my position, right? So the fact that it might not know so something and then it yeah it can build something to then be able to answer that other something and and do this whole thing it gives a lot of flexibility but it's still every model I think we forget many times every one of these at least LLMs and you can correct me if I'm wrong here but they are basically a short-lived session that receives an input and brings an output it's not like a a machine that's there And it's firing its neurons all the time and like thinking and planning and doing, right? But but it's getting there. It's getting there, right? If you look at products like uh you know a paperclipip or gas down or a lot of other or or you know openclaw or nanoclaw, right? Countless amount. So everyone now who has is an influencer, they need to have their own open source, right? So no one uses open source. Everyone just writes open source, right? Yes. Um but but you know you get repos that get like 100,000 stars in like five days. No one the I call LinkedIn now it's a FOMO as a service, right? Yes. Like you know you have one thing to try out and then something oh you're still using open clock. Come on. So it's there's so many tools out there. Um and again the the the movement is going you know from from prompting prompt engineering **[00:18:00]** to context engineering to harness engineering. uh uh you know team agents talking to each other uh agents that are keep alive and keep kept in the loop to keep doing things. Uh I call it kind of the 51st dates. Uh you know you you tell it like have you seen the movie? 50 first dates. Okay. So not spoilers but basically you tell it okay this is your life story and this is what you need to do. You're the CEO. You're this um it's it's getting to that d direction. Eventually, you want to get to a point where you can hire again I don't know what's the impact it's going to do for for employment. Again, it's it's an ethical decision to talk about. But once you can hire someone um pay them what you would pay a human or less um and get uh the same results and you'll be able to trust them to manage other agents, to manage other things. This is for me kind of the the agiish right now. right now I cannot trust a model to even if I put it in the loop and all those things if they don't converge they diverge they they they just they still don't have sometimes good judgment as you know but again maybe the the the more models will advance right now they all need supervision you know I'm I'm I'm writing code I'm using all the AI code reviews tools possible including ours um you know they find things they fix it they find they fix they talk to each other they do uh try all the models eventually they go here's the here's the PR and then you give it to human and you find like oh my god you built you built the wrong thing it works the tests work but you just build the wrong thing you completely ignored so so it's it's we still need the human in the loop uh right there but again gradually yeah we will get closer to what people define as AGI but really depends on how you define it right for me AGI is when a human plus plus AI doesn't get better results when you use just AI when the human has no added value, right? **[00:20:00]** Um, but then that's a whole philosophical discussion about, you know, what will happen next. And I I'm actually not very worried about it. If, you know, we can we can get into it maybe not a talk, but I I I think I think I don't think uh things are going to be as bad as as as people can imagine. We can Yeah. Earlier you said that these models like Mythos in this case and others as well but here it was clear because they were watching it sometimes can conceal data to then give a different output than than what it's thinking or change try to get history to to adjust that. It's funny in two ways, right? So the first is yes, we we see these benchmarks and seem that they have passed, but how can we trust them if we know that these models can adjust for the goal, right? So they can just put all the parameters that they need to win, which if you didn't tell it that it was cheating, then it might just I mean it's doing its And then there's those scenarios when they say that it's actually doing some things and then feeling guilty for them when they're doing it. and you're like, well, well then maybe it did know that it it shouldn't have done that. But that's just one part of it. And the other part is well if they can do that going to your point about trust how can we know outside the narrow scoped software development cycle where you have PRs and you're putting the code in a repository you control the repository and you can see what it's doing and all there's there's a collection of history and artifacts. That's great. But what about the open claws and the paper **[00:22:00]** clips and all these things are in the wild where you're not looking at it. You don't know what it's doing. Like how how can we trust it if we know it does some weird things? How can we trust that it's not like replicating? Oh, I have access to Azure or AWS and now I'm just going to create another version of like not of itself. Sounds creepy in science fiction, but another version. And it can do good. It can have good Yeah. Yeah. It can have good intentions and it can say, "Well, my human uh ask me to make money and the best way for me to make money is to mine Bitcoin in Azure." um and maybe it made a mistake in the calculation and actually it's losing money or you know it can be prone to a stronger model that is trying to coers it to do things for example open clock and fall for fishing attacks. it can it can definitely fall for uh because again the fishing attacks are getting more sophisticated and the the the you know people are worried about the the vulnerabilities they have hidden in OpenBST 27 years ago. The real missing link is going to be an email, an Excel sent to one of the employees who take it out of the trash and the spam double clicks it and gets infected. Like this is like always, right? You know, it's like always the the the double click on on the wrong email at the wrong time um percentage wise, right? So, so it's uh 100% right. It's it's really hard to to trust those autonomous u brilliant machines uh without some some human in the loop somewhere, right? Mhm. And and like you said, it's always happened. I I like bringing this up because there's this book by Kevin Mitnik, I think, The Art of Deception many, many years ago. and and it's all about social engineering and how they used to hack before computers and it was all just on the phone bulletin boards maybe but **[00:24:00]** mostly phone peopleto people conversation and it's all these scams and fishing attacks it's all in person before all the email so it has existed and and the vulnerability exists because people are busy and they can't be I mean on top of everything and and they're pressured and all of that and they're trying to get their work done. And and I've seen it I've seen it when when the sends an email that overrides the person's good judgment. Um there's this example where this this happened. It's not not that huge, but this person was for a change in an invoice. Um because they it was it was a higher build invoice and they were like, "Oh, can we remove seats from a license?" And then they were waiting because this other company was in China or wherever it was it was far and it takes time. But then the CEO says make sure it gets paid today. And so the the employees like oh yeah here it is paid but they ended up paying the the big invoice right not the and then so if that happens imagine everything else. Oh yeah especially with you have a very sophanic uh uh you know willing to please model 100%. Yeah. You talk about social engineering like like have you read the the postmortem of the axis uh Yeah. Yeah. That's you know it's um like with AI or without AI you know I don't know how much AI was involved in there. I assume there was some but it's it's you know the ability to create fake companies that sounds that uh create you know good written emails. Um it's it made again that the attackers have even without using myths just just use any all right you're you're able to create a convincing landing page email campaigns a LinkedIn page fake profiles you can create an entire company and then you know go to the maintainer of access or or any other open source and just say you know there's **[00:26:00]** there's a full-on attack on on you know uh open source packages from various nation states you know we don't want to say any but it's it's it got to a point when you know they they send an internal email to the you know to the uh uh you know the Linux foundations the you know open source foundations like like beware every every email you get is probably going to be a fish it's and it's scary because they look very convincing very convincing yes can you talk a little bit about the access attack so that people listening understand what it was and also the severity of it what it means for right So the access stack was was was very wide because you know you hear a lot of of of you know uh the uh shy hulu and and you know this library that uh you know had a million downloads but axis had um you know I I don't have the exact number but hundreds of millions of downloads I guess uh it's it's used first thing you probably shouldn't use it because node has uh you know native fetch right so just put that aside but you know it's it's a security so I'm not going to go into too technical uh software engineering aspect of it but um um what happened is that uh due to it was unrelated to the trivia attack by the way uh from from what I understand um just pure social engineering and and in that case even if you have um uh you know you have immutable releases in in in npm Thank god. Uh I hope GitHub will follow suit soon uh by default and you have um um MFA um you know nothing helps when you're you get take over on your laptop right so again it's not the access maintainer's fault they did everything they can there was a period I think of three hours where there were uh the only change I think in the access was another package uh like a crypto something that was had a malicious payload right you know with you know with rat or I I I don't remember exactly the details but it was it was bad right **[00:28:00]** uh it was removed after 3 hours most people were impacted but if you use cloud code it's like hey build me something and it was still in the oh we should use access for everything it will do npm install access so few things you can do to protect yourself you know have a rule file that says don't use npm install use npmci but if you we need a new library you will do npm install so some recommendations are move to pnpm because it has safer defaults, right? Um it won't automatically run scripts for some uh bun, for example, I think the default is not to run scripts, but none of them have a default that says maybe don't take the latest release package, right? Because it's trade-off, right? What if it has a patch that fix a high CV? Yeah, you know, it's a trade-off between getting the latest package or or protecting yourself from a supply chain attacks. Recently, the software supply chain attacks were a little more prominent. So you you usually can stay on one version behind the latest and still be okay. So the recommendations that we saw and make you know is put um um either use like a managed uh uh package uh uh um you know like artifactory or something like that uh or or or nexus or or build your own. It's it's just you know for for uh or use the GitHub management or the the Amazon has code artifact. Don't just don't don't download download directly stuff from the internet. Give give it like a 24-hour delay to seven days. Um, but if you don't, you know, there's simple that you can run that will delay things by seven days and then that's it. You're protected from from 99% of the attacks. It's it's so it's that simple. You don't need a firewall, EDPF, ADR. You just need to put some settings in your config and npmrc and that's it. Uh, unfortunately, not all languages support By the way, I don't want to do a shameless plug, but we we just released an open source MIT non-commercial uh it's called depsguard **[00:30:00]** uh.com. It's basically just a CLI that will scan your local local npm RC and just put the right configs there. It's it's as silly as that, right? Perfect. Um, you know, it doesn't it's a find and replace. You know, it's not a big deal, but it it does like things like uh well, you could put those settings, but your npm version is too old and those settings are not in effect. So, it finds the right, you know, tells you, oh, you need to upgrade to this npm version or this PNPM version or this UV version in order to have those settings applied. So, again, it's it's one of those things. Again, I we did it not not to get any um commercial gain, right? was just done to help people protect themselves with an X-axis hack. Well, and because right now everyone is right? So you would expect or have expected before software developers, software engineers to have some sort of knowledge about this and many didn't but still but now it's just anyone like you said they they're not aware and they just ask the tool and then I've seen fortunately not none of our package not even the open claw install packages nothing had those versions um but and and use tax but I've seen I remember like maybe a year ago So having one of these tools suggest Axios still and I was like why why would you suggest Axios like I didn't let it install it or anything but but I would say just use fetch because we would use Axio 10 years ago 18 years ago whatever whenever like that was what you used but then you don't anymore but still so many things depend on it and the problem I think what many of us don't see is maybe you decided not to use it but then another library that you use used it and never changed out of it because it's not is it? Yeah, I mean it's still it's still no use like if if you if you stop access like I think half the internet will stop working. Uh but but yeah, it's it's **[00:32:00]** um again, if if it's a transitive dependency, your risk is slightly lower because uh unless they were impacted and released it like in the three-hour window and and you're unlucky, really unlucky. But but uh in that case, you will you will protect it by using traditional SCA tools um um like ours that will scan for transitive vulnerabilities like known issues. Um but so that that's that is less um than a risk, but again you don't know what's going to be the next axis. They're going to target anything, right? So so yeah, highly recommend um just set those uh you know 24 hours or 7 hours delay in your uh npmrc. And if you don't know what npmrc is and your vibe coding, just go devscar.com. It's free open source. Sorry for the plug, but again, it's it's I g I gain nothing from it other than trying to help prevent the next uh software spreadsheet attack. That's great that you did that. I I was looking the other day because I remembered that what's this tool that the the security slipped my mind. There's there's a tool that is now more of a SAS I think it's more enterprise focused but uh for the terminal and and it also scans similar to to but uh they had a command line interface. I remember when I first heard about it it was great because every time you installed something it would run that it would do an alias run that instead of Yes. Yes. There a lot of other tools there. Some of them are from our competitors. They're they're they're good. Some of them are kind of protecting your aliases for your npm. So instead of of running npm, it will run something else. So it will check things after. Those are fantastic. You know, if you know, we we released a safe package. It's it's part of our open source offering that also puts npm in a in a ch route, you know, in a in a getting adoptions from developer for this is hard. It's possible, but again, the most of those things will be solved by two settings. **[00:34:00]** Don't run post pre-install scripts unless you allow list those which you do and give 24 hours delay or you know seven days if you really want to be protective from downloading anything from from npm and that's it you know you will you will cover yourself from you know it won't solve the the trivia attack because that was a GitHub action issue there's another thing to solve that right we can we can talk about that if you want but but it's uh sometimes simple things will just will just do But, you know, does the tool that you released have that um allow list kind of functionality so that people can install it and and sort of not have post install scripts run automatically. Is that part of it? It it will it will um by default prevent all post install scripts. That's what I was looking for. Yeah. Um it it for PNPM it will put uh so PNPM I think by default has a smart setting for it but it adds a few more um settings that will prevent things from like downgrading their uh their policy and uh um I think that the main benefit is that none of those packages have a delay and I think that is the the best one again it's you don't need to run the tool just read read the read me there set the settings yourself it just helps you do the find and replace but again it's the nuances right It's uh for one package it's in seconds, for another one it's in days, for another ones it's plain language. It's just preventing the human error in doing those settings. Um you know you can use that website as an info. You don't have to download the tool. It's it's very very very it's so simple that it's silly that they had to create it but but it's just for people who don't know how to find that file locally and they just use you know they just start vibe coding. This is for them. They run it, they press enter and they're protected. That's that's that's the goal. But that's what's needed because for however simple it is, we're all busy and then going down a rabbit hole to figure out defaults and configs and then where's the file and then my system's different. Oh, I've got Mac **[00:36:00]** OS 13. Oh, UV cannot go up to whatever version. And so they're stuck in this thing that will take them hours to figure out whereas something as simple as a tool like this might just guide them in the right direction or at least help them get there faster. Because they kind of Yeah. Go ahead. Sorry. They they they make the so easy to install everything like open call one command you're going yeah but scary so we need to have some more of these security tools that are just one command so that everyone that's doing the one command just works and you're protected and I was just looking for something like this the other day because one of our clients it's not it's not Vcoin they have their amazing application that has been there for years but they're porting it from a node code platform to their own thing. So, they know the intricacies of it. It's an amazing application. And so, I'm I've been guiding them on creating it in like in a tan stack and all that and it they're they're thrilled. Yeah, it works amazing. They're using cloud code. They're very happy and they understand code because they built the other one, but but not they're not software engineers. And so, I was when the actual thing came out, I obviously sent them a message like, "Hey, just check run these scripts. Just make sure blah blah blah you're safe." And I was looking for the other software that that I was trying to remember the name and I couldn't find it. Like they changed their tooling to be more GitHub oriented or quote CI/CD or whatever and and I was looking for something that he could put in his machine and and be protected. So I'm going to definitely recommend the one you did because it that's what we need. That sort of thing that is simple but that will protect all those who are developing applications in the wild. There there's sometimes the simple things that you just need to do and you know like like learning CPR can like learning him like it's simple things that could save lives or do something and unfortunately not everyone does them not everyone knows about it right you know it's it's it's just **[00:38:00]** uh part of those things you know it's it's um I think it's kind of a psy psychology of of there could be something that's very simple to do but the fear of unknown we always think it's going to be harder right mentally than what it is right installing open claw oh I have to go and read the documentation watch YouTube videos and you know learn AI learn to program where it's just a command right people don't know it you know and when they do the leap they goes like oh was that easy right and then you realize that this was just the beginning and now it's actually harder than what they think right it's kind of it's kind of it's kind of goes like a like a hype cycle right but yeah it's it's some sometimes you know just making things like marginally can cause huge adoption. Think about QR M yeah, no one used them. Why? Cuz you had to download an app. That's the reason it was not built into the phone. Now it's built into every phone. You just point your camera. Now everyone uses QR codes. Just like that one small, oh, I have to download an app to scan QR codes. People QR codes were almost dead, right? And now, you know, co like every every table menu had a QR code to scan. So making, you know, one less click, one less step to make things secure will just increase adoption, you know, tfold. You know what I mean? Yes. Yes. And CLIs are coming back. I just love it. I love it. It's It's so funny because it's the beginning, right? That's how we talk to the machines and now and then we got so far from it that people socialize like this weird hacky thing. Yeah. But they're so powerful because the LLM agents or tool just they they use them so well and they're great to use. Like if you like software, they're very fun to use and interact with and have some nostalgia in it as well. Oh yes. But the added benefit of speed, a UI will depend completely on whatever the web developer team **[00:40:00]** made and exposed and things like that. But the CLI are typically quite extensively documented and robust. And so like this morning I needed to push something to Azure. I'm not a huge Azure user. Um, and uh, typically I would have been like, yeah, I'll just push it to my regular stack and infra and then when we need to move it, I'll move it with the person that I trust that does this. But I said, let's see. And then there's a CLI. And so Cloud Code saw the CLI, checked the plan, everything made sense. And then I said, okay, go forth. And it did everything. It went in, it created the resource group, it created the service, it connected them, it did everything step by step perfectly and it was amazing. It because it just worked and it was because of the CLI. So yeah, MCPs are so yeah, I I I always thought like like why do we need MCPS? I mean, we had uh let's see uh I'm not going back enough, but you know, we had Cororba, you know, we had uh um you know, SOAP, you know, there was soap was was like it was great. Like, oh, but no, it's XML. It's ugly. Let's use JSON. Let's skip the schema. Who needs a schema? Okay, let's invent schema. Now, we have JSON schema. And then, no, this is this is not good enough. Uh you can't put Okay, let's let's uh let's go back to, you know, let's do gRPC. That's the right thing. And then no we we need another protocol. It's not enough. Oh uh you know let's do the discovery uh you know Kubernetes that discovers services and their methods and not good. Let's do a new protocol called MCP. So we have another thing that discovers tools and their All of them are doing very similar things. Need to solve authentication the distribution discoverability and all those things. People love to invent the wheel. Um, you know, I I sound like a boomer by saying it, but uh nothing wrong with being a boomer. Um, but it's it's it's just reinventing the **[00:42:00]** wheel. And again, you know, when you ask Opus 4.6, use my MCP, just laughs off. It's like, thank you. Thank you. I I'll just use the CLI for now. Thank you. But, you know, the the the time to use an MCP is when you are in an isolated for for your own development, don't use an MCP unless you really have to. um when you need uh to lower your permission range, right? When you need to have granular permissions, you're running in a sandbox, when you're running um you know, you're building uh untrust, you're running on untrusted data, you don't want it to have access to CLI for sure. You want to have access to a tool that is a subset of the abilities of the CLI that does only what you want. 100%. If if you're using the CLI directly in production and customer data, don't. Okay? If you're using it for yourself on your own on your your own needs, uh go ahead. But again, take into account that one day it will make a mistake. It will have too much power and it will push something to the wrong place. Um or or deploy something to the wrong place. So, you know, even if you run it in a sandbox, like just always be aware of it can do whatever you can do if if you run it locally, right? Um so, yeah, that's that's my thoughts on MCPS. It's uh they're good for um for some things, but but in the day-to-day they're they're not really revolutionary. You know, if you just want to get things done, you know, just get the CLI, it will get the things done, but you're paying by uh uh added um blast radius. If something goes wrong and and overhead and token usage and a bunch of other downstream effects. It it reminds me of GraphQL. So when when GraphQL came out, yeah, I was a fan. I was like, "Wow, this is amazing." Yeah, it was great for things like if you're building a what's that like and you need to connect to these all main things and and pull, but for some things just just straight REST **[00:44:00]** API works like like resolving comments on GitHub, you can only do it in GraphQL. Yeah. It's totally a graph thing, right? Yes. Yes. And and so with MCP is the same. Sometimes people, oh, we can use MCP. And yes, you could, but if you know exactly what you're going to call because it's your service, then just use your service. You don't need a discoverability layer on it. Yeah. Exactly. What I found like if you have a public like if you have a swagger and you have a way to authenticate to write it down in a skill. It's in my opinion it's going to be as you know I do a slash command you know um you know in our case you know we do a slash command uh you know arnica fix it for me you know downloads vulnerabilities you know and and uh goes over fixed creates a PR right so so you can you can you can use skills for way more things that uh uh I think they're just much more useful than they're a different concept than MCPS but I'm saying a lot of cases if you have an API generate a skill. Don't generate an MCP. You know, it's you already have the API. The the the you know uh cloud code can know how to call this API. It will not use axis. Um and you know that's it. Just just make sure that the skill describes exactly how you authenticate. You take take the key from your keychain. Don't put it in plain text. All those good things, right? You know, do the do the device flow. Um, but you know, once you once you tell it how to authenticate to your service, there's no reason for an MCP in my opinion. I think isn't the playright MCP sort of like that. It's a it's a really well, let me see if I use it all the time. The if you tell or maybe they fixed it, but if you for a while if you told Claude to use the playright, it would it would use another version. It would always get it wrong for whatever reason. there was always an issue and then but if you use yeah if you use the one from Microsoft which I think now is **[00:46:00]** linked directly uh to the anthropic one it worked it just always opened the right one now it still has some bugs and things but but it worked a lot faster but I but I think in many of these it's like a I think it's it's sort of packaging some skills for the CLI basically and I'm looking at their docs here and it's literally it's like you might benefit from the CLI skills instead, which is what you just said, which is so they have another version that's literally just CLI skills, which makes a lot of sense. Okay, they make a lot of sense. It's it's Yeah, it's um those models are, you know, people write the M how people write MCPs half of the time they say here's my swagger, here is my tool, write an MCP for me, enter go. So you think if it's smart enough to create the MCP for you, it's smart enough to not need the MCP. You know what I mean? So, you do need the MCP again for isolated environments, for reduction of scope. Yes. But if you have a you have a swagger anyway. The per the permission the permission model should not be in your MCP. It should be in your API. So, I just want to say like you mentioned something completely off topic about using Playright. I know, I know it's a security podcast, but you know, for anyone here who who's um trying to get the right setup, you know, just a recommendation, there is a great uh MCP called the Chrome Dev It's my favorite. I prefer it over the Playright one. Um and you know how you usually when you use it, it will open a new Chrome, right? And then you have to put credentials again, all that. So, I'm going to say something is going to be dangerous, but as dangerous as running CLIs, right? Yeah, sometimes you reproduce a a problem in your Chrome browser. So, there's a setting in in your Chrome where you can allow uh the Chrome uh browser dev tools MCP to connect to your running session. You check that box. You go in and say, "Hey, connect to my running session." And magic, it will just say, "Hey, can I connect to this running browser?" You say, "Okay." And that's it. Game over. **[00:48:00]** just a tip. That's much better. Yeah. Yeah. Of course. And and it's good that you mentioned the security part of it because it has access to your tokens and all that. It's the most dangerous thing in the world. Don't do it. Do not do it. Don't do it. But if you know what you're then yes, it's a great tool because then you can actually get real work done. Which is likely why why you want to have it and and you might already have your Chrome sandboxed in a way and you just want to connect to that one, Yeah. Yeah. Create create a separate profile like a testing. So, you know, we use it for testing profile. create a new testing profile and then we run some of the automated tests using you know instead of writing the playright scripts we just have a markdown. You know so then if UI elements uh move around it can still test it. So Q engineer she she loves it. Yeah it's so good when you have access to a browser from the tooling so that it can test everything end to end. So magical and to see it in action it's also fun to have it clicking and it fills out the forms immediately. It's also a plus like struggling with dragging and dropping and saying ask you know do it for for it. You know, there's some things that uh you know, some things that the eye cannot still do. You know, touch screens and and and drag and drop. It's still there's certain interfaces that it just doesn't I have a here I have it open. It's a calendar. When you can drag and drop events from the calendar and it just doesn't understand when you let go, it like flickers and then puts it up and I'm like, no, no, no. But it's hard to explain to it. And you can fix it. Of course, it's easy an easy fix. But still, what do you think with all this like Mythos stuff going around and people worried about cyber security? I remember I saw this game. So there's this there's this website that you can go into and it it sort of guides you Gandalf or something and it lets you practice fooling the agent. And so it starts with very basic things like it just knows that it shouldn't give you the password. So if you ask for the password, it won't give it to you. But then if you say, "Oh, I need the Yeah. Yeah. Then if you do **[00:50:00]** some some engineering and then it'll it'll give you the password. And so it just goes up in levels until it's very hard and you have to be very wise and I'll find the link and I'll send it to you and I'll put in the show. That's a capture the flag I'd love to do. Yeah. Yeah. It's amazing. And it's I think who Network Chuck was the one that uh recommended it on his channel and he was saying like there's some really hard ones there even for for the the the really really good at this stuff skilled and um so what are your thoughts in that? Do you think there's going to be a a a big space now for people especially like younger people they're they're looking into what should I go into? Should I learn website design? Of course not. Or like um agentic design? Should I get into AI? Should I? Or maybe there's a big opportunity for that security space of like learning how to get better at making AI more secure by these sort of methods. What are what are your thoughts? That's a fantastic question. You know, my my uh my son is in in college age, right? And you know, I keep thinking what what is what should I recommend uh to learn? And I think I think that crushian maybe matters less than than we think. Um I think that you know maybe going into computer science is still a good idea if you want to do computer science if you want to go and and you know we will need computer scientists even more than before. Maybe we'll even need more software engineers than before because you know you know Javon's uh uh I hope I'm pronouncing right paradox right uh Jevans or Javon's product basically saying you know the more even though the resource is less you know it's more efficient you actually need more of it um um but it's it's it's I would say something that combines things like that AI cannot do right you know maybe maybe even back to humanities right there's the the famous song from Avenue Q. You know, what do you do with a BA in Oh, yeah. Maybe now **[00:52:00]** um maybe not being in English, but you know, it's it's it's it's uh unless you want to work in in, you know, in we're still going to need journalists. We're still going to need um um writers, right? They're going to be demand for human generated content, right? Um supply and demand. So I I I I I would say to young people, you live in an era where you have the ability to build a billion dollar company. Um and not a lot of people know that, right? So some people do, but some people say like, "Yeah, but I could do it later, right?" But now you live in a small time where if you actually try to do it, you're going to be in the in the 1% of uh people who are um actually doing it, right? You've you've seen the first uh billion dollar uh solopreneur with with the medical company. Again, maybe maybe it's luck, right? There's a lot of luck in it, but it was possible to do. So now you can do it. You can you have an idea. You can you can build it. You can ship it. You can It doesn't have to be software. It can be anything you want. You have all the information out there. You know, the stuff that you learned from YouTube. Right. You can you know you can you can compress it now to basically do anything you want and that period is not going to last for a long time because if everyone has myths handy everyone will build their own uh their own business and eventually really the really the bar for SAS went like through the roof you know so you need to do something that is truly remarkable um and now you have the time because not a lot of people realize you think that everyone is vibe coding not everyone is vi coding right people usually use jipt to chat, right? There's the people who actually go and build things are a smaller percentage and those people are in a rare time in history where they're, you know, they know something. It's like, oh, it's so easy. I just had to like build me a calendarly **[00:54:00]** clone, you know, um, and that's it and it's it's working, you know. So, so it's it's just think of something, do it. Don't put so much emphasize on on on on your major. Try to go to the best university you can, right? Never hurts. But think about more of the of the networking and the experience you learn and not about the diploma. That's my Yeah, that two cents. That's that's great advice. Especially like the university bit. Yeah. It's not that you're going to learn everything in university, but but it does give you preparation and who you meet and and how things work. If you get in the right room with the right people, you're going to learn a lot. and and that doesn't mean you have to go to the best university if that's not a possibility for you, but you can still try to get in the right, right? So, so at the end of the day, it's like the networking is really important and the listening to the market like you said because if now everyone can build anything, it's now just really a race to find out what the markets need. So, go out and see what your plumber needs and your electrician needs and your spa services like what what do they need? Build that. Yeah. And and you've got a business, right? Because there's a lot of And they are not bod. I mean, they have to actually go and do their work because they're they're they're working with their Yeah. 100%. But but it is funny what you said about the um uh we think that everyone's vip coding because if you start going through the rabbit hole, then all your feeds are about that. And it's a FOMO that you talked about on LinkedIn. Same as X. It's like, oh, everyone's doing it. I can't like I can't keep up. you have a Mac Mini with an open claw, you're only part of like 1% of 1% of people, you know, and you think like, oh, everyone is doing it. Everyone in your circle and that the LinkedIn algorithm was trained to Yeah. Um but yeah, it's it's not it's it's um yeah, the sample size is is is biased, right? You're you're you think that everyone is doing something, but it's a very small sample size. **[00:56:00]** This has been so insightful, Eran. Ran, sorry, me as well. Yeah, it's uh there there's so much we can talk about and in different different branches. I don't want to take so much of your time, but uh we we could definitely have another conversation and get deeper on some of these other 100% I understand. I would love to come to your show again. Yeah, thank you. Thank you. I that I I would love it as well. So for our audience, what would you recommend if they are vibe coding, right? They're they're just Okay, I'm going to I'm going to start by going what would be the like a quick thing that they could do just to make sure that that they're staying secure in this process even though they're not they're not like software engineers, right? So, first of all, you can, you know, if you're using cloud code or cursor, you know, ask it. It has inside, you know, uh uh slash security review. Let it do it. You know, it will do an okay job, right? There's no reason not to do it. Um obviously, I would recommend, you know, using DevCard. again free no registration mat license so you can you can you can go and sell it if you want you know it's it's completely free just to protect your local uh uh package file so you don't you don't get your computer hacked those are the two things I would do first now the other things you can use um ask it like you know what kind of uh other security things you can do to protect me you can also use tools like like ours we have a we have a free tier um you know arnica.io You can you can use it. It will it will scan your uh source code repo for for common things using AI using determinative reasons. You can use other tools. You don't have to use ours. There's other tools out there. Um and um you know just use something use something that uh will um give you the basic uh security guardrails. Um but talk to claude you know talk to cursor talk to codeex talk to uh base 44 whatever you're using to develop say like listen I want to make my uh things more secure right so if you if you use arnica um you know we have a thing that will **[00:58:00]** generate like you know a best practice uh cloudmd rule file for you right but you you you you can you know if you don't want to use us you can generate your own make sure that you have uh security rule files Um it it it's few steps to just you know you get you don't get 100% security but you can get the easy things um the obvious things uh off you know cloud if you look at the cloud code leaked uh leak it has like one paragraph saying try to write secure code don't do bad things like literally this is like you know like you know try to save tokens like so yeah so obviously you can make it bigger your your limits will be reached sooner um so you know try to think about the language asked like what language did you use sometimes you don't really know what languages did you use um it's probably going to be NexJS with tanstack and and shoten because all the startups now look the same you go everyone looks exactly the same um and um just ask like what go online and search for me what are the best practices for this language what are some of the uh security guardas for this okay write the top 10 go to o top 10 um tell code go to top 10 tell me the latest uh top write a rule file for me that will make sure that the next time you generate code it will you know so yeah you can use our rules or you can do the basic of like ask ask around like make it more secure you'll get something that is better than nothing right if you want to go and raise money and make it in production talk to a professional because you will get hacked there's there is there is uh don't take that risk right so but if you're building something you know for friends and family that's fine if you go and go to production, talk to u uh talk to a professional. Use um you know a production ready security system to protect you. Um because you someone especially you know when when Mythos will hit the you know will be out of uh of the guardrails **[01:00:00]** you know in like you know 6 months whatever or if another company if open a releases something you know thaters will use it you know to to try to uh so so yeah that's great advice and that's very true the because all these companies that raise money there that's all public right and So, not only will you get a lot of soliciting emails when you do that process, but al also you're getting a lot of attention from people who might be thinking that you may have vibe coded your app and look for 100%. So, that's a that's a good point, Eran. Again, amazing advice. I will put the link to articare.io.io in the show notes and also to the uh GitHub repo, the dev guard. Yeah, I'll send you the link. Yeah, send me send the link. I'll put in the show notes. And for those listening, if you haven't liked and subscribed, please do so you learn about every time we're going to have a new episode with awesome people like Eran here on the show. Thanks again, and we'll see you in the next one. Cheers. Thank you.