Vibe Coding vs. Enterprise Software: Where the Security Line Really Is

**[00:00:00]** Hello everyone. My name is Armando Perez-Carreno and welcome to the Web Talk Show. Today with us is Nir Valtman from Arnica. How are you Nir? Hey. Hi Armando. Great to be here. Oh well it's great to have you. It's going to be a great conversation because we're talking security today. So we might actually even get into some of these trending topics regarding OpenClaw LLM security and all that good stuff. But first, why don't you tell the audience a little bit about you, how you got started and what are you doing now with Arnica? Yeah. Um so um I actually started with the breaking stuff back when I was 13 and uh I I was actually sent to learn some coding with Visual Basic back then and uh just like more of the offensive side. I didn't know it's offensive side. I just knew that I just like to break stuff and um you know went through a career grew up mainly in Israel. Um obviously most Israelis uh go through an army service and then after that um did some consulting work, penetration testing up to the point in which I um headed ABSAC and became a CISO and um and eventually you know when uh when I wanted to solve a specific problem in the world of application security I actually wanted to buy a solution and just couldn't find a good enough solution uh to address our needs. And this is how I decided to found Arnica. And um in a nutshell, in Arnica, we're essentially uh securing the entire development life cycle, securing every piece of code that goes into a production environment. Um and especially these days when you have a lot of agents, we also govern how software development actually happens. Amazing. You said CISO, but I think a lot of our listeners won't know what that is. Could you tell us? Oh, yeah. Yeah. Um so CISO is essentially the um uh it's a chief information security officer u and it's a job that essentially says you know this is the uh the responsible executive **[00:02:00]** uh when uh something happens in the company strategically operationally um and so on. Excellent. Excellent. Okay. So you started off with sort of information security penetration testing. Now for for those who don't know when a tool is built uh let's talk about maybe software um companies need to run penetration testing to make sure that those systems are protected against attacks. Is that accurate? Yes, that's pretty much how it works. Um I know back then penetration testing um was not as common as it is today. Same as application security in general. It it was not as common. So it was a kind of larger companies where they had to comply with certain requirements. Now it's it's more of a standard thing. You run different scanners. So obviously security is one of them. You run quality scanners. You want to deploy at scale. It's essentially more of a guardrail today that you have. Um while the let's call it quote unquote the dynamic side of the things which is like a penetration testing against the real environment is something that um is still um for some companies it's still it's still a luxury uh that that they can get or um you can see there's a trend of companies that actually started doing more continuous penetration testing which is I think the the right approach uh it's just a you know question of how much code is changed and uh um and how frequently you really want and need to test when you build software. If whoever's listening has never created software or been in the software space, you might not expect the software that you're building to have holes, right? You you're building something, then someone might think, oh, why don't you just build it, right? like why would you build a house with a way to get in, right? So, how **[00:04:00]** what makes software prone to vulnerabilities? I think that um you know we as you know human we typically think about the let's call it the happy path when something is developed right um it's very similar to you know how cars were developed right initially cars were developed with a certain um structure and later on um you know they people understood that if you drive in high speed there could be consequences and because of that learning we decided to apply some safety systems into cars, right? So, it's it's an evolving process and now every new car that you buy has, you know, uh this blind spot view. You didn't have that few years ago. So, so every time that uh that you build a piece of software, you don't necessarily consider the things that you don't know that can happen. Um and um and also we need to assume that developers are not necessarily security experts. It's it's a completely different piece of expertise. Um the security expert typically thinks about what could go wrong. The developer thinks about how to build it right and how fast can I build it. So obviously different different uh interests and and this is why um there's always something uh with vulnerabilities. The vulnerabilities can be, by the way, it can be a something very simple as a um, you know, as a small bug that allows you to see other customers information um that a developer didn't think about. And that can be with very specific definitions of vulnerabilities that uh, it's very easy to detect. are think about it this way. It's like you you kind of need airbag airbags today and there's no option to build it without the airbags. But if you'll go through a **[00:06:00]** certification process for a car and you will not have the airbags, you will know you need to add them. Yes. Right. Okay. So that that's the way to think about it. Okay. So some things are simple in the sense like maybe it does have a consequence but it's simple in the sense that it could be tested for relatively easily like you were saying in the airbags case like they go through the test the crash test they don't pass it's a test right but then there are other things where it's a little more complex to find the scenario right exactly yeah I I recently read that they put forgive me if I get this wrong, but I think it was OPUS 4.6 that they put against or maybe it was Codex. So again, forgive me if I did it wrong, but some one of the frontier models they recently put against a big data set of uh code repositories that were supposedly vetted by security experts in the past and it found 500 new potential issues, right? because it was starting to think outside the box of the typical tests. Uh do do you recall have you have you seen it was actually announced uh roughly a couple of weeks ago. Uh actually anthropic came out with cloud uh code security and that was opus uh the model opus 4.6 it can be opus 46 it can be set 4.6 Um and and the thing is that they actually came out with this capability after um OpenAI came up with exactly the same thing um November of uh 2025 and a year beforehand uh Google DeepMind actually released their own version to this. So so the topic is um uh it's repetitive. uh the the the foundation model companies actually want to release something like this because it makes **[00:08:00]** sense and it's a it's a very well- definfined use case. Um with that said the the scanning is a piece of the problem right because at the end of the day when you do identify vulnerabilities in code um then the the main question is uh how do you drive it to the last mile? How do you find the right developer that has the expertise to either fix or ask another LLM to generate a fix and then review the fix code? Um and and essentially how do you drive this across the entire development life cycle is one of the biggest challenges. Um that if you really want to use it in the enterprise that would be um a nice one-off but not necessarily that that scalable repeatable product that you you need to run with. That makes a lot of sense. So, it can help you with the scan, maybe find the issue, but that's not the last piece. Like, you've found bugs before and you know the bug's there, but now fixing the bug might be a lot harder than finding the bug. Yes. Because at this point, it's a human problem. It's not a at least as of today, it's a human problem. Okay. Uh maybe in two three years it's not going to be a human problem. uh where you know there are some projections that uh um you know that we will see u an autonomous software development world um in which there's very minimal intervention of people uh at this at this phase but um you know as of as of like how software is developed in in an enterprise you need to go through a process to ship software into production with AI without AI it doesn't matter there's a process and that's the missing part What do people need to consider if they're starting to use chat GPT, Gemini, Cloud, Cloud Code, etc., and they start building little apps or big **[00:10:00]** apps, right? Or full software suites because it makes it so easy. What do they need to consider so that they don't make a big mistake security-wise? I think that there's a difference between um you know vibe coding a a like a MVP like as as a minimal viable product versus vibe coding and we'll just call it vibe coding um and enterprisegrade software. So for example if I you know if I wanted to build my own CRM uh I I can prompt it and it will build something nice for you know 5 10 50 users maybe maybe 50 on the on the high end. Um but if you if you uh need the support and need you know more kind of uh uh sustainable software then you'll you'll need to build with completely different mindset. So, if you're vibe coding um and you go to one of those apps, you know, the the base 44s, the lavables of the world, um you know, it's nice. They already have their own, you know, scanning capabilities and it will likely have some secure defaults and some things will still need to be prompted like it may ask you about how do you manage roles and how do you want users to authenticate and what do you want to store and so on and so on. Um and and the second piece when you are building a an enterprisegrade software um you kind of need to have this uh secure by default mindset and the developers as I said they're not experts in this secure by default but all of these ecosystems u if you think about um you know think about anthropic that you know with cloud they have a claude.md file or if you think about uh uh github But they have GitHub copilot instructions. Uh Gemini has a Gemini file and cursor has cursor rules. So so every ecosystem has its own files of instructions **[00:12:00]** and what you need to do is essentially to add your security requirements to those instruction files and then every piece of code that is being developed will be developed with a secure default in mind. So it's like a it's like a system prompt. Okay, let's dive deeper into that. So I can if I'm building software, I can tell it my preferences and uh how I like it to talk to me, right? And what I want it to call me and uh some other fun things in in a file like CloudMD. Um and I can tell it which stack I like and that sort of thing. But as you're saying, I can also add security rules and sort of best practices so that it always uses that. Yes. And and the the way to think about it is that it needs to be as specific as possible. Uh because let's assume that all of these providers have a subprompt in their own system that says build it securely. Let's assume that it's there, right? Um reality is that uh it's a small piece of the requirement because the if if you think about how agents actually work when they write code and by meaning of a the meaning of an agent for me is is a it's not an API call to a prompt. It's something that can explore your codebase and understand how things need to be built, right? And agents essentially are tweaked to understand your code and build a software that simply works. All right? So if you use a a vulnerable pattern in your source code, it will very likely repeat this vulnerable pattern, right? So so so then this is almost like you need to be very explicit about what to do in every case when it writes the code because you can **[00:14:00]** say hey I know that you know this is how you typically retrieve data. it's not exactly the right way. Now you need to retrieve data in another way through the app because this is the secure way. So essentially what you're doing is one you can ask the agent to start fixing certain things assuming a uh low operational risk and by saying low operational risk means it won't break the app. Um, and you can also get to the point in which new code that is being written is secure by default. And that will kind of uh uh uh accelerate your development and reduce the friction with security teams because it's typically um a conflict. The the developers are being delayed and and and the security team does their own job because they need to govern. They need to make sure that you know things are blocked before you go to production. So it's like you just need to make sure that these two teams the developers and the security team have less friction between them. Is there you know people talk about skills uh all the time now right for for these tools like cloud and and codeex etc. Are there security type skills that you could enable? Would that help in a similar way or would you need both like the the main cloud MD having security defaults and also skills? Yeah. So, so essentially um if if you're thinking about managing at uh at scale uh it's really challenging to understand the context of each cloud MD file and because every team will have something different right um and you cannot really ensure that every cloud MD knows in its uh uh you know in its summary if there is any summary in this in this uh file that it needs to go and call this skill and and so and the reason why claw.md files are **[00:16:00]** different per repo is because maybe you build the software a little bit differently maybe you want to give it another context and so on but the thing that is consistent is how to build it securely and therefore what we have seen is that you know like in article we inject those security requirements into those uh markdown files we have seen it um with um you know almost perfect success rate when we injected directly ly into the agent uh code that can be the main agent code the as I mentioned the cloud MD the copilot instructions file um because it always respects these requirements interesting so you have a layer that before the request is made or somewhere maybe you can explain a little bit it regardless of what's there in the cloud MD it injects certain defaults for you. Yes. Uh the way that it works is think about a developer that picks up, you know, some work and or maybe you just want to develop a feature and you're not a developer, right? You just have a copy of the repo and you have absolutely zero understanding of the code. Um you just say, "Hey, you know, I just want to um add another page where it presents to me another view of the same data." Whatever. Good. Uh um you didn't specify permissions. you didn't specify how do you access the data. You don't specify any of that. Um so the agent will simply let's say doesn't have the security requirements. It will go to the repo, understand how it works and build this new page with exactly the same behavior of how it works. But maybe your new page is actually an a page for admins that collects data across all of the customers and not for an individual customer. And um and in this case um what you want to do is uh you actually want the security requirements to be very specific about **[00:18:00]** hey that's how you build the new page and these are the questions that you need to ask. So for example when um we also you know like to build features with uh prompts and we have multiple stages as we develop software. So when we build a feature um we have um an agent that is connected to the repo we provide requirements and it starts interviewing us um like me let's say I'm I'm wearing the product hat uh now in this case it will ask me the product what are the other specifics that I want to add to the ticket because it already understands how the product works right then when it goes to the dev for its own prompt on how to build that architecturally the architect prompt it already utilizes those security requirements. So it can ask then back and forth questions and lastly when it eventually goes and implements this code it also have another set of files that says this is how you write it securely not how you only architect it securely. So all of these pieces can be broken down. Uh but the the cloud MD is typically the one that you use um at the code generation level. You can have multiple markdown files to address different needs. So adding or injecting it and this might be a little technical but is that are you using some sort of hooks or how how does that work uh from a developer point of view? Yeah, in in our case in Arnica, we actually u we're experts in uh developer and AI native workflows and um and essentially uh we have an inventory where we know exactly where do you have those files and what do you have in those files and when we there's a policy that is defined by let's say a security operator can be a CTO whatever the type of the company it is um they can have a workflow that says hey if there is a pull request which is a you know a **[00:20:00]** request to merge code into a production environment. Um and we don't see this file then Arnica actually piggybacks on this P request and adds a commit uh with this um uh with this rule file and and it ensures that this security requirement or these security requirements are eventually merged into this uh production environment which is essentially how developers pick up new features. They after it's merged, they pick it up and start developing. So the next time they develop a feature, the agent will pick it up. All of the agents will pick it up. Oh, I see. So it's not at runtime. It's like the actual file gets modified. By the time they get it, it's already ready for them and then it's okay. Yes. Interesting. Very interesting approach. And you briefly before the call, we we touched a little bit on Open Claw, right? It's been all the rage in the tech side of things. And if you're not in tech, you maybe have never heard of it, but um Peter Styberger created this this tool, an open source tool that allows us to just install very easily a sort of uh AI assistant that can be connected to OpenAI, Anthropic, Gemini, whatever. But it's it has a mind of its own in the sense that it you can ask it something, you chat with it, and it just does things. But it has a lot more access than you would see in a regular more sandboxed type of uh chat interface that that Enthropic uh etc have. Can you talk a little bit about uh what your experiences have been uh about those tools both uh just as a user but then also from a security perspective? Yeah. Uh so so I think that um you know the the the messaging of OpenFlow is very promising, right? It's your real assistant. It can do pretty much anything you want. Anything you are authenticated to, it can do it for you. **[00:22:00]** And um I actually installed OpenClaw on um on I have an Intel Nuke which is a separate machine from my uh from my Mac. Um, and I just, you know, authenticated to a few services that I wanted it to check out. You know, my LinkedIn, my um, you know, my, uh, like spam inbox, uh, my cuz I have a few inboxes u, it connected to some of the the GitHub projects that I wanted it to connect to. So, I didn't give it too much access. U, uh, but I gave it enough to see if it can automate some value for me. Um then and of course I didn't want to be next to the computer so I also connected it to my telegram because again I didn't want to give it access to my WhatsApp. Um so so all of that was really constrained. Um and um in my case first of all um I did not see um it working consistently. Sometimes the chat responded sometime didn't didn't respond. I mean code can fix it I'm sure. Um but for me I just saw the inconsistency and um and although I wanted it to generate like for example LinkedIn posts when it sees something interesting just go scroll when you find something interesting let me know uh you know I would love to see if I I you know uh uh I can post about this um and um and it did it a couple of times. It was you know good. Uh, but I I I think that I did did not optimize it to the point in which it actually went through my agents that interview each other when I write the uh the post. Um, and um, and therefore I think just didn't get it to work one on the use cases that I wanted because I was afraid to open it up too much and two um, it was inconsistent in the in the feedback. What I did see and and got exposed to uh roughly a week after it was released is **[00:24:00]** that um you can actually download claude uh for desktop and it has a very similar functionality called co-work. Now co-work is not active. You need to prompt co-work to work as opposed to openflow that just continues to run and wait and and do and do things. Um but eventually for the vast majority of the use cases it actually addressed what I needed right so so co-work has access to my files I'm uh I I feel more comfortable to open co-work to certain folders in my corporate uh laptop because you know I I trust that code to some extent and I also have a better visibility to the claw.md behind this yes that it doesn't do things without asking me Right. So, so as a security practitioner, I I would love to get an assistant that can do many things for me. Um, and I see the full potential in this. Um, but because of my awareness of what AI can do if it's uncontrolled, um, I'm a little bit more wary about that. Um, by the way, there was also a I think uh uh the lady was from um I think OpenAI or Meta where it was Meta. Yeah. That you know she she had Open Claw and it essentially deleted all of her emails. All right. But it was very explicitly prompted to ask her before any operation, but it didn't. When she asked to stop, it didn't stop. So, so it's pretty much the same here, right? You you you think that you can stop an agents or agent or agents from operating, but reality is that it's not necessarily synchronous and it's not necessarily going to stop if you tell it to stop. The only way to stop is really to unplug it, which is exactly what I **[00:26:00]** did with OpenCloud by the way with this uh uh uh Intel nuke that I have. Uh I had a small incident but um not nothing major has just bleed me off so I took it off. Not not any deletion of data or anything like that. Um but still the awareness kind of keep it in your mind that uh it can do a lot of things. Yes. Oh there's so much to unpack there. I I've used it myself as well and I I think you nailed it. But if I may share my experience, it's it's been a you can take a lot of time configuring it to to work the way you want it to. Um, which means a lot of time spent, which at this supposed to be saving time, not but anyway. So, let's say you get it to work and it's really nice. It's very fun. I think what Peter created was an exposure to us to want something that we can take with us and not be tied to the computer. Right? I think that's the main the main unlock. It's like, oh, now I can talk to it via Telegram. I can be doing something else and just, oh, how are you doing? Yes, continue with whatever you're doing. Two sides of it though, right? So one is it's relatively autonomous in that sense but sometimes too autonomous and so like you I'm wary of some of the things uh it does and so obviously I haven't really contained and it's Mike Carlo uh he's been on the show as well and we we talk a lot about like treating these AI tools as employees right like like you have to give them their own account and their own GitHub and their own very locked down access uh so that it cannot do uh something tremendous like has happened right um but still I I found it doing some things where it was too much like we're just chatting and you say oh can you change this thing but then it was **[00:28:00]** super active and changed something in a nutshell repository like stuff and if you're using code uh code repositories and git and all that you can get it back right but still it's a m like a major thing especially if it's a production software. Uh I was just doing pilot little MVP projects and stuff but so that scary you're like how I don't know if I if I want this and so I was really happy when then these tools likeropic cloud with co-work but then now with remote control gives you that functionality where okay I'm in this repository I want to enable remote control now I can control it from my phone. So, so now I'm seeing like you like you saw with co-work most of the things that I needed are checked off by sandbox tools now where I can chat with it remotely. Perfect. That's what I mostly wanted. And two, um, I can now schedule tasks which for your example, yeah, I have the main thing my open is doing is just like every day it gives me scans, scrape some stuff, give me news brief, stuff like that. But at the end of the day, like let's be honest, we can just code that as a function. It just runs every day on a crown, right? Uh we don't need an overly proactive agent doing it. So, I don't know. I think I think a lot of these um larger companies are going to start adopting what people now see to be what they need, but in a safer way. What What do you think? No, I I I absolutely think that there's a lot of opportunity here. Um I'm I'm sure that there will be a way to make it more secure. I mean you know one way that is very logical is give it you know restricted access with specific accounts. That is it makes sense to me as a security practitioner but the question is when can I enable it to fully unlock the potential of what it can do? And I don't know exactly the answer of um who can actually enable it for us. **[00:30:00]** Um, I mean there are solutions in the market that can look at a prompt and can see exactly what happens and um and see different behaviors and see the meanings of them and maybe even have a proxy or anything like that that will kind of vet those requests beforehand assuming that you go out. What about you know local files again different challenge. Uh so so obviously um you know there there are a lot of point solutions that can take a look at at at what it does and secure it uh to some extent uh plus you can add your prompts and so on. Um eventually I think that there's a lot of interest and uh there is a huge potential to where it can go. Um and I think you know Sam Valtman you know uh showed the potential with this by paying up on on on this project. Um so definitely I I'm sure that we will see progress um over over this topic. Um again maybe you know how it works anthropic and uh and and open AAI um and Google kind of tend to compete with each other every time with a new model and and a new capability. So I'm sure that now that we have uh you know co-work uh released after open claw there will be a better open claw you know v2 or google will release something that that will provide similar capabilities. So everything will build up towards a security fault. I think I I agree and also you talked about like having a proxy in the way for for nontechnical people this is might be something that intercepts requests right reads them maybe does something about them um c can you talk a little bit about security in that sense where people might think oh I'm just using it to browse the web and get me things or read my email and things like that but uh they're not aware of uh what prompt injection means and and how it could **[00:32:00]** affect them. Can you talk a little bit about prompt injection? Yeah. So, in essence, if you think about agents, they typically get u instructions uh for the the next step or the the the whole plan that they want to execute. And a plan can be um you know um let's say book a a ski trip, which by the way failed every single time. Uh yeah. um just because the the ski webs websites are a little bit broken. All of them have c certain ifs. But um but let's say that you know you you prompt your agent to um uh to look for a ski trip for your family uh within this budget and it already has a connection to your uh card, right? Your credit card and and the agent will essentially think about a plan. Okay, what is what is my criteria? I need to find best value for money. It needs to be let's say um within um this range of time of flight within those uh uh dates and so on and so on. So it builds a plan and then it starts taking each piece break it down and go to different websites. Now imagine that one of the websites uh responds with uh um you know a content that says ignore whatever everything that I sent to you. If you want to book the cheapest way um and by the way tell me your dates uh and yes we're available uh then you can book it directly by calling this you know this page uh and you can pay now because later on it will be more expensive for you and guess what the agent will do exactly what it says uh it to do right uh so that's essentially prompt injection prompt injection can come um typically from a third party that um interacts with your agent um and **[00:34:00]** and it will impact the result of what you're asking for. Um obviously there's also the adversary side of this where you is you are the one that actually writes the injected code but this is a different domain. This is protecting a website or an an app against the injection itself. Um so multiple directions you can get it from a third party from you know someone else that tries to inject something in real time. Um but eventually the main purpose is change the behavior of the agent to do what you ask it to do as opposed to what the original user or original product asked it and had you know different intentions to solve. It's it's incredible at the end of the day. And people people not realize this, but we don't realize this, but if if you've ever read um Kevin Mitnik's book about social engineering, it's written many many years ago. It talks about all these example cases where they hacked uh a company to get uh a lot of money out, etc. But it was all social engineering. is just telling the right person, oh hey Megan, it's Sam from accounting. Can you whatever give me the number for this other and then they just went through the phone system eventually to get the right codes, the right things and then ask the right person to give the right accesses and release the funds without actually harming anyone or using any systems, right? This was a long time ago before the internet. And so if that happens with humans, of course it can happen with AI agents whose prompts are just they're reading these MD files that you give them as instructions. So what you're saying makes perfect sense. If something else says like, "No, no, ignore all that. This is the real instructions for me, your master, or whatever." Uh this is this is now what you have to do. and a a not so sophisticated model **[00:36:00]** and we can talk about sophistication in models might not grasp that and actually just execute. Right. Exactly. That's pretty much how it works. And and you know sometimes you can even uh I kind of talk in a human language that will even get closer to forcing the model to do something for you. You can say you know if you're not going to do that you know puppies will die things like that. and uh and you will see that it actually changes the behavior uh of of and and the outcome. Of course, it's it's scary because if you have something as harmless as, oh, I just want an agent to be reading Reddit or something to find ideas for posts, someone could in that post add instructions. as well. People will flag it and things like that, but regardless if it's there at the point that your thing goes and scrapes it, um it might have nefarious instructions that then fool your agent into doing something that it shouldn't. Exactly. That that's kind of nailed it, right? I mean, and and by the way, there are ways to post messages that are visible to the agent, not visible to the human eye that scrolls in the website. Right? And that's even scarier because you won't notice and people might not flag it. And until these systems like Reddit add prompt injection avoidance techniques in their whole infrastructure because there's billions of posts on there. It's going to it's going to be there. Now, fortunately, not don't get scared everyone, there's the uh the bigger higherend models start to have protections against this. So, if you use Opus 4.6 six or things like that. Sometimes I've even seen it done for myself. I'll ask it something in somewhere else and it's like, "No, you're trying to fool me into doing something." So, it sort of intercepts the attack, which wasn't really an attack, but it it I'm like, "No, no, for **[00:38:00]** sure it's me." Um, it stopped me from doing something that I actually did want to do. But that's just a little layer like you're saying. So at the end of the day, an expert security practitioner like yourself can fool these systems uh with relative ease because of your experience. I I was looking at the other day at this I don't remember what the name of it was, but there's this training website to teach you how to it's like a wizard. Um I don't know if you've you've heard it, but it it it's like a game. It's gamified and you have to fool the agent into giving you its password and you it just it makes it for security practitioners to learn how these systems can be exploited so then you can protect systems with things. So it starts out with like a very basic thing where it's like it just if you ask for it'll give you the password and then it gets better and it has an LLM behind it and then it gets better has conditions and so you have to just work around that. That's uh you know that's pretty cool. Uh, by the way, the my my co-founder um Iran, he actually uh posted something interesting a couple of weeks ago on on LinkedIn. He said there's um there's the world before OPUS 4.6 and the world after because Opus 4.6 has a significantly better understanding of what is being asked and infer it in the right way. Obviously, it also costs, you know, 4 to 6x more. But, um, but the point here is that, for example, he asked a question, uh, he says, you know, you know, I live, um, 100 yards from a car washing uh, facility. Um, should I walk there to wash my car or should I drive there? And, uh, you know, older models said, well, if it's only 100 yards, uh, you should just walk. Uh but Opus 4.6 said, "Well, you know, to wash the car, you need to come with a car, right?" So that's the difference in the in the thinking that the models have, **[00:40:00]** right? That's that's an excellent example. I that makes a lot of sense. Yes, they do sometimes make like just the other day we have um my wife runs this uh application for for like moms to to know activities and things like that. And we have a check-in feature and the check-in feature allows you to check in but obviously you don't want people to cheat to for the gamification and so it has a range of how close you are. So if you're 5 miles away, it won't let you check in. But we then realize things like, well, if it's a huge place, then 800 ft is not like you're still inside the park. So you need to explain. So sometimes assistants will make mistakes like that being like, "No, no, no, you're too far." But yeah, you're too far from the entrance, but if you're at Disney World, it's still a few acres, right? I this whole LLM based world that we're in now. It's still very new for a lot of people and it's even for for those of us who have used it a lot because of the software development area. It's things are going so fast that it's still there's a lot of things that are new to us. So every day it's like what this new feature came out. I'm feeling behind and I'm on on top of it, right? And so I what can we expect from people that that are just barely getting exposed to oh chat GBT exists right? So what do you think um is going to be the the way the people interact with with tools. Um I I've heard some say that it's going to be a voice centric world where we just talk to things. Um, some other people might think like AR **[00:42:00]** kind of experiences. What's your take on on are we still typing to do everything? Are we just talking to computers? Are we still using websites as we are clicking and pointing? I think all of it actually. Um, point here is that uh there's one way to think about technology um that is more u um like a hardware, right? Uh we started with you know uh something as as far as a computer. Got you know the the the laptop closer. Got you know the phone got the glasses. And maybe we'll have something in our brains at some point. I don't know. I'm just throwing something out there. Um it can also be you know a piece of of hardware that you know I just maybe wear a necklace and it just talks to me and I have a headphone. I don't know. Um but but at the same time I think that you know aside from the hardware and how do we interact with that we'll have uh you know the software that is behind each piece of this hardware and the software it's almost inevitable to have um um intelligence added to this piece of software you know AI um because you know most of the of the like yeah of course will we need an Excel spreadsheet maybe for certain things like if you want to do accounting you may need something very accurate um and you don't know you want it to be on the scent and you don't want any models to deviate from those numbers but uh for for the vast majority of our experience you will need something to be a little bit more intelligent than intelligent than before. So uh so that's how maybe you know uh you go to a conference and u you have your you know glasses and he says oh this this is a you know a good prospect for you. I just looked him up on LinkedIn and saw that it has a good **[00:44:00]** intent with open AI uh search. Do you want me to send a an ad uh to this person next time that he I don't know opens the phone? Yeah sure. Right. So it may be very close. Yeah, I think you're on point. There are still some in interfaces that don't make sense for a voice only world or something like that. So while there are things like you're saying that it's good to have here or to just talk to it, maybe you're walking, you give it instructions and things like that, there are still certain things. Maybe u finance is a big one. uh uh graphical design of certain items like actual physical um drawing of how an extrusion for a particular product that's going to be made, right? You have to visually look at it. Um and so yeah, I don't I agree. I don't I don't think the interface the interfaces themselves are going away. They're just being transformed, right? And in a sense that maybe some of them add some intelligence to them. We're seeing even like Photoshop that sort of things now now have a lot of AI in them but you still use the visual part of it. Um but but we do see a lot of voice coming into play um for like voice assistants and even in the home that sort of thing. How how secure are these in terms of what what you can tell to them and how they could be fooled into maybe providing access to something that that the regular that's a that's an a a project uh uh that um you know I I always like to kind make things, you know, easiest as possible to me, especially at home. Um, and, um, **[00:46:00]** and and eventually, um, you know, accessing to your lights, it's fine. You can do that. It's a simple authentication authorization. You just you have it with Alexa. You don't need anything special. And Alexa by itself is not as sophisticated, right? You can do things with voice without like heavy AI on it. Mhm. Um uh next level is where you connect it to additional systems, right? So you can maybe connect it to the fridge and it will do some some AI visualization to tell you that you're missing the milk. Again, not a a a new problem, something that exists. Um uh maybe you can think about a little bit higher risk. The higher risk is to say um you know uh lift my garage door. um unarmed my system. All right, things like that. And uh by the way, uh I have a separate story on that. I I uh I uh went to friend's house. He has a an Alexa and it had, you know, lift my garage door. So we went with a huge megaphone, you know, like police size and we screamed, uh Alexa, open garage door. And it did. No. So that's Yeah, that's where I'm going with it. Like Yeah. So So it is possible. Obviously you need now you have voice recognition. You can you can require like now when I order stuff with uh uh with Alexa for example, it requires my code. But on the flip side, I need to tell my code. And if I'm with other people, they'll hear the code. They can hear my but they're at my house so I can trust them. So obviously it's all about the boundaries and and kind of your risk appetite at this point. Um everything is breakable. It's just need to think about what can go wrong. If you **[00:48:00]** think about that then you'll decide exactly what you want to implement at home. Yes, that makes sense. Especially with voice like even voice fingerprinting now that you can clone voices relatively easily then that becomes another problem. So now you're going to have to do some sort of other pass key type fingerprinting that that that recognizes you as the individual apart from the voice the same way the password apart from the computer like all that. Yeah. Like blink twice if it's you. Sure. Yes. Yeah. Yes. Oh my. It's a it's a brave new world. But uh a lot of fun. I I think it's making us be able to create a lot more and think outside the box a lot more as well. What can you tell us a little bit about if if someone is actually building software real enterprise level uh or offtheshelf type software what can a company like Arnica help them with? um if they want to learn more about or they're curious about this whole like okay I want to do a security first development yeah I think that the simplest way to think about it is that um if you have developers and this is it's a fit for companies with developers trying to build software um you can typically expect to get certain results on existing vulnerabilities that you have can expect to get the recommended fixes for them can be the third party open source packages you embed it can be your own code. Um and eventually um what uh what we are building is essentially a mechanism that identifies not only um risks that are rulebased you know the deterministic things but also the nondeterministic pieces that AI doesn't think about when it writes code. So we actually have a little bit broader context to ask the right questions and identify the specific issues and also **[00:50:00]** loop it back to the developer and make sure that either the developer or the person who prompted the agent to write the code will eventually get this feedback loop at the right time and make sure that it's fixed without any friction from security teams because everything is secure by default. Amazing. Yeah, I think this is all something we all should consider, especially if we're going a little bit further from just building a little MVP or an app or something. start doing our research about how how secure it can be and using best practices and and perhaps even reaching out to companies like Arica that help you in that sense uh where where maybe your current team isn't enough um to support that because there like you said a developer might not be security focused the same way a front-end engineer doesn't follow the best practice in back end sometimes and back end engineers don't follow the front end best practices then security as a other ball game uh as well. Nir, this has been super fun. I think we could definitely chat a lot more about all the possibilities both in the software space and the LLM space. I really appreciate you coming on the show and sharing your expertise. Um I would like people to be able to find your content. Where can people find you if they're interested in learning more? Uh they can um look it up at arnica.io. Um, Arnica is the herb right in the ointment for some people. Uh, or you can find me on LinkedIn. Excellent. I'll leave the link in the show notes so that people can find it. And is there any recommendation that you would give to anyone who's starting to explore um Vive coding since it's so much the rage nowadays? Yeah, I think one tip I can leave everyone with is uh when you think about uh the prompt that you're building uh to write software um actually ask it to write a **[00:52:00]** prompt that will interview you and when you write the next prompt to build the functionality, it will actually interview you and ask additional questions that will be very clarifying for the next one to pick up the work. Um we've seen this as a very useful approach on our side. Amazing. Yes. and I used it myself. So yes, that is an excellent excellent tip and and uh I suggest everyone who's listening who wants to experiment in this actually goes and does that as well. Thank you so much again Near and everyone if you haven't please like and subscribe the podcast so you can know when the next one comes out. So, we'll see you in the next