Latest / Prompt to Prod / He Hasn't Hand-Written Code in 18 Months | Prompt to Prod - with Wyatt Roersma

Transcript

He Hasn't Hand-Written Code in 18 Months | Prompt to Prod - with Wyatt Roersma · Prompt to Prod

• speaker-0: The more I prompt with AI, the more I interact with the models, the more I understand like what they're good at, what they're bad at, how to phrase things to where I can let them do stuff and I know exactly what's gonna happen because I've done it enough times. And that's like some cope, but that's like humans' cope at its best, right? Like you don't drive through ice because you're good at it. It's because humans have this ability to make risk decisions on some blind data. Experience speaker-1: When people really realize how much these tokens cost, the shit's gonna speaker-0: Right. speaker-1: Hey everybody. ⁓ welcome to the Promptoprod ⁓ podcast. I'm so excited to have you here and listening with us. you know, this is a new podcast. We're really excited to ⁓ bring together ⁓ a bunch of folks who are building with AI, thinking about AI, ⁓ in the context of cybersecurity. And ⁓ my mind immediately went to Wyatt ⁓ Rosma, who is ⁓ somebody who I think of as as cutting edge of AI. ⁓ so first of all, Wyatt, thanks for being here, man. I appreciate it. speaker-0: Yeah, happy to be here. Thanks for inviting me on. speaker-1: So you're a a ⁓ man of of many talents. ⁓ you're a security guy, you're a developer, you're a detection engineer. ⁓ I think of you as all those things. ⁓ but maybe you can give everybody just a quick minute on your background, ⁓ so they know who you are. speaker-0: Yeah, so a long time ago, ⁓ I first got started with volatility back in like twenty twelve and started contributing to that project for memory forensics. was doing some of the largest memory dumps at the time with Hyper V, so it contributed a little bit to that early on. ⁓ helped a little bit with the project. ⁓ ended up through career the most notable like Fanny made a Target. I was at Target Post Breach. Then I worked on their threat detection operations team in which I developed a custom like cuckoo, sandbox, dockerized, all that ⁓ from like 2015 on 2019. ⁓ and then I went to a range developer where I worked on doing like automated range analysis, deployment, user emulation, attack automation, stuff like that. I've also worked on a part-time cybersecurity ⁓ testing and training platform on the side while I started back in like twenty eighteen. It was the analyson known cyber range. ⁓ and I ran like GERCON challenges on that. So I've been running GERCOM challenges for digital forensics and response since about twenty thirteen. So I've done every single one since then. I've done a lot of challenges. And I really enjoy the perspective of everything from development attack design of a network to exploiting it and then trying to tell a story through it. speaker-1: Awesome. So you're also somebody that I think of as a builder, right? You're part of these projects. ⁓ you're not a person who sits and analyzes data from tools that you build tools. And that's one of things that I'm most intrigued with, just sort of where we're headed, right? Because I think a lot of us, you know, we scripted through the years or built little side projects, but like the what you're able to do now is incredible, right? Like the way you can scale up capability and obviously costs, but also capability. ⁓ so I know you have a couple of projects that you are open sourced on. ⁓ like what are you mostly building these days? speaker-0: ⁓ so I have a couple open source things you can find for Threatflux. I moved everything to that organization. in terms of my open source development, which is pretty light at the moment. I mean, granted, there's like probably thirty some projects on there that are open source, a lot of Rust S D Rust SDKs, ⁓ things like that. So but most of my time is actually focused on ⁓ what I'm working on at countermeasure security, and that's where I'm at right now. So ⁓ And ⁓ really built a pretty sophisticated Rust backend to kind of build and deploy and manage our like tenant infrastructure environment, ⁓ the all the all the Rust backend being that where it can handle pretty much a whole isolated ⁓ environment for like messaging, tasks, ⁓ deployment of infrastructure. ⁓ like a dev secrets and like a vault type of thing. the the main product that's coming out of that is our coverage platform that's that's focused on helping people aggregate like their visibility, intel and detection and say, hey, like like 10% of your devices don't even have like your defender or CrowdStrike on it. And you could just like easy win right here. ⁓ because like you know, like I was talking about kind of before we started Where everybody's talking about vulnerabilities and when we look at the defender posture in the real world, a lot of things don't get patched that fast. ⁓ blah blah blah. So like what what's really gonna help infrastructure and blue teams out there is like what they can do with the least amount of effort to get the biggest wins. ⁓ and if that is like identify like you just absorb this company or whatever, and you need to like deploy, you know, some security telemetry so you can actually just know what the hell what is going on. Like that's gonna be huge. because you're not like it's always kind of been like an attacker like can always infiltrate if they wanna pay enough money to have enough motivation. It's just more true now. ⁓ so the defense needs to scale up and go, how can we like protect our assets better, get better visibility and coverage? So when something does get hacked or O day is used, or a vulnerability we've known about for two years, so we still haven't patched the server. has now been used against us, which is sadly very common. ⁓ you know, common, you know what mean. ⁓ but yeah, so that's kind of like what we're helping you build out and track it all by like miter mapping. So it's kind of easy for people to visualize. Like, hey, you you don't have anything that detects data exfiltration like at all. So like and then on top of that, you only have 50% of devices that even if you put this rule in You still wouldn't detect half the data like exploitation on your environment because you just like can't. Cause that's like I think what most people kind of need in terms of like ⁓ to unify a lot of the security stuff. It's kind of like what a sock was or not a sock, but ⁓ a SOAR was supposed to be and all these other things where it was like, Well, how do I do the easiest things? And they just gave you very complex tools that are not intuitive to the people on the other end, ⁓ that end up just generally being middle people between like the infrastructure, the programmers and all the parts of their company, which becomes very painful because I mean I can recall a time at major company where I requested like the firewall rules as like an export and the dude told me no. Like on the detection team. And I'm like, what? speaker-1: What do you mean, though? Yeah, we work for the same company. speaker-0: I just need to check to see if this this this thing is possible. And they're like, nah. So I had to get my boss, ⁓ at the p time, Paul Nelson, to like facilitate that, which was just like a kind of a stamp of how the industry kind of like they don't like all the parts don't talk to each other, be it like red team and blue team, or Intel and the blue team or detection, like. So on and so forth. Infrastructure programming. Infrastructure just cares about their thing. Security cares about their thing and everybody's just working towards their objectives. And it's not really like unified in a way that helps everybody. So speaker-1: So this is probably the longest conversation in cyber this year that didn't have AI in it yet. So you know, I I would love to ask you, you know, in this world that you live in, and obviously you're building solutions for customers. That's awesome. ⁓ you're building these side projects. Like, how does AI fit in your world? Like how do you think about it? ⁓ you know, I I know in some of your stuff you and I I I never know with you if you're being cheeky or being real when you like co-credit AI with some of your code. Like it'll say in there like co-written by Claude or something like that, which is I think a cool shout out, but also kind of funny. Where are you at? Like is it a huge part of your day? Is it part of your day? Like what what how much you working with AI right now? speaker-0: Yeah, I haven't written code in like like hardwritten code like 18 months. Plus. ⁓ I've been automating everything ⁓ ever since like the origin of like Cline Roo code and then Claude code and then you know like Codex and all that. And you know, figuring out that well for me I just need an orchestra of systems to like give it a dedicated sandbox environment, let it build and test a thing. And the more you can build and test locally without anything external, ⁓ which is kind of the reason behind like the unified system. It's like it can kind of test all the points that matter, and not have to like have secrets or anything. It can just stand up a thing, deploy the thing, and test it without, you know, shall just affirm all. So things like that. ⁓ where you can prove things through like screenshots of like, hey, it's changing this UI. Have it crap capture with playwright screenshots. that record the the interactions and the changes and the visuals that change on your UI UX and then have it compress it all into a GIF so you can watch it in the in the PR. ⁓ right. So speaker-1: Wait, so you're having AI build you demos, basically? speaker-0: ⁓ just like what would be the visual changes. So it's like, hey, I changed this to do this and now I've done a playwright test that confirms it works. Here's the visual play of it working. So you can watch it, click through like the interface and see what happens and confirm that it's working. All on side your tests, because I mean I don't want to waste my time by standing that up and and testing it, right? Like that's a lot of work. ⁓ so I just have it do it. ⁓ I actually picked that up from OpenAI. They They at one point posted a video about how their PRs come with videos. ⁓ so I was like, wow, I don't know why I'm not doing that. Cause I was doing like a picture thing and I'm like, I've run into the problem where like the pictures get too big. So like a compressed GIF is amazing. ⁓ cause it just like stitches it all together and you can kind of get the idea of what's going on without like taking up a ton of space. Yeah. ⁓ so yeah, and like because I use like I I pivoted to like a very tight Safe language. Obviously, Rust is very strict. I moved away from back when I was like writing the arranged stuff with like Python and Go, ⁓ that I did all myself to ⁓ Rust with AI because I found that it was very good at producing what I wanted ⁓ when I put the things together and the plan and the context and the skills or however you're gonna do it. ⁓ very specifically so it It would build what I wanted and needed, and then I could verify it. ⁓ so really I just spent all my day in code review looking at things that are like obviously wrong, and then you know, visually checking things out, which I mean does still happen where it's like, well, I see you wanna do this, so I'm gonna try to open up this port. I'm like, no. Stop. Stop. We don't we don't do egress, stop it. If you need a new service, I know it's harder here, but you like So those are the things that like the more you understand about the bigger picture, the better you can be at it. Cause you don't really need to understand coding to see a rule that's going allow any, any or allow this port, right? Like you need to be able to put those together and still be able to conceptually understand where your risk is at. ⁓ and what you expose to the internet, right? Like ⁓ you just put everything behind Cloudflare, zero trust. you you can improve a lot and then put another thing on top of that. just onion it up, right? Like that's kinda what I did. and and inside the platform AI will ⁓ runs in a whole pipeline thing, but that's ⁓ that's pretty depth. So if you wanted to get into that would be pretty interesting. But speaker-1: Well, part of it for me is like, you know, obviously there's specific things that you do in your day to day that are relevant only to you. But I think there's a lot of people who are in cyber who we know who, you know, are thinking about these things. Like, do I need to buy that solution for doing this function or can I just build it in AI? Or, you know, how much relevance, you know, if you think about like detection engineering, like Do should I build my own model that builds detection rules and validates rules? Should I download one and run it? Like, you know, there's always like I feel like it's interesting because I was talking to somebody about this. I just went to the CrowdStrike comp conference. I was talking to somebody about this. And, you know, I've been in cyber for 20 years. And I'm just gonna build that was the rarest answer in cyber. Like you heard it occasionally, but like basically never, right? Like it was like, I'm gonna buy it or not buy it. Like that was kind of it. And now it's like, well, there's this whole third option of like, I'm not gonna buy ASO ASOC, I'm gonna build it. I'm not gonna buy this, I'm gonna build it. And that's weird from a group of people who aren't developers, right? For the most part, cyber teams are not developers. So how do you think about that distinction? Or like the boundaries, are there any boundaries? Or do you think like just YOLO, like build it, see what happens? You know, if somebody comes to you and they're like, I need X, you know, should I build it? Is your answer yes? Like try it, even if they're not coders or You think that's maybe a little overhyped? Like how do you feel about that? speaker-0: Say I don't know. ⁓ and it's it's kind of a cop out, but I will say what I do know. ⁓ and that's it, it's gonna be very dependent on the people you have implement it. Do they not as coders, but as risk analysis, right? You can let stuff happen in a sandbox environment if you know what a sandbox environment is, but if you don't, and you just like things can get out. then you know, it really just depends on how good that that is and how much you invest in it. ⁓ and how much risk people are willing to take on isn't something I can just like blanket say for everybody. What I can say is I absolutely think everybody should be building as a practice with AI. ⁓ because I see I see people like, well, I go back to coding sometimes, so I can I can practice it to feel it. And I'm like, right for me that it To me, makes no sense at all. I mean, I understand where people are coming from because that's like a a thing that's as true as time. But like, unless you're telling me you also go and like do like punch cards and like you wanna write the bits yourself, like what what are you practicing? You know, like we've just been attracting things for years. Like, why is this the the fork you're gonna put in say in to say, hey, I'm done here. If I'm just gonna stand on this hill. I'm like, I don't I don't get it because for me, then they can drive through ice, right? Like it's one of the hardest things for AI and ice and rain and stuff. Anyways, ⁓ you know what I mean. So it's like imperfect information is a real problem ⁓ for AI. So you you need to be able to like catch those gaps of like what applies to your company, how an API works functionally, not necessarily what code does. I mean Do you not understand some logic? Yeah, yeah. But like that, that really comes down to like input and output. ⁓ if you you know, like, and why I'm encouraging everybody to go build is not only can you like solve something for yourself, but you can learn how that process works more. You can work with, like, say you want to publish it open source, you can go learn some CI C D stuff, you can go learn all those things. And that's where AI is a superpower. And that's why I say to go build is because you can learn all these technologies now. Because AI is really good at explaining stuff to you. Is it a genius? No. Is it perfect? Absolutely not. But for somebody that's never done something before, it can absolutely give you amazing step-by-step instructions. It's actually crazy, right? Yeah. Like, and if it's wrong, sometimes you just gotta be like, you gotta know, like, ⁓ this this company updates their docs like so fast that it's not gonna be in training data. So you'd be like, hey, double check the docs, and if it's not publicly available, you're You gotta like sign in and be like use the browser to go through this right quick. because I ain't gonna read it now. I'm gonna like have AI summarize it for me and be like, okay, this is I see what you're wanting to be. I get it. I understand what they're going after. Be like set some placeholders here, run the script. Does it work as far as you need? Okay, then I'll take over and do what I need to do with like anything secure wise or whatever. ⁓ but as far as everything else, you just Yeah, use it to explain things to you. Have it like take a paper and explain it to you. like, yeah, it can hallucinate, but like, okay, so do humans and like humans are wrong all the time. I mean, just look at the art of memory forensics. There's an errata and like that's because I would have never read it and caught half those errors. But Joaquin Metz, man, he like he caught a ton. That's because there's people that are like so smart technically that like the rest of us are gonna make a ton of mistakes talking about like. protocol and file formats and stuff like that because we just don't get it at that level, which is fine because we all need experts, but not we don't all need to be expert developers. And something that got lost there is the people that became developers weren't the people trying to solve the problem. And if you know what problem you want to solve, you're the most useful person to solve it because you know exactly what you want. And so AI can facilitate that. speaker-1: Yeah, I mean that's somewhat been my experience with it because you know, I've been in cyber for twenty seven years. I've built almost nothing in the sense of like coding, but I have told developers to build hundreds of things right over my career and stood in front of rooms full of people and been like, do this, do this, do this, and then you know, some of it happens, some of it doesn't, you know, some of it gets built bad, some of it gets built good. But like I've I've overseen a lot of those efforts. And there's always been to your point, like a gap of like, I see it and then it doesn't happen. Right. And then, you know, with this new generation of AI where I can go, yeah, I really want ⁓ you know, the A either AI to build me this or AI to do this. And it and it does it at this like, let's call it 80 20 level. Like it's good enough. It's not probably spectacular to your point. I'm sure there's some genius who could code it better, but it's like fast as hell. Like and I can iterate really fast. And I have ⁓ I I've probably like you on like jump between models a lot, right? Where I have started with obviously we all started with GPT and then I spent a lot of time with Gemini. Now I'm spending a lot more time with Claude now and sort of different things and and you know, models have personalities and so you kind of fall in love with one and and hate another one. What I have found though is is as a non developer, the technical infrastructure part is the hardest part. Hosting and ports and and code to CSCD to pipeline, like It's actually the part that I find AI sucks at. Like, I'll give you an example. Gemini doesn't know how to run GCP. No, it doesn't. What what the f how does it like, how does you not know? I just faced this the other day where I was like, I was trying to stand up as some standalone Python workloads to run this NVIDIA spill spec skill specter thing, which is like the NVIDIA open source project to run skills and and see if they have threads. You wouldn't believe how dumb Gemini is. It's like I speaker-0: Mm. speaker-1: It it couldn't do anything. So I had to like cut and paste a thousand commands in a cloud shell and try to get it to build out. And like I'm not a coder, but I'm like, I know how to run cloud shit. But I'm like, this is kind of dumb. Like it's it's a little stupid. And, you know, I know the hype people are sort of selling us that it's all this magical thing, but if you go to Claude or you go to Gemini and say, build me a secure infrastructure, it cannot do it. It has no ability to do that. And so our bosses think it does because they don't know anything. But like magic wand, make it happen. So in your world, if you're starting out and you're like, I wanna I wanna build a project or I wanna I have this great idea and I'm in cyber, ⁓ whether it's using AI to code the software or not, like how are you thinking about your stack? Like what does your stack look like when you start a project? Are you like cloud code plus you know, this like I'm doing manual cloud work or like how does your projects work? speaker-0: ⁓ it's a lot of codecs. ⁓ yeah. ⁓ Claude is Claude Design Heavy. ⁓ I use that for a lot of the visual creation before I tell an agent or anything to go on do anything like that where it's gonna impact a UI or UX to where the Claude still kinda got an edge in that. speaker-1: Codex, okay. speaker-0: But it isn't by much. And the design feature they have is still pretty good. ⁓ but as far as like on the client side with the chat to PT agent ⁓ encodec, that you can like the annotations and stuff, the web browser was a little bit faster to release than clots. and the it does a generally better over longer, harder tasks. ⁓ where but I I also have like Claude run Codecs and Claude at the same time via CLI or Codecs run Claude and Codec CLI at the same time. So like six agents doing one thing for from Claude or Codecs underneath it, and then six agents for like Codec script running. ⁓ it just depends on what the task is and how much can be paralyzed or that kind of stuff. And like depends on the work, right? ⁓ what you're trying to do and like how many tokens you want to waste. So it's really kind of like super dependent. ⁓ sometimes it's like a very simple task, you know, whatever. ⁓ it just I make a lot of those like calls where you just like you have so much experience with it, you're like, this is gonna be fine, you can just do this. And ⁓ so in terms of like how they would go about it, I'd be like, give them a sandbox environment, whole project, whatever in GCP, and then just give it G Cloud access, tell it to build a framework around it. That you can use to push G G C P that then locks the AI from having to do any of that shit. ⁓ so you can test everything you need do in this like environment that doesn't matter. You can confirm everything, you can security test it, and then it's just like a package that you can then use and that's what I did, right? So I can then deploy into GCP or really anywhere, ⁓ through an abstracted Rust ⁓ backend that Can take a spec and go, hey, you wanna just fan up this GCP project? It needs these permissions, these APIs need to be enabled. It's gonna need this with account ⁓ set up from GitHub. And it can you know initialize the thing from GitHub and have everything just created from GitHub and the WIF provider. So that's how I did that. speaker-1: So you are you finding yourself building, I guess I'll call them agents, but like whatever you want call it, that like for example, like something that manages GCP to spin up and down services and understands GCP at a deeper level. I would assume Gemini would be the best at that, but it doesn't seem to be. But like, do you do you build that once and reuse many? Is that kind of how you think about architect like as a person using AI, do you spend time writing an agent that does that and then you just keep calling to that? Is that how you're architecting it? Or do you kinda start speaker-0: ⁓ Yeah, like do you have a GCP? It's no longer an agent though. It just becomes code. Okay. It's not it's not like an agent that would know it. It can look at the code and see how GCP is deployed ⁓ through the serverless account. How what is referenced from the GitHub action that so once it has access and you kind of call back to that driving through the rain problem, once you take away that imperfect information, the problem you're having. ⁓ with any AI, doesn't matter who it is what it is, if it can't read it, it's gonna try to search for it. And if it can't find a good good reference or it doesn't have API or internet access to search, it's just gonna make some shit up. But if it has access to the code, and say you've proven out that like this code based on this spec will deploy something to GCP, it will enable these ports, disable the shit, and that's your pipeline, then you tell AI to build a, you know, a spectrum spec that adheres to your pipeline and then it can see all the code that interacts with GCP. And because you already tested it and all this stuff, it will make you a good spec. ⁓ because it doesn't have any of that imperfect information where it's like I can just look at the code and see what's going on. And then I can conclusively give you an answer and how this works. ⁓ so that and then once you have that like kind of like baseline built. You can just kind of go on building whatever. ⁓ I mean I've had I've had AI write like a connector. We could even like for like claws, like we wrote a connector for it and they don't like I don't have access to their API and it it worked for the most part. There's there was like 10 bugs I had to fix past that, but like okay. I mean it doesn't even have docs. I don't get it, but I'm like fine. Like there's some stuff you can get off their site, but Yeah, most of it was just like, okay, I'd run into a condition trying to ingest and I'd just fix the bug and then nobody move on. Like okay. ⁓ 'cause yeah. So it's like once it kind of got to that layer and that's why you see a lot of SDK type stuff on my my threat flux repo, ⁓ is that I built a lot of that foundation to where it doesn't need to like assume a package version or any of that. It kinda takes it all away because it it has like the whole the whole thing to see, like whether it's update Najira or any of that stuff, like I have layers for it all. So it's like file system, object storage, Kubernetes control, even like ⁓ like the way it deploys locally, it can just like cargo run and then spin up a tenant with another cargo run and spin up like a Docker compose for the Postgres, like just optionally instead of running full K3 stat. ⁓ because there's like a abstraction layer above that that allows it to just like execute whatever and and test it can just like you know make run or whatever. It's technically like a a CMM command that does it but stands it up, builds it. It can run without Postgres, like totally just in memory. that's kind of annoying because then restart anyways. Yeah. So That's how I kind of orchestrate it all, because then I can I can have it test everything locally. It can build up and spin spin up a tenant, whether it be like a local K3, is if it really needs to test something like that, where it's like, ⁓ we're starting to change like mappings or something like that where it actually gets into stuff that matters, but half the time that shit don't matter. Half the time it's like, you know, a connector bug for claws and you don't need you don't need all that shit. You don't need no Kubernetes or GKE to test it. You just need to know, does this like key work? Done or does this bug fix? Like it can stand it up, test it, do that screenshot thing I'm talking about, provide you a screenshot, and then you can go test it. right. And as long as you're like you can read code at a level of like don't allow crazy bad things to happen. You just need to know like the environment is saying. speaker-1: How does MCP fit into this? Are you an MCP guy? Are you kind of looking at what vendors are providing with MCP and using it? Or you kind of think you'd rather just build those specs and talk to APIs? speaker-0: So I built my own MCP manager ⁓ that that can import MCPs, that can, you know, interact with custom built in MCPs, blah blah blah. ⁓ I built a whole system for that. ⁓ you know, I think th a lot obviously a lot of the like Claude, ⁓ they use a lot of that in the back end. I mean it's just an API spec for it really. ⁓ know, as a use case for a lot of like easy integrations, but also like a lot of the internal stuff I would do at an agent just wouldn't need access to anything because it would all be like an internal service that you can just have it spin up and get literally isolated access to only essentially like a a forge is not GitHub, but all those types of things where it's still got like its own tasks thing. It's got its own it could just build up a platform and run that too and test it. ⁓ So that allows you to kind of quickly know where all your gaps are at. So I I use MCP in there and allowing to call like the shared threat intelligence service in library that we kind of collect and and that's part of the way we offer to like, hey, you got this gap. Here are some rules that based on the MITRE mappings, like you could deploy in these environments based on that threat actor or just that technic technique in general to base based on what rule it is and what environment they support. speaker-1: So I assume, you know, when you're talking about coding, you're mostly using the frontier models. So in some of these things you're talking about like a building a threat intelligence MCP with an AI behind it that sort of notes all the relevant intelligence like that. Like how are you thinking about like private models? Like are you using a lot of, you know, non frontier stuff in your regular like obviously you not coding, but like the rest of the world of things happening. Is that something that you're messing around with or interested in? speaker-0: Yeah, so ⁓ There are things that the the frontier models do worse. ⁓ like in my testing I found ⁓ with like miter mappings, Claude and Codex both are pretty bad. th they're frontier models, I should say. ⁓ miter mappings due to their over their over report. ⁓ they're very they're very tuned towards even providing an answer. And if there is no mapping, it's more likely to provide a mapping than say like Quen three point eight twenty seven B. Like so for like f Fable or I did only I only tested five six, then I tested Astra to be clear. ⁓ but Haku did better than than Sonnet, Opus, and Fable. Four five, you know, like because it just didn't overreport. That was it. Like the other ones just over report so hard. And I've been fine tuning into like the Quen stack. ⁓ I was talking about like the Quen family, 'cause you can take like their point eight model, point eight B, use as a baseline, kind of see where your training data is at, get signal with that before you start to tune against a bigger model. ⁓ you can actually technically take some of those snapshots and transfer it directly into a ⁓ one of the other models depending on what you're doing, which can save you some like training and And and fine tuning time. ⁓ so I'm looking at that, and that's like the way we're looking at ⁓ to take a a customer's rule set and in an offline environment have it map miter based on a number of factors on hey, these are you know essentially ⁓ a light rag with some re ranking. for what the customer's rules might already do for like their standards and stuff. And then have like a fine-tuned model that is really good at like you've shown the entire chain of like, this is an attack, this is what happened, these are the exact mappings, and you get good training data, you can push that signal to even higher than frontier models ⁓ at that task because you're you're really just tuning towards one direction. You don't care how good it is at. coding or any other shit. Right. Using it for a specific service. And then you can you can do it really fast. You can I actually like I've taken, for example, the 0.B model, I've taken from a 0% compile rate for just generating YAR rules on 10,000 tests to a like 70 something compile rate with like a 69% detection rate on 10,000 tests. That is a point 8B model, ⁓ taking a Spark, just one, ⁓ fine tuning it with seven hundred and thirty-seven SFT samples, which is tiny. ⁓ well, I say that, but there's also like three million tokens in that. ⁓ so it's not it's not really tiny. ⁓ there are some of those are really big ⁓ samples where it ⁓ it kind of talks about like you really want to diversify what you put in there. You don't want to just like focus on one type of thing. ⁓ you don't ever want to repeat yourself. You want it to have extreme diversity in the way the data is presented and answered. is the best way I can say that. You don't like you you never wanna like say you shouldn't have too many of the same file formats in your example sets or too many ⁓ malicious rules in your first set, training set. ⁓ that's something I really big models really ⁓ like frontier models struggle with is they just They think this is all malicious when like they're really bad with intent. Right. Right. And like just because you're writing a detection rule doesn't make it malicious. It could just be a policy violation. It could just be heuristics detection. It could be an alert. Like, well, what does it mean? Right? Like I've written just really stupid ass rules for Yara at Target, Target.com, right? Like speaker-1: Right, yeah. speaker-0: That's not I'm gonna have a lot of false positive, but you what? It turned up some really cool stuff in Virus Total. Like it just really depends on your your goal. And that's where AI is like, well, this would be terrible. And I'm like, but it but what it where are we deploying it, bro? Like it's got my domain name in it. I kind of care. And then I can filter out the rest of it like another layer. Like that's where it's like, ⁓ I don't know false positives. I'm like, but well, what's your goal? Right. Like speaker-1: Right. speaker-0: How are you gonna track if they change like a campaign or a tactic if you don't have like a wider scope you're tracking, you know? Like you don't. You need some samples to go test again. So I'm like, ⁓ that's how I used to do it back at Target. I'm like, here I'll write a basic rule that tracks like these basic imports. So I whenever I need to like hunt for a sample change, just not fire it anymore, I can just pull that set from Boris Total and like compare them, ⁓ and I can easily generally find where they they pivoted. speaker-1: Right. speaker-0: ⁓ so it's like, yeah, just those little things about models, whether like smaller, or you can kinda make it towards a specific task, if you can do that to drop, you know, a lot of cost, you can do that. And it doesn't take a ton of experience or knowledge because like I just, you know, I mean I had like Sonnet three point five back in the like two years ago helped me learn how to start fine tuning back with like Lama three, with like an A V model. And I didn't know what the hell I was doing. I never fine-tuned shit in my life. I just asked it what to do. I set up a fucking vast.ai interface with a fucking ⁓ H two hundred and I I started fine tuning. Right. I didn't know what I was doing. But you what I did know is I knew YAR rules really well. I knew like what was a good YAR rule, like, ⁓ what you should use for a private, like the thing that will throw them all for a loop is like ask it what a private YAR rule is, and I'll be like, I think it's just something they keep private. And you're like It's it's a reference rule ⁓ that doesn't flag. ⁓ so it's like those types of things you can like target in training data when you're working with a model that you know it's bad at, ⁓ as opposed to like just the general stuff that might just get right every time. You don't really need that in your fine-tuning if it's already doing that. speaker-1: Right. Okay. speaker-0: That was a really convoluted answer. Hopefully that helped. speaker-1: No, I it we're all navigating this, right? And like, you know, you think about you know, and and this is so interesting because obviously you layer things like security across it. Like I I the other day I was listening to ⁓ the guy who started open router talk about it. I'm like, you're the fucking devil. Like you're just talking about routing our data to China with just YOLO level of security. And I was like, yeah, not that he's not the devil, but like they're doing bad shit with data. Like they're just whatever's cheapest, flow it there. Who's running it? I don't know. I don't care. You know, it's a mess. And so, you know, the model conversation, and I don't know if we should get the warm fuzzies from the frontier models either. Like I don't trust them either. And, you know, I don't think they're doing great things with our data, you know. And I like fundamentally, like it's all built on theft, right? It's all trained on shit they didn't build. So like I also have a problem with that. But ⁓ you know, just thinking about it, you know, from a security perspective of like, you know, the amount of private data that's being sent into these models is horrifying. Right. Like we're just breaching ourselves all day, every day with this stuff. So there I I believe there's going to be a movement to like some of this has to come private. Some of this has to come local. Like should we really be sending our financials to Claude? Like probably not. Right. And so I just think we need to get geared up to start building these solutions securely, quote unquote. ⁓ so that's something I think a lot about you know, as a security professional, not necessarily as a model expert, just as a like You know, we're in this frenzy, mania, hype, whatever you want to call it, but like it's gonna end. Like it's like it's we're getting so subsidized. The analogy I always use is like, I don't know if you remember five dollar Uber rides, but like there was a period of time where you could get an Uber for like eight bucks to go to the airport because they were subsidizing it. Then all of a sudden it was twenty and then it was forty and now it's eighty. Like that's gonna happen with AI, right? Again, I know things are getting cheaper at the same time, but like, you know, at the end of the day, compute's compute and it costs money and it's not gonna be free. So To that end, like with models and and harnesses and all these things that you're working on, like what have you done and you've like, Well, that was stupid. Like that didn't work, or AI isn't capable of that yet. Like where are you sort of seeing the edges of like ⁓ that really wasn't that great? speaker-0: Yeah, it's on like the intent stuff. ⁓ it's obviously it's gonna be more confident in answers. It's trained on what is the general consensus versus like what might be like technically true. Right. So there's like things like that that it it has a lot of biases in more than one direction that you might not like consider. ⁓ there's like obvious ones that people talk about, ⁓ and there's some that are like you wouldn't necessarily expect based on what's in it or how it like Without going too much into the type of data research that I've done with AI and and how it it behaves, you there's patterns where if it's if it's confident enough in something, it will just respond with an answer. And I kinda talked about how the code breakdown where like if it has all the code, it will go look at it. ⁓ and there's like sometimes it just like won't find what it wants on the internet and it'll just move on. right, like ⁓ so They The more confident they are in a certain pattern, the more they're gonna follow it. And and those can be like super toxic in a coding paradigm. ⁓ because as soon as something's like integrated or not integrated, but has been put into the context to your environment in your code base, trying to undo that can be really hard. Right. ⁓ it really just depends on the perspective that you did it. But if you like, you go to like, you know, ultra secure, ultra blah blah blah. Sometimes it's gonna get crazy about stuff when you're like, dude, like, this this ain't possible. Like everything, yeah, like what you just described would take like twenty things to be broken. And then even then, like, like bro, the admin could do it too. I don't like in perspective of like ⁓ it was like it was like basically okay, I can change the URL of what my connector goes to as a tenant admin, and so can any other tenant admin. And I could point it at a malicious one and potentially get the save token to be leaked. I'm like, well, yeah, yeah, you could do that. Sure. And, you know, that's why we don't have like allow egress to all. Like speaker-1: Right. speaker-0: Yeah, but it it's like those things where it's like, okay, but if egress is broken, yeah, that's like a problem. So like you still need to like think about like, okay, we can still backlog this and make a solution that says like even though they're configuring a host, it still has to be like a certain type of host or somehow or from ⁓ like a higher layer than just like any tenant admin that could just like go in there and change it and redirect and say, Hey, like I want to steal this password, even though like Why would anyone do that inside an environment? I don't know, but like it's not wrong, right? But like at the same time, it's like such an edge case that you're like, yeah, we'll fix it. But like at the same time, like it's it's trying to treat it like a tier one issue to stop like a PR. And I'm like, bro, we'll put this in the backclug and we'll get to it. But like this ain't gonna this ain't gonna break. Like, you know, they put in their token once and nobody touches it. I mean, like, yeah, can another one ⁓ the two people they have have admin to configure those connectors go in there and pop it. Yeah. But they both have access to like the thing that generates a token in the first place. So this is all it's like those things and like that's not necessarily true for every single organization on the planet. So like it's not wrong. But like at the same time it doesn't know. And that's where I'm getting to is like they don't know intent. They don't know what what means or who has access to what and you can just tell it and it'll be like fine, I'm moving on. Cool. You gave me a green light. speaker-1: So who cares? Yeah. It's ⁓ speaker-0: So if you really want to like constrain it, you gotta like build in some strict ass guidelines to where like I got my bot will tell me to go punch sand if it can't, like it just won't override things. It won't even like it can't because of the way in like a skill review it has like a whole isolated container, it stands up, it like feeds the context to the agent, provides everything. The agent can't go outside of it, has no access to anything else. so like even if it like got prompt injected. somehow. Yeah. And then you would have like, okay, the the the follow up agent would it have that same prompt injection, ⁓ because it would get somehow mangled through, like you'd have to get them to like prompt inject you to put the new prompt inject in, which like, no, like I have scripts that will catch that. ⁓ or not scripts, Rust code. There's actually I have a Rust laboratory for it. ⁓ anyway, so it yeah, it's like just just know what you speaker-1: To go where, yeah. speaker-0: you know what access is and that's where like models struggle and you've seen this especially with the release of their cyber security models. They're just terrible at things now. Like they didn't used to be so bad with authentication. But as soon as you touch an authentication bug, they're like, bro, I'm out. Can't I can't do it. And you're like, bro, like this didn't happen a year ago and this just hurting us. Like that's the kind of stuff that like, you know, like speaker-1: Yeah. speaker-0: Does everybody wanna be in their like trusted plan? And and do they have an incentive to like train on your data? Yeah, like why wouldn't they? I mean, if they get to AGI first, the financial incentive to do that versus the lawsuit you could bring to them for training on your data is minuscule. ⁓ speaker-1: Yeah, whenever I hear that like your data is not being used to train, I'm like, ⁓ okay. maybe, but like I I somehow but then like it's almost so like you can trust me not to steal again is the way I read that. Like I stole the thing that got you t to be able to build this thing, but like I'm not gonna do it anymore. Like all right, maybe, but like that doesn't seem the way that that works, but what the hell do I know? ⁓ I was funny, I was talking to a buddy of mine at Netflix and he was like he basically the same problem of sort of unlearning and you know, it's like it's not the what is it? It's not what you know that gets you in trouble, it's what you think you know and just ain't true. Like that's the scariest thing. So like it's he called the wisdom, which I thought was pretty smart. It's like teaching AI wisdom is really art because it it is it's still a computer and it's still a prediction engine. It's like it kind of does dumb shit. So I I think that's that's very valid, and I've had a hard time unscrewing Projects and sort of you give it one rule and then two weeks later it's still following it. We're like, damn, I don't, that was just one time thing. And like I think we all wrangle with that. So ⁓ I just want to be respectful of your time, but like the the last thing I'm curious about is like, what are you gonna do next? Like, what are you thinking about building next? Like what's exciting to you? ⁓ obviously you do it for a living, so maybe the side work isn't as exciting. ⁓ but like what are you looking at? You're like, man, I could I really want to do this, I really want to push the boundary. speaker-0: ⁓ well the biggest thing is to really get ⁓ like a like a blue team model. ⁓ that's more capable. I mean like a blue team model to make it more capable. ⁓ well I I I shouldn't even drop that 'cause don't I kinda think it's all the same thing. Like it's it's development. ⁓ blue team and red team are both just like branches of development. ⁓ Right, like follows mail at the end of the day, whether people want to like agree with that or not. For me, ⁓ a rule is a rule, a log is a log, and file is a file. they don't it's all about intent of whether a file should run for a certain person to access a certain set of data. and same thing's true for AI or threat actors and and and what have you. ⁓ So like Building that paradigm of allows you to give it just what it needs, trained in a way that isn't kind of with those, like where it's okay with saying I don't know. ⁓ which they're not they're kind of not tuned at all to be very bad at that. Right. ⁓ and that's like in cybersecurity, that's like a a pro. Like if if you can identify like what you talked about is like what you don't understand, that's where you can can slag if AI is ⁓ kind of like off track or it needs more information or there's like a gap. So those those are like the frontiers of ⁓ where I would love to if I had the money in the the but we're not financially backed by any any organization and we haven't taken any ⁓ outside investments for anything. So We're just trying to build slowly and properly ⁓ you know, using like the cyber simulation type of building, ⁓ to create its ability to like set up an environment, understand it, its intents and purposes, and be able to test and train data off that, all just pure generated simulated training data to make a model that that can actually be. what these people aren't providing because they're not they're not looking at training at any of this stuff. They're not training at like YAR rules and KQL rules and Yar L rules or any of that other stuff, like Sigma stuff's in there, but it's not it's not really in there. Right. It wasn't you could tell like, ⁓ it's got it's got good Rust training because it can like get so far on Rust or Python and a very basic example. But you go to any of this other stuff and it's gonna start just being shit real fast unless you have a lot of guardrails. ⁓ if you just like sit there and have like these things where you're like you're like me, where you're like running these tests. Well, every test I run is essentially data I can use for training, where it's like I can look at how Fable or whatever did and say, these are the gaps in the leading models, and they're answering this wrong. Does my training compensate? Is it a problem for my training model? ⁓ and I can use those as signals. ⁓ without actually using anyone's actual data because it's like I just built so many ranges and stuff. All it needs is like here's the context of an environment, a potential incident, right? And then you give it that. You need the answer. So like customer data is not good enough for that. It's it's imperfect. It's like giving a black box. You want to give it literally fucking everything that says whether this is right or wrong. ⁓ that's it because that's the key for models, and that's what they're starting to accelerate with coding. It's because they started using all these like people that don't turn off code sharing whilst going through these coding sessions to use to train them up. So if you have something similar where you do that for like, hey, here's a simulated attack, go and then you can collect the data and then like here's the policy from the environment. Here's like what the infrastructure is supposed to be doing. Here's like you give that all as like a set of turn data that it can go through and train on. ⁓ you can build a lot better understanding of like networks and stuff that. You know, they don't have in training data. They're not taking PCAFs and explaining no ⁓ to the model on what a PCAP is. I mean, it's got that from like base shit on the internet, but it's not it's not being tuned for it, so speaker-1: Yeah, just because it ingested some Cisco, you know, instruction book doesn't mean it knows what any of this crap is from a security perspective. So do you think that's ⁓ a s a a solve at a model level, or do you think that's a enough contextualization in the frontier models? Like like which one do you think it is? Or both, I guess. Not to put you in a corner. But like I because there's a lot of security companies going like, we're gonna wash all your stock data through a frontier model through, you know, cloud. And then spit out summaries and do all this great work. I don't know. Like, and then there's I don't really see anybody pushing models. Like, I don't see that in our market. Like these guys are raising all this money and they're just building harnesses for LLMs. Like they're not. I have not seen a good private model for security. And even companies that have raised a shit ton of money to your point. I just haven't seen it. ⁓ so do you think it's a model thing that needs to happen versus what we're seeing now, which is mostly you know, context forwarders to L. speaker-0: Yeah, I think it's the model. I think I think that writing's been on the wall for a couple for a while now. ⁓ and why because like like I said, until Anthropic or OpenAI decide that they want to hire the best blue team, red team, purple team people on the market, they they're not gonna be capable of that training data. And the only reason they're hiring those people right now is for general cybersecurity. Not not for for training. ⁓ so until they decide to actually care. ⁓ and go, hey, what what are we not doing with our data? What's wrong with intent? Because they talk about alignment all the time, but they they miss a lot of shit where it's like, dude, you're you're forcing it to give you answers. ⁓ you've made it singly its most important objective is to finish what you've prompted it to do. Right. It is not, they're not leading it towards the I don't know path. The I need more information path. They did a little bit to have Claude prompt for more like questions, but they still like are very like gonna make an assumption on a certain point. And unless you've like encountered all those parts, you're not gonna like have a skill or harness layer for it. And then it's still shit at intent. So you can see it. Like they're just rejecting everything because they're like, Well, you might be look fucking delicious. And you're like, What? Like there's nothing like, bro, you're looking at You're looking at my own fucking code that like I have speaker-1: Why is this phone? speaker-0: Yeah. Like we're not like you're telling me you can't fix a security vulnerability in my own fucking code that like isn't public and you can like look at and know it's not public. ⁓ and you're like, bro, we're like really just ⁓ I see what you did. You put any of these buzzwords in here and all of a sudden it says F U and you're like, that's not that's not sustainable. That's not for blue teams. ⁓ I finally got access to daybreak and in open or open AI. It's a lot better, but it's still not great at like the refusals in certain certain points. So I speaker-1: Just I saw ⁓ and I was like, I feel like that it's a different guardrails, same training data is the way I felt about it. Like it's it's still got the same answers in its DNA. The guardrails are just I guess I would say better. ⁓ the one I'm interested in that I I was just at at Felcon, CrowdStrike's building like a red blue based on CrowdStrike data, which I think like they have good data. So I'm excited to see what they're doing. ⁓ the you know, the problem with that is always cost, right? It's like It's great, but if you're a top fortune fifty, but if you're a regular dude at a regular company, you can't afford this. Right. So I'm I'm interested to see where that goes. I actually I called him out of LinkedIn. I was like, ⁓ the last thing we need is another damn lab selling super expensive tokens. Like you give it away or don't do it was a kind of my point. ⁓ so yeah, that's very cool. So ⁓ thank you so much. I wanna give you I don't wanna c I already kept you long. ⁓ I love hearing your take on things. I think you think about the world in a in a very cool way. speaker-0: Yeah, sorry. speaker-1: ⁓ obviously people can find you on LinkedIn ⁓ and then your ⁓ threat flux is a is a project. and then countermeasures your company that you work for. ⁓ anything else on the agenda, giving any talks or do anything that people can find you at? speaker-0: Yep, kind of security. ⁓ I'll be running the GERTCON D Digital Friends Exit Response Challenge. and that will be a hosted by a version of our platform that will allow people to work in a simulated environment based on that coverage, visibility, Intel, miter mapping stuff. ⁓ that's like part of the challenge is to improve like a simulated tenant environment. as if it was connected to like your infrastructure. Obviously not your infrastructure because it's it's gonna be like a micro sized environment. So I don't actually have to pull in real connectors and data. But ⁓ it will essentially be a a version of our like, you know, your team will get a version of our our tenant plat our tenant user environment and and they'll be able to compete with each other for that. So hopefully speaker-1: Kirkan is Grand Rapids. Is that where Kirkana? Yes. And that's when is that? speaker-0: ⁓ the twenty fourth and twenty fifth of this month. speaker-1: ⁓ coming up. Awesome. ⁓ some of that content makes its way online. ⁓ this should be up by then. ⁓ hopefully everybody can get a chance to go or or listen to your content. ⁓ awesome as always, man. Thank you so much for for spending some time with us and and teaching us about ⁓ about some AI stuff. ⁓ appreciate you, man. I'll talk to you soon. All right, bye. speaker-0: See ya.