Ep 21: The role of AI in security measures with Casey Ellis
In this episode of the Boring AppSec Podcast, host Anshwan interviews veteran security founder Casey John Ellis. The discussion opens with Casey’s recent health challenges and his recovery, then moves into a deep dive on his founder journey—creating Bugcrowd, co‑founding Disclose.io, and shaping the modern bug‑bounty ecosystem. Casey reflects on the early days of crowdsourced vulnerability research, the cultural shift needed to treat ethical hackers as partners, and the economic logic behind bug‑bounty versus traditional pen‑testing. The conversation then pivots to the current state of the security market, critiquing short‑term vendor solutions that act as band‑aids rather than addressing root causes. They explore secure‑by‑design principles, the difficulty of translating high‑level standards (e.g., CISA SBD, OWASP ASVS) into actionable practice, and how crowdsourcing could be used to enforce accountability and continuous improvement. A substantial portion of the episode is devoted to AI’s dual role in security: AI as a tool that accelerates both offensive and defensive work, AI as a new attack surface (prompt‑injection, model exploitation), and AI as a broader systemic threat when integrated without rigorous design. Casey presents a taxonomy—AI as tool, target, and threat—and discusses real‑world examples such as AI‑driven XSS discovery and autonomous agents like Expo. He weighs the hype versus reality of AI‑generated zero‑days, the limits of current LLM reasoning, and the need for niche, well‑prompted AI applications. Throughout, Casey emphasizes the importance of conviction, focused problem‑solving, and building security solutions that raise the cost of attack rather than chasing fleeting market hype. The episode concludes with take‑aways for founders, investors, and practitioners: prioritize fundamentals, narrow scope, leverage AI thoughtfully, and foster transparent, accountable security practices.
Chapters
Foreign.
Hello and welcome to another episode of the boring AppSec podcast. In this episode, I'm super excited to talk to Casey Ellis. Casey is a security industry veteran with more than 20 years of experience. In this conversation, Casey shares his personal journey through health challenges and his insights into the cybersecurity landscape. He discusses the evolution of the bug bound industry, the importance of secure by default design, and the role of AI in both enhancing and complicating security measures. Casey emphasizes the need for accountability and the potential of crowdsourcing security, while also addressing the challenges of implementing effective standards. The conversation concludes with reflections on the future of AI and security and the necessity for focused problem solving in the industry. I hope you enjoy this conversation. Listen in.
Welcome KC I am super excited to have you here. Why don't you give me a quick rundown of what you've been up to recently and then we can go from there.
What have I been up to recently? Yeah, so hello, thanks for having me. My name is Casey Ellis. I'm probably best known as the founder of bugcrowd, also the co founder of the Disclose IO project. What have I been up to recently? That's a good question. I actually had a kind of an unexpected health issue midway through last year that kind of put me on enforced sabbatical. So I've been sort of working my way back from that, kind of exploring what's happening in the world, continuing to do a lot with Bug crowd on the strategy, the evangelism and kind of the market side of things. But alongside that, really digging into what's happening in AI and kind of this convergence between AI security research and the overall kind of security environment, just in general, as well as a bit of angel investing and advisory stuff. One of the things that I've always loved and kind of got into pretty quickly after bugcrowd started working is this idea that founding is hard. We're all learning things in a million miles an hour and getting stuff right and figuring some things out the hard way kind of along the way. So the opportunity to have folk come and sort of learn from some of my experience and you know, the cool part about that is that I get to learn from them in terms of what they're seeing on the cutting edge. So that's kind of where it's been lately. Aside from that, you know, bit of podcasty stuff here and there. I'm on the review board for Black Hat and defcon, so I've been kind of finding ways to keep myself busy as I've been sort of recovering out of the, the surgery stuff and. Yeah, that's sort of where it's up to.
Yeah, no, I, I'm sorry, I had no idea that you are going through some kind of a health issue, but I hope you're well. Like.
Yeah, I mean, just a, just a quick, just a quick double click on that. I had, I had to have basically open heart surgery midway through last year. So I had a. I had a genetic issue that we figured I'd gotten away with that. It turned out I hadn't. And it was, you know, definitely not on the bingo card, but ended up on the operating table in July of last year and then had to go back in, in August. So, you know, everything's worked out well in terms of the surgery stuff and I'm super grateful for that. But like the. Yeah, the recovery time from that sort of surgery is not, not trivial. So it's been a balance of like, trying to keep myself entertained and get back into, you know, what's happening in, in the industry and building and all that other stuff and obviously treating recovery with the right kind of respect.
Yeah, yeah. No, man, I appreciate you hanging in there and you look fine.
Yeah, thank you.
Appreciate it. I am glad that things are on the mend, which is always nice. So awesome. So, Casey, I had some questions for you. I have. For the audience to get some context. I've known you for a long time now. I think I was probably one of the first early researchers on the bugcrow platform when Jake Ryan and a few others came up with this idea of crowdsourcing bug hunting. And I was really fascinated by it because I was into security and the idea of getting paid to find security volume just prove too good to be true. So I appreciate the fact that you've basically created an entire industry like all these years and you know, you continue.
I do remember that too, like, you know, back in the, like the very early days, like scratches, trying to figure out how all of this works. And you're, you're kind of part in that. So I appreciate that on your end as well.
No, absolutely, man. I think the entire bug bounding industry, you know, has a lot of gratitude to you and others who have basically started this industry and it has. I've seen people's lives change because of the amount of stuff that they can do with all of the awards and whatnot. So again, thank you so much. We really appreciate it. So, you know, in terms of your founder's journey, right, like you've been a founder, you've founded Disclose IO like you Mentioned. You obviously founded Buckcrowd. You know, I'm kind of curious to know how your founder's journey has been. Like, things that you've learned over the years, things that you wish you did differently and things that you wish changed in the way our industry works. Just kind of feel free to go at it, you know, take different numbers.
Yeah, I mean, it's a super broad question because. Yeah, like, I've been, you know, I've been at the bug crowd thing now for. What are we, 20, 25? We kind of kicked it off in 2012. So 13 years. Bug Crowd's not actually my first rodeo from a founding standpoint. I have done businesses prior, but it was the first time, you know, I kind of took a swing at like, let's invent a category. Let's kind of try to disrupt, you know, a problem at a system level. And then obviously the venture backing and all the stuff that kind of comes with that or that we kind of executed on that journey. A lot of that I had to learn very rapidly for the first time. I think the thing is, you know, once you've in the early stages, like, we didn't invent bug bounty or vulnerability disclosure programs, like, that dates back to like Netscape in 1995 and prior to, um. But having come from the hacker community and sort of understanding the mechanics of vulnerability research and just how vulnerable things are, you know, really the problems that I wanted to solve, it was two things. One is like, how do I. And. And you know, the early team, obviously, because it wasn't just me doing the thing I scribbled on the napkin, but it's ultimately the team that makes this whole thing happen. Right. How do I change the operating environment for hackers that work in good faith? Because I think prior to bug bounty, you know, this notion of a hacker being potentially good was not really a thing out there in the broader kind of. Even in the security industry. I think there was a subset of the security industry that got it at that point in time, but everyone else just assumed that if you could hack a computer, you were probably a bad person. So, like, that makes helping really difficult. And I wanted to basically, you know, throw a wrench in that and actually create an operating environment with folks that can do security research that makes things safer. You know, give them the ability to operate without a chilling effect and with better understanding of what they're doing from. From the industry as well. So that was one part of it. The other side of it was this idea that, like, you know, as defenders, really when you think about it like the, the entire reason we're here is because of the bad guys, right? Like, if they packed up and went home, we, we could arguably like all pack up and go on vacation ourselves. Like, they're sort of the leader of why we're all here as a defensive industry and they're creative. It looks like a crowd. It looks like people that are incentivized by, for results. It looks like people that have an incredibly diverse array of skills and motivations and all those different things. And meanwhile, we were trying to compete against all of that creativity and resourcing with like a single pen tester being paid by the hour or whatever automation's available at the time. And I was looking at that problem, you know, running a pen tester company at the time and just thinking like, this is stupid, like, this is like the math is wrong. Do you know what I mean? It's not got to do with how skilled or unskilled the individual pen testers or the teams are. It's the fact that they've got to outsmart all of this and they're just not going to be able to do that. So meanwhile, you've got this entire kind of security research community sitting behind all of that, kind of waiting for their invitation to the table and kind of the, you know, the idea behind bugcrowd originally was like just to plug those things in and kind of normalize it so that I think, speaking about the founder journey and all that kind of stuff, it was a pretty hairy, audacious vision at the time because we had to literally reform the opinion of hackers, turn hackers from being this sort of toxic, scary thing to something that's potentially a product that companies can consume and that actually being the precursor for productive engagement, that was, that was a big boat to turn around. And yeah, you know, I think probably the big thing, you know, that I talk about with founders a lot these days is that there's no shortage of opportunities to chase a quick win in cyber security. Like, there's definitely a lot of, you know, kind of short wind money to be made and basically putting band aids on broken arms, right? And you've seen me rant about that on LinkedIn more than once. But this idea of if you're trying to attack a fundamental problem, developing just a rational conviction around what the future should look like, testing that and making sure that you're not kind of smoking crack and off in fairyland with that conviction, but if it's valid, if it's real, then really kind of almost marrying that thing and going all in. That. I think conviction is the thing that can, you know, take you through, like, the bumps and wobbles and the hard times and the good times that sort of follow, especially when it gets going. And I think that's a. That's a really important thing, and it's ultimately the thing that your team will follow as well. Right. Like, there's a leadership component to that in terms of your ability to actually present that vision and have people, like, adopt it and run with it. So, yeah, those are a few thoughts on. On that side. You asked about five different questions, and I kind of launched into some backstory. So we might come back.
No, no, I think. I think this is awesome. So, yeah, I think you definitely shared things about your journey and things that you've learned. Curious to know how you think about the things that currently exist in the industry, especially when it comes to security vendors trying to sell products that might not necessarily bring value to teams, but they're still vendors at the end of the day. So, yeah, just kind of walk me through how your mental model has evolved over the years with. With respect to what really works and what does. Anyone?
Yeah, that's a. That's a great question, because I think, like, there is this sort of, you know, sales and marketing trope that. It's not even a trope. It's actually, I think, good advice. But when you apply it to cybersecurity, it gets a bit weird. Like, the whole idea of, like, sell pain pills, not vitamins. Right. Like, if you're trying to. If you're trying to sell. If you're trying to create a product that you're trying to fit out to a problem that you see in the market, you've got to, you know, attach it to some sort of pain or some sort of urgency. And I think that's absolutely true, because people don't buy stuff unless they feel like they really should or need to. Right. But the problem is that if you just do that, then you end up kind of doing that thing that I was just talking about before. It's like, I'm going to create a band aid for a broken arm. I do believe that in security, just in general, and this is really a product of hacking the Internet at scale for the last 12 or 13 years. It's the fundamentals that are screwing us, right? It's not all of the fancy things that we talk about on the edges and whatever else. Those all matter. But it really is the basics and the need to make secure Easy and insecure. Obvious that we've sort of failed at, I think as a security industry, just in general. And I think, you know, the kind of founders that I like to work with and that kind of, you know, all back from like an angel investor invest in, sorry, an angel investment or an advisory standpoint are the ones that actually want to go after the root cause. It's like let's, let's try to alter the fundamentals of the math that we're dealing with here when it comes to defense. And that's hard to do because on top of that, know, you've got people that like want to treat their headache and you're trying to sell them vitamins. So you got to figure out how to market that, how to position it into use cases, how to, you know, take the success of that and create reference marketing like network effect for reality. All those different things. Yeah. Which, you know, all of those things we've, you know, I've had exposure to and I've, I've kind of had the opportunity to lead throughout the course of building bugcrowd because it, you know, it was a absurd idea at the outset. The folks that were looking at, it's like that's never going to work. People aren't going to accept the feedback of hackers. Like, you know, we're not that vulnerable anyway. There was all of these different objections and I kind of knew all of those were wrong. And I had the conviction to press through that and had the team with the same conviction to actually support that. But in the middle of that, you've got to actually take this sort of big picture and then sort of distill it back to things that folk can access in the present as well. Yeah. So going back to your question, I do feel like the industry has a really kind of short, like a fairly narrow focus on whatever sells best in, in the short time frame and whatever can sort of, you know, demonstrate the, the greatest kind of hockey stick of growth. That's important. But I don't think it's the only thing, I guess, is what I'm saying. Yeah, yeah.
No, actually I'm glad that you hit this sort of point of, you know, we're not necessarily making secure things easier and we're not making the insecure things obvious. I think I really believe in that. And I think this is also a good segue into my next question I had for you more about the secure by design space. Right. So, you know, as a practitioner myself, having been in the industry, having worked for multiple enterprise and early stage companies. One thing I've realized is that, you know, folks generally want to buy certain things or try to solve problems that they see in the industry. They, they don't necessarily try to analyze the root cause of why this is a problem in the first place. Right. So this is where the whole security design comes in, right? Where yes, I think people are still trying to make sure that this sast their dast tooling their programs are effective and that is essentially scanning for vulnerabilities in code that is already shipped. Right. Now if you really try to analyze all of the issues that you see in your program and if you try to identify trends and see if there's a significant root cause that could be addressed, that eliminates entire vulnerability classes, I think there's much more value there. Right. I think this is where the world I see more evolving towards because as an application security engineer myself, you know, I don't want to be dealing or triaging the same kind of vulnerabilities day in, day out. Right. I want my mental.
You want to ideally improve if you can, right?
Yeah, yeah, yeah, yeah, yeah. So, but, but on the flip side, you know, at least from my experience for the past few months, I found out that it's really difficult to sell or build a product in this space when it's not quite obvious that it is what needs to be solved in the first place. Folks and companies are still dealing with vulnerabilities. They want a solution to manage them, but they're not really thinking about, okay, why are these volunteers getting introduced in the first place? Can we do something about it? So you've built companies, you come up with this idea of crowdsourcing. So I'm curious to see how you think about, like if you were to crowdsource a company in the secure by design space, how would you go about it?
Yeah. Okay. So specifically thinking about crowdsourcing around secure by design, I think, because I agree with everything you just said, I think the other, the other piece that often gets missed is, you know, this idea that like security practitioners greatly overestimate the ability for, you know, normies, so to speak, to think about risk and process the prioritization around risk in the way that we do. So the fact that we can just identify something that is non compliant or violates an SBD principle, I don't even think that's necessarily enough. I think there's this aspect of reducing the cost of that feedback loop to as little as you possibly can so that you can fix things like encode pre production or even fix things in the ide. I think that's an approach that, you know, I think AI has actually opened up a lot of opportunity to provide that feedback real time to developers. And there's some companies that I work with that are taking that approach. And I think that's an approach that's sort of dawning on the zeitgeist at the moment, which is kind of cool. But the other side of it is like, why do we care? Why is this actually important? Like, are we able to, like, properly articulate to, you know, the person who's ultimately signing the check to go through the additional inconvenience to do things securely, why they should do that? And, you know, I think we've sort of rested a little bit on this idea that, like, oh, cool, people get security now. So, like, we should just tell them what's wrong and they'll just automatically understand why that's important. Golf and fix it. And I don't think that's right. Like, I don't think that's accurate. Or actually, even better put, I don't think that's a resilient kind of like market thesis for the security industry. And you can see the evidence of that over the past couple of years is there's been kind of, you know, economic headwinds, and a lot of this stuff has sort of gotten deprioritized in the process. So, okay, there's a sales and marketing issue in there somewhere, right? And there's a. There's a value story problem that we're missing. This is why I rant about marketing a lot in context of security, because, like, you don't go through the expense and the effort and the pain of putting up bars on your windows unless you know that you live in a bad neighborhood. Right?
Yeah.
And that whole moment of realizing you live in a bad neighborhood, that happens because you've been broken into. It's because your neighbor's been broken into, or because you've been at the local coffee shop or bar or whatever and you've heard a story that helps you actually realize that that's an important thing to do. So that, to me, that last part is the bit that was sort of missing in terms of prioritization of vulnerability, discovery, mitigation, avoidance, full stop. And where I think the crowd comes into that is this idea of being able to demonstrate this is how easy it is to actually exploit this stuff. Crowdsourcing. The idea of trying to create this story of what a bug is actually worth. Because as defenders, our job is Basically, in my mind, if you boil it all the way down, it's to increase the cost of attack to the point where the kind of adversary you're likely to face, it becomes economically irrational for them and they go somewhere else. This whole idea that, like, you can reduce vulnerabilities to zero, I think that's a, that's a fool's errand. And I think the market's starting to properly cotton onto that now. But it begs the question, well, now what? Right? I think using the crowd to be able to actually help them really understand, like, how easy or how inexpensive it might be to exploit failures, anti patterns, vulnerabilities in the systems, and then what kind of potential economic damage they can do in the process. I think that's something that crowdsourcing's. Obviously, I'm sort of bringing it back to a bug bounty use case, which is not the goal here. But I've seen that kind of be a byproduct of bug bounty in a lot of organizations that I've worked with over the years. And I do think that sort of feedback loop, like, hi, we're your friendly neighborhood locksmith, like, we could be a burglar, but we're not. But let's tell you about what we would do if we had malicious intent. I think that's, that's a really important kind of step in the process. It's probably missing right now.
Yeah, no, absolutely. 100% agreed. You know, I think the, like, the simplest example I can think of is a lot of companies have programs these days, right? And I'm pretty sure most of them, they don't go the extra mile and they try to analyze root cause. Okay. So they keep seeing the same kind of issues every day. But have you ever taken a step back and considered, okay, what can we do about it to eliminate this class? Nobody does that just because we are still in that, you know, there's a fire.
Some do. But I completely agree for the better part. The better part, they don't. And, and if there's a. If pressure comes on that team to deliver things quickly, or if pressure comes on that team to kind of reduce the cost of delivery, then those lessons can have a tendency to go out the window pretty quickly. So. Yeah, no, I fully agree with that part. I mean, I think the other part here is that, like, we're talking all about stick and like, ah, this is scary. Yeah. The reason, like, I worked with Jack and Bob and a bunch of folk on the CISA side of things around the development of secure by design. And there's a component in there that includes vulnerability disclosure programs and you know, Safe harbor being included in that. And it draws a lot from Discloser, which I worked with them on around election security and a bunch of stuff in 2020. You know, I think the reason that it didn't, I think everything that's in the CISA SBD kind of document is the right, like they're the right set of things to be working on. And this goes back to that idea of like, we can't, like, let's not worry about trying to figure out how to cure cancer until we've figured out how to remind everyone to wash their hands after they use the restroom. Like, let's do the fundamentals first. Because like, their lower cost is a lesson and they're probably. The likelihood of them actually creating an issue is higher in the process. And it's just economically rational to approach it like that. I think they're all the right things, but I think why it didn't necessarily get the traction that it could have is that there's no, there's no penalty for saying that you were going to do it and not doing it. And I don't necessarily think there was a strong enough kind of reward out to the public market to say, hey, this organization is more mature. Like we're not like attesting to its security or any of that other stuff. But what we are saying is that based on these things, they actually are ahead of the pack when it comes to security maturity. So therefore, as a user, be IT consumer or B2B, you know, you can, if you're factoring that into how you're buying stuff, like you might want to look into these guys as a less risky option and actually creating kind of a positive feedback loop on the marketing on the market side, I think that's a piece as well.
No, absolutely. 100% agreed. You know, I. When Sisa came up with this, like this whole secure design, I saw a few companies like sign the pledge. And I think Okta is one of the few companies I've seen who are actually sharing how they are working towards that pledge. Pledge, which is. Which is very good to see, right? Because we need more such enterprises that come out and say, like, look, we signed this pledge, this is how we are working towards it. And I don't see that enough off. So I think, I think it makes a lot of sense from that perspective. There was one other thing.
I think you can crowdsource accountability. So going back to your question of how would you Bring crowdsourcing into that.
Yeah.
This to me is one of the fundamental things around vulnerability. Disclosure in particular, like, forget about bounty and just think about, see something, say something. It invites feedback and accountability from the outside world in a way that like, kind of demonstrates transparency and maturity inside that organization. I think if. If you apply that out into going and looking after, looking for specific things that, that sit within kind of an SBD framework, that could be. That could be an approach there.
Yeah. I think one thing that I've been thinking about quite a bit is, you know, so you know how there are different compliances. There's like SOC Tool. Some companies do these things because there's a compliance mandate. Right. With Secure by Design, you know, there are things like OWASP, ASVs, Application Security Verification Standard. It is, I think, in their fifth version or something, but I wasn't quite sure. And maybe I'm just curious, like, why weren't these standards mentioned in that secure design? Because, you know, if we start saying that, look, you can get to a secure design state if you start considering the ESV standards for each application that you build. Right?
Yeah.
Do you have any thoughts around that as to why that didn't happen?
Yeah. Not to speak too far out of school on that one, because I want to. I want to be respectful of that, but as a general design principle, I think a lot of. A lot of the thought that went into the CISA SPD document was to make it as kind of simple. This whole idea of simple is strong if the fundamentals of what are trying to be achieved are there and you haven't reduced them down to things that don't matter anymore, but you're removing as much prescriptiveness and complexity north of that as you can, recognizing the fact that, you know, everyone's going to have their own snowflake approach to doing this thing because every company's different, you know, every application's different, like ASVs, you know, I know Dan and some of the folk that work on that really well and like, they work their asses off on trying to make it as broadly applicable as they possibly can, but it's not 100% because there's just edge cases that pop into the mix and create a problem. So to say that if you don't meet this bar according to this particular standard, you therefore fail, it kind of trips off.
Right.
So if you just bring it back to the fundamental and say, no, do this thing, there's lots of different ways to do it, but make sure that you're doing the Core thing and demonstrating improvement there. I think that was a lot of the design kind of theory that got applied to the creation of that document.
Yeah, no, I think I've heard the same argument from multiple people. Is that saying that you can build secure by design, saying things that you can eliminate entire vulnerability classes is easier, but when you actually start doing it in practice, you don't, you don't realize the disparity of systems, the heterogeneous environments that companies have, especially big enterprises. And it's so difficult to address each and every edge case, each and every.
Well, yeah, and you call out an important thing there as well, is that we do, you know, especially like if we've been kind of, you know, bouncing around Silicon Valley for any, any period of time. Right. Like there's, there's different kind of technology, generations of attack surface that we're dealing with here. And I don't think particularly kind of the, the, the cloud and the app sec and, and now kind of the AI security group as well necessarily consider the fact that like the bulk of the infrastructure that runs the Internet is like 20 plus years old. Like, it's not how. Yeah, like CICD pipelines and all these kind of fancy ideas that we, that we kind of throw around as just the easy solution to the problem. Like they don't have that available to them because they've got this massive incumbent set of infrastructure that they can't just bolt that stuff on top of. Right. So, so again, like this idea of like, how do you bring it back to the principle that could be useful for those kind of more like legacy, larger organizations, but also can be applied. You know, the easiest way to do this stuff is if you're building something from scratch and you can just build it in from the ground up. Right. But not everyone gets to do that. So it's. How do you kind of, you know, come up with a solution that applies across the entire spectrum of attack surface? Not just the, the kind of, the, the, the shiny new stuff.
Yeah, no, yeah, I think it all resonates with me and based on what I've heard and for experience, yeah, I think it's the right direction to go towards, however, how we get there. You know, the implementation, the table is in the details of the implementation. And I think each company is a snowflake, like you mentioned. And yeah, it's, it's kind of very difficult to solve it across the industry, across different verticals. Right. There's healthcare, there's financial, so you got
to bring it back to the primitive and just make sure that everyone understands that and that everyone is at least directionally working towards that, because if they can demonstrate continuous improvement, then all of a sudden, you know, we're all getting better as a whole at that point.
Yeah, yeah, no, it makes sense. Cool. I want to quickly jump topics here. I want to see what you think about AI and how AI has had this huge impact on both sides. Security for AI and AI for security. So, yeah, just feel free to go at it. How you think is AI impacting, whether it's good or not, and how do you see us in the next six months to a year?
Yeah, good. Like, good or not is an interesting question because I, you know, I think, like, I view the Overall impact of AI on AppSec in particular, but just, you know, if you think about, like, system level, like information security and cybersecurity as a whole, I don't necessarily see it having. I. I don't sort of see it as an independent actor in that sense. It's. It's more that.
It's just.
It's gonna basically surface a whole bunch of that we've been trying to solve for a really long time and kind of make the obviousness of those things way more kind of clear. Because it's an accelerator like that. That's. That's literally what it does. It speeds absolutely everything up. Like when. When ChatGPT first dropped, I was doing some work with the White House, throw on CD on creation of the National Cyber Strategy and then the AI executive orders. And it was funny because when it dropped, it's like everyone just kind of freaked out all at the same time. They're like, this is super scary and powerful and really exciting, and all of the upside was obvious, right? But because it was so obviously powerful and because it was usable by anyone, regardless of their skill level, it became an issue of retail politics really, really quick. So, of course your legislators, their immediate reaction to that is, we've got to write stuff to make sure that this is safe, which is what happens. But the thing that was interesting that came out of that was people were saying we need to write policy around AI security. And if I'd asked them what they meant by that, no one could really give a straight answer. It's like, oh, we're worried about this particular aspect of it, like bias and the data side, or we're worried about it using. Getting. Being used as tools in the hands of the adversary, or all those different things. So I cooked up this taxonomy of basically, AI is a tool, AI is a target And AI is a threat. So AI is a tool is basically the use of AI to decrease time to success for attackers or defenders. AI as a target is really all of this kind of net new stuff that's exploiting LLMs and kind of AI systems themselves. So prompt injection, like bias monkeying about different things that can happen there, all that kind of stuff. And then AI as a threat is kind of the unintended consequences that come from integrating this stuff into what we've already got and how we already think. Right. I think the last one honestly is the one that kind of, I would say it's bad in the sense that anytime a new technology stack hits the market and there's a ton of pressure behind it, and then everyone competes to integrate that thing into what they're doing. It's like MCP. Yeah, MCP. I mean, you go back historically, you think about IoT as a really good IoT.
Oh yeah, I used to remember those programs on BugCloud.
Yeah. You think about sort of in reverse this same phenomena. You've got this new flashy technology that's genuinely awesome and transformative. But what happens is that all of the people that want to use it, you know, get really busy just cramming it into whatever they can because it becomes competitive. Right. And then, you know, security is kind of, well, security is the child of quality and quality and haste don't tend to play nice together. Right. So it's just you end up kind of accumulating technical debt that you end up having to pay back in some sort of trash fire in the future. And like the Mirai botnet was a really good example of that with IoT. You know, there's been a bunch of similar things that have happened with cloud, like, you know, Docker and kind of infrastructure as a service had a similar phenomenon. I think it's no different for AI. So yeah, the fact that like we're not very good at, at, you know, providing like design and implementation feedback, you know, early in the development process, I think AI is going to really accelerate the obviousness of that. Like the good news is that I also think that it's a really powerful tool to actually help solve that problem to some degree. But I think overall it's kind of bad on that side. And you see like, yeah, you see a lot of different examples of that. Like on the, as a tool piece, you know, I think it was last week there was a CVE that dropped that actually got found by Expo who are doing like AI pen test focused on web apps. Like Ooze and Nico and all those guys, which are. They're awesome. Um, but they found basically a. It's an XSS that can be triggered by rewriting a URL. Like, it's bad design and then it's vulnerable to implementation, to sanitization issues in terms of how it gets passed on the back end in a Palo Alto networks firewall around a vpn and like, AI found that, and then it got picked up and dumped into a NUCLEI template. And at the moment, like every man and their dog are just spraying the Internet trying to find this thing. So, like, I think the frequency of that type of event is going to increase. And that's old code, that's not new code, like the Palo Alto firewall. So there's just a whole bunch of different things that were already sort of broken that I think are going to become more obviously broken and we're going to have to think differently about how we solve those problems going forward.
Yeah, no, I think you bring up a very good point is, you know, we are still dealing with sequence. Those things are not going to go away. I think they're only going to get magnified and much easier to find. And then this is just the old issues. When we think about the new issues that AI brings, right? Problem, context, stealing, whatever it is, the different attacks that are coming up pretty much every day about how different MCP servers are getting inspired and whatnot, I think the world, especially from a security angle, I, I don't, I don't know. I don't know how I feel about it. It just feels like we have to do better as an industry and focus more on what really matters, which is, you know, building cigar systems to begin with.
Yeah, yeah, I, I think that's. I think that's right. I think, like, because, you know, part of, part of my job is to like, do like stochastic analysis, right? So you take sort of data points and try to infer a future trend based on like a limited set of information. So you don't necessarily have a trend yet, but you can kind of extrapolate where things are going and just really seeing around corners. That's like the founder geek. I think in a lot of ways one of the things that keeps on driving me back to this whole idea of attacker versus defender economics is that historically, as defenders, we've identified a problem or identified a systemic weakness or threat, come up with a defense to that, and if we're successful in implementing that defense, then the attackers are like, this is too hard. And they Come up with something else. Right. And there's this sort of cat and mouse thing that goes on that sort of defines, I think, a lot of how we prioritize security just in general. And it's not just cyber. Like that happens in the real world too. Right. The thing that's happening with AI is that like that kind of OODA loop, like the, you know, observe, orient, decide, act. I think traditionally it's been kind of long enough for people to fit into the middle of it and figure out what's going on and start to predict what's going to be needed in the future. I see that loop getting so tight that we kind of can't anymore. At which point we've got to think about like, what's the actual economics of what we're doing here? Like what's the, what's the economics of attack for the adversary and how do we start to like impact that on, on their end? Like how do we take the fight back to the bad guys? You know, within reason. But thinking about that as a system level thing is like we're just probably going to end up needing to go in that direction because the things that we do right now are going to fail to work at some point in the future. So yeah, there's some interesting kind of thought experiments there, but that's a lot of what I've been sort of, you know, thinking about in my spare time over the past little bit.
Interesting. Yeah, you spoke about Expo. You know, I've been following Expo's journey.
Yep.
Pretty closely. I think, I think it's fascinating to. I think I just saw yesterday on my video yesterday that they, they are the number one on the US leaderboard last month or something like that.
Yeah.
So that basically shows, and this is like a real world bounty programs, they are no longer sort of the VDP leaders anymore. I think, I think the whole industry is very worth, very fascinating to begin with. Now I think it's even getting more interesting. Right. Like the fact that autonomous agents hack agents, go and find vulnerabilities, exploit them, you know, create reports. All of that stuff is pretty wild to me. Right. So yeah, I guess, I guess my question to you is, you know, as one of the founders of the leading bug bonding platforms out there, how do you see this sort of impacting the bug bond industry and just the offensive security industry in general?
Yeah, I get asked this a lot because it's. Yeah, there's, there's definitely like it's gonna, I mean the net net will be that it will trigger A, a need to uplevel, I think across, yeah, the, the entire bounty community and, and the entire vulnerability research community just, just full stop, right. If you're playing that particular game where I don't necessarily think it's gonna like make the entire like bounty space irrelevant overnight is that like there's, you know, Soul Typhoon did their thing using like end day in, you know, edge edge access equipment. Right. Like there's still a lot of vulnerabilities that have been reported to organizations and have either been fixed and not been deployed out into the wild or haven't been fixed because they're just on the queue somewhere or whatever else. Finding the bug is not the only part of the problem that we need to address here. So that aspect of it, I think the space itself and what people as program owners see as being valuable, I can see kind of incentives shifting around to think more about the behavior of threats, more than just kind of the enumeration or identification of vulnerabilities. So there's that. But yeah, on the actual like skills competition side, you know, it's everything I've seen out of, out of. There's a bunch of different vendors playing in this space now. Like Expo is definitely, I think, the kind of the noisiest and sort of the furthest along from a funding and execution standpoint. And like Oosh, the founder there, he's the guy who wrote Semel, which eventually became co pilot. So that dude knows what he's doing. From a CEO standpoint, he's a great guy too. So there's all of these things in the mix. But then in the meantime you've got all of these different students coming out of universities writing papers on similar things. You've got, you know, different sort of versions of Expo kind of popping up all over the place. And the thing to me that they all have in common is that they, they sort of lack the creativity to go kind of second or third order with, with vulnerabilities that are more difficult to reproduce. Right. Like they really, really good at, you know, the, the, the bug that I just mentioned. Like you've got like a URL with an instruction in it and they went through that URL and tried to figure out like how they could get XSS to fire. Once they found something, they then doubled down on that vector to see what could be done from an impact standpoint. Like that's, that's cool, like that's really creative and I don't want to diss that, but you don't need to know that much as a security researcher to get to a point where you're able to do that type of research. It's, it's just that like this stuff is able to do it very quickly at scale, without sleeping, you know, all that kind of thing. So I sort of see it kind of clearing off the bottom tier of bugs when it comes to complexity and ease of discoverability, if that makes sense. Thinking about vulnerabilities or attack chains that require more context, for example, just as a practical AI limitation like context, windows are a thing, right. So if you've got multiple systems interacting with each other and you're trying to exploit their relationships, it's going to be a while before AI gets good at that, for example, and it's going to need focused effort towards that particular type of problem before you'll see a system pop out that can just kind of walk through the park and get that done. And I think it's just going to be an evolution of that over time. So, yeah, for hunters, understanding how this stuff works, making sure you're using it in your own workflows, I think the biggest piece is just getting AI native. If you're in security research, like if you're kind of treating this like a fad or treating it like a thing that's not that relevant or whatever, then I don't think that's a safe assumption in terms of, you know, what you can do in the future and getting comfortable with it. It's probably the best idea.
Yeah, no, absolutely, 100%. You know, I've, I've tried to build a few hacking agents myself and I was mind blown at how easy these things have become. Like anybody can just do things now. Like there's no longer restricted by, you know, you know, to go find an expert to go speak to, like Charge Cloud, all of these sort of services, they're there to train you, to teach you. And I think using them in a way to actually enhance your skills as opposed to, you know, just kind of ask them questions and copy paste that, I think. So there's a fine line here and how to use AI systems effectively versus how we just kind of. So I think it's fascinating and I think from what I've heard from you now is there's a lot of competition, especially in the offensive security sort of AI native space. And I think what will really matter is how you differentiate by, you know, how deep can you really go. Like the second or the third order sort of vulnerabilities, Right?
Yeah, yeah, well, and I mean, even you Know what kind of like what is the use case? What are the things that you're going to develop an expert at doing? Because I do think a lot of the offensive solution stuff that's come out over the past while has tried to be. It's tried to sort of take from like oh yeah, I'm talking to Claude and I can ask it anything and it can tell me the thing and it's better at certain things but it's kind of a generalist. People have been trying to build out solutions that sort of mimic that. I don't think like AI is a long way off being good at that. So the whole idea of like yeah, I want to have something that, you know that example that I just gave before, like understand the full context of all of the systems within an environment, understand their communication and trust contexts and then go off and exploit those like that's a very specific thing to do. Right. So actually building for those use cases or use cases that look like that I think, I don't know for folks that are building in that domain that's how I'd be thinking about it because a mixture of agents and mixture of experts approach with a particular goal or a particular kind of strategy in mind that might have multiple different paths within it, but you've got this sort of set of guardrails that you're trying to build around and follow. Those approaches seem to be the most effective. So we'll just see more and more of those pop up I guess is what I predict there.
Yeah, no, absolutely. I think Apple released a paper I think last week or maybe the week before that about this illusion of thinking where I think the gist was that they don't think LLMs are good at reasoning and then I've seen arguments on both sides. I played with some of these models myself. I'm just curious to know what is your take that? Can LLMs really reason and do you think these offensive sort of security automated security platforms using AI they can actually find some zero days and whatnot? Because I still haven't seen any company finding zero days day in, day out. They're finding stuff in open source stuff and they're finding stuff that are low hanging food source speed but haven't seen any groundbreaking sort of discovery. I know there were a few blogs, there was one by Project Zero, I believe and then there was a new CV recently released which was mostly done by reverse engineering using AI. Right. So I've seen those kind of use cases but I haven't seen a company or a vendor or a tool kind of claim that they can find zero days every day. So, yeah, just kind of curious to know your thoughts on how this is, this space is evolving.
Yeah, I mean, there's a lot of work going into that and I think there's like the progress that's coming out of the work that's going into that is probably not turning up in papers at the moment, is probably one thing that I'd call out there, because there is definitely research and build around this type of problem happening inside larger organizations. And I think there's a lot of work around solving these types of problems happening on the high side and kind of defense and military and the intelligence community as well. So the true state of the art, I think for the purpose of the podcast, we can only kind of really surmise on that. Right. But in terms of, you know, where it's actually up to kind of on this side of the wire. Yeah. Like, you know, like the O3 thing that dropped last week or the week before, you know, someone finding a kernel vulnerability.
Yes.
And making use of O3 for that, like it found the vulnerability autonomously, I think it was 7 or 8% of the time. And really what happened was the researcher actually used it as a tool to kind of fit and kind of supercharge their own knowledge of how to get things done and then was able to like successfully complete that chain and weaponize and, you know, do all that. Right. I think that's sort of where we're really up to when it comes to more complicated issues. I think on the flip side of that, if you're talking about go execute a past the hash attack, bounce around inside an internal network and get the main admin, that's been happening for a while now, that's already kind of off the leash in terms of offensive deployment of LLMs. And to me, what that says to the whole reasoning ability of all this stuff with zero day or domain admin or root as a eval. Right. Is that it's very good at inferring what it should do next from what it's already been trained on and things that it's seen actually generate success in the past as a function of its training data or whatever's in its rag data set or whatever.
Right.
But in terms of this whole truly novel reasoning thing that like Apple was, was talking about. Yeah, I, I kind of agree with the Apple paper in a lot of ways. I, I, I do think it's a little bit academic at the same time because, like, your ability to reason is limited by what you know, and that's going to vary based on the operator. Right. So this whole even I. I get squawky about this even when, you know, the definition of AGI comes up because it's like, well, according to who?
Yeah. What is AGI?
There's lots of different ways that you can kind of decide to set the. The cut line on that. And this is no exception to it. So, yeah, look again, I think it's really going to be a matter of these tools being worked on and refined around specific kind of domains of purpose. And then for the better part, I think having human operators be the training source for that and kind of the beneficiary of that. And ultimately you'll see learning loops where the operator is ultimately training the machine to get better. Like DARPA and the AICC are doing a ton around that type of thing. Yeah, there's a whole lot going on. I mean, the sky hasn't fallen yet, I guess is probably my big call out there. But it has in some ways and I think will in terms of the acceleration of risk when it comes to end day and not these like, oh, there's a ssh, like unauthor privesc or whatever else that you can kind of get from the outside that we found with AI and there's hundreds of them everywhere and we just don't know they're there. That might be true. I don't think AI is going to get us to that answer yet. But there's a whole bunch of bugs downstream from that that are everywhere that it's very good at exploiting at this point. So, you know, as a defender, that's what I'd be thinking about there.
Yeah, absolutely. Sounds good, Casey. We are very close to wrapping up.
I feel like we've solved all of it, though. So that's good.
Yeah. You know, I think, yeah, only if conversations and ideas could solve problems in the world, you know, I would be doing that every day, all day. But at least, like, we need more people to speak about these things. Right. And sort of share how they're solving some of these challenges in their organizations and whatnot. Yeah.
So, yeah, I think the directional signposts that I guess, you know, as takeaways that I'm trying to throw up here is that if you're working on like, this is an area that I'm digging into a lot. There's a lot of interest in it from the Venture side. I think there's a lot of interest on it in it just in general. From the market right now and people that are working on it, like, share your research, reach out, do whatever. But also this idea of, like, don't try to be all things to all people. I think narrowing down the set of. The set of goals you're trying to achieve with your. With your systems, like, I think that's, for the foreseeable future, going to be the best way to approach actually creating value there.
Yeah, yeah, no, absolutely. I think that's been my experience as well, is if I try to solve too many things with AI, I absolutely fail in 100% or the cases. But on the other hand, if I know a problem really well, if I try to solve it, you know, using AI by giving it the right context, the right prompts, you know, it can actually really solve difficult problems, challenges that I spend days, weeks on. So, yeah, yeah. I think focusing on a very niche or a very specific area and solving it in a way that works consistently, reliably every time. I think this is going to be the key of how folks can build companies and whatnot.
Completely agree.
Awesome. Casey, I had a lot of other questions, but I think I'll have to have you again at some point, but this has been a very wonderful conversation. So, again, thank you so much for joining me in sort of discussing all things AI and cyber security. I really appreciate.
Excellent. Yeah, thanks for having me on. Anshwan, good to see you again.
Yeah.
Thank you for joining us. If you like the show, please share this with your friends and colleagues. If you have any questions or comments, you can reach Anshwan on Twitter at anshumanbh. That's a n s h u m a n b h and me Sandesh at jubba on jeans That's J u b b a o n J e a n s See you next time.
More from Casey
The Founder Who Convinced Google to Let 10,000 Hackers Attack It | Casey Ellis
Black Hat 2026: Why Threat Researchers Are Hoarding Zero-Days | Casey Ellis, Bugcrowd
Security Teams Are Gaslighting Hackers (Vulnerability Disclosure) - Hackers On The Rocks Podcast
