TL;DR: AI asymmetry is real. While AI is clearly helping attackers find weaknesses across the enterprise attack surface, is it helping defenders just as much? Where is the benefit of AI felt the most in cybersecurity? Join us and listen in on the discussion. Guest Anton Chuvakin (he's baaaaaaack) YouTube Video: https://youtu.be/PCOBz2O9GG0 Description AI is getting blamed and praised for everything in cybersecurity, but the real question we keep coming back to is timing. Right now, attackers can point a model at a target and move faster, while defenders wrestle with a messier reality: “defense” spans SOC operations, vulnerability management, application security, identity and access management, and the unglamorous work of fixing what’s broken. We talk through why that asymmetry makes AI feel instantly useful on offense and frustratingly vague on defense, even if the long game may shift. From there, we get concrete. We dig into defensive AI use cases that survive enterprise constraints: relentless asset discovery and exposure management, smarter detection and response in the SOC, and using AI to reduce patching risk by predicting blast radius before changes ship. We also challenge a common pattern we’re seeing in the market: expensive “AI SOC” products that replicate what good scripting, better queries, and better process should have done years ago. If you’re evaluating tools, this helps you pressure-test vendor claims and define measurable outcomes. We close by debating where AI might matter most beyond real-time response: application security that removes classes of vulnerabilities, identity security as the next frontier with service accounts and agent identity, and whether the crowded AI SOC startup space is headed for a sharp consolidation as major platforms bundle “good enough” features. Subscribe for more security realism, share this with a defender who’s drowning in AI promises, and leave a review with your answer: where should AI help defenders most next? Have something to say? Let's hear it. Support the show >>> Please consider clicking the link above to support the show! -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= YouTube home: https://youtube.com/playlist?list=PLyo0dkKRvfVtWXjRxNISrhme1MgBj3C2U&si=scHDiTuLXSEQ9qHq LinkedIn Page: https://www.linkedin.com/company/down-the-security-rabbithole-podcast/ X/Twitter: https://twitter.com/dtsr_podcast
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_00They say they say we should have known better than to also deep down deep down into this have a second. I
SPEAKER_01didn't think of a new stream featuring Mr. King. Taylor on the one and only Down the Security Rabbit Hole podcast. Plug in, listen up, and enjoy
SPEAKER_00the spoon. Don't let the fact seemed the last time that we were.
SPEAKER_02I love that intro.
SPEAKER_03That rabbit.
SPEAKER_02I love that intro. I do. I enjoy it. Ladies and gentlemen, welcome down the security rabbit hole to another edition of the Down the Security Rabbit Hole podcast. Raf here. Below me is James. Over that way. That's Anton. Anton's back. And if you're a if you're a steward student of 80s films like I am, and you don't recognize the name, the title of this podcast, we said we'd Anton, we'd let you talk a little bit about the AI defenders or or or attackers, right?
SPEAKER_02From your uh your RSA talk. Is that what we agreed to?
SPEAKER_03I think we just wanted to revisit this otherwise painful topic. But while I was thinking about it over the weekend, I think I actually cracked it. Because I've noticed a lot of people agonize about it. And I was one of those people. And I've heard really good takes for both sides, whatever. But I think that the critical element is when. Because when people say, like, I'm gonna be a big fan of Phil Venables, I am a big fan of Phil Danables, right? And once Phil said AI benefits defenders more because they have more data and more compute.
SPEAKER_03And I'm like, yeah, I buy that, but it's a future theoretical argument. I it is to me, it's a convincing argument. But the question I want to ask is who benefits from AI more today? And my answer is attackers. Attackers.
SPEAKER_02Okay, okay, so so just to be clear, before we get into all that, I I wanted to like I actually look up this quote because I wanted to make sure I got it right. For those of you that are not familiar with the Princess Bride, the actual quote is, and I think this is very well applied, by the way, to AI is you fell victim. Uh so this is this is Zini talking to the dread pirate Roberts, as we find out later. Is Wesley, but uh you fell victim of one of the classic blunders, the most famous of which is never get involved in a land war in Asia, but only slightly less well known is this never going against a Sicilian when death is on the line.
SPEAKER_02And I think we replace Sicilian with AI and then death with hacking. I I think we're good there, right? You just I just I've I'm learning not to bet against AI in in our current market condition, not not probably not for the reason a lot of you think. So if you think that I'm coming at this because I'm I'm coming at this from a uh the technology superior and everybody just needs to bow and and and give up to it. No, I don't think that's the case. I think certainly a massive uh a massive boon, certainly a massive support.
SPEAKER_02I just think that like between uh the real the hyped and then like the prospect of the future, the the the opportunities are too big. But now the question is is it uh benefiting our like we've done this, is it benefiting attacker or or defender? We done we've done this a couple of episodes back. I I think we came to the conclusion that one that you did is is attacker at this point.
SPEAKER_03Yeah, correct. I think that it's it's at this point, but I am really, really anchoring on at this point. This is
SPEAKER_02uh
SPEAKER_04it it falls under the same to me, it falls under the same thing of pretty much everything else, right? Like attacker versus defender, it benefits more. I agree. I think it benefits attackers more because it's easier to implement in an attack scenario than it is for people to understand how to implement it in a defense scenario.
SPEAKER_03Like
SPEAKER_04it's easy to go out there and tell it, hey, go hack this. But just like trying to patch, just like trying to everything else we've been trying to do in the defensive side, everybody said, Oh, go do it, it's great for defense. Nobody shows you how to do it. It's not as straightforward as being able to go say, here's a URL or here's a code base, go scan it and find me all the security vulnerabilities versus go fix everything, or you know, even from a I think code may be a little bit easier, but still not super simple.
SPEAKER_04But all the rest, like you're not just AIing to go out and patch all your systems and all that stuff tomorrow. It's not that simple. And I think that's where we see that difference of I think to your point, Anton, today it definitely benefits attackers more than defenders, but more so from a I don't think we know really how to use it because more people focus on attack than they do defense if you're talking about publicity standpoints.
SPEAKER_03So you're almost saying that the attack is if you say if you if you wake, I'm gonna use a 3 a.m. example from that I used on my podcast for quite a bit. If you wake me up at 3 a.m. and says, Anton, what's an attack? I would be just I would easily describe you a compromise scenario, somebody's breaking in, somebody's sending fiction emails. Would be very clear what it is. But if you say, what about defense? I'd be like AppSec, GRC, NetSec, SecOps, firewalls. There's no AI for attack, it's pretty clear. Faster compromise and better lateral, whatever.
SPEAKER_03But AI for defense is like a somewhat amorphous blurb, right? That's the that's your argument I'm hearing.
SPEAKER_04Yeah,
SPEAKER_02that that I think that's that that's the argument everywhere, though. Like, think about it's I I this is how I got into this industry. We knew that like everybody, James, we how many times have we talked about every wants to everybody wants to go be a pen tester or red teamer now because you just you know you have it? I think it's more fun. You have a a lot of opportunity to go do something cool, and even when you're wrong, you you still find something cool.
SPEAKER_04It's less accountability, right? Like you're the guy that points out the crack on the bridge, but you have no responsibility to actually fix it, fund it, or have know-how to fix it. All you have to do is say that's bad.
SPEAKER_02Yeah, but defense, I think, is hard for the the reason that it's as Anton's pointing out, is it's it's actually relatively undefined, right? How do you how do you what's the what's the goal of defense? It's to prevent successful offense. Right? And and when the goal when the methods and means of successful offense change, the definition of defense changes, at least the nature of it changes. Like how the how. What we're trying to do is prevent our company from being our company resources from being misappropriate, our company secrets from being extra extracted and stolen, and our uh resources and and people from being exploited.
SPEAKER_02How?
SPEAKER_03Yeah, how exactly. It's the you know, use AI for that. Like, well, well, the what is that? I I mean I want to know what that is that you're referring to, right? I mean, of course, a lot of people slide immediately into sock, but I feel like that's I I'm sorry, I am the SOC guy, but I can probably tell you that using AI for sock is not the most impactful use of AI for defense.
unknownOkay, maybe.
SPEAKER_04Well, it changes the it raises the bar too on the whole, like what's everybody always said, you know, we you're raising the the cost so that way they'll move on to the next piece, right? Like, I mean, it's the whole like put the security sticker on your window, and you know, the attacker will go to the next one. Like, we raise it so it costs too much for them to want to do it, and AI erases that a bit.
SPEAKER_02Yeah,
SPEAKER_04right by being able to do stuff quicker and fast. Like it it removes what was that simple barrier to I don't know, what we were using as kind of that definition.
SPEAKER_03Yes, because if you have uh if you have a way to eliminate misconfigurations on the perimeter and somehow that process involves AI, that's probably a good use case, right? Because you remove I I I have a specific scenario in mind, but I'm being kind of a little vague here. So if you if you have a way to optimize asset discovery using AI and make it kind of relentless, that would be a very impactful use case for defenders.
SPEAKER_02Yes, because only because one of the greatest challenges that any enterprise faces is knowing what assets they have.
SPEAKER_03Yeah, correct.
SPEAKER_02So like when your foundation is at best half identified, right? It's wasn't there uh wasn't this some this wasn't this some like takeaway from the uh from Troy is that you know impenetrable walls, all this stuff, great army, and then they had like they had to go build tunnels out to the ocean or sea or something that ocean to sea to get water from or whatever, whatever, and then that's how they were broken into because nobody knew that those were there, and then you know they got they got busted. So I I think I saw a story like that somewhere, but anyway, but I but that's that's the challenge for about at least if I if I had to put a a dollar figure or a or a percentage on it and put put money to it, I would say at least half of the compromises that I am aware of or have been personally a part of in some way, all of them started with somebody going, wait, that's not ours, is it?
SPEAKER_02Or or what is that
SPEAKER_03thing? Right?
SPEAKER_04It's it's also interesting, Raph, the way you brought that up to be able to say, like using AI from a defensive standpoint to identify all your assets. That's great. Take that back to the original question. Is it benefiting attackers or defenders? Attackers, that's that's most of what they're doing. They identify and do it, right? We are great. We still got to do the attack side, which is identify, but then we have to actually do the resolve, which is if you're building that wall, if you're sealing that tunnel, that's gonna take more time natively than identifying the vulnerability and exploiting it.
SPEAKER_02So, Anton, do we break this apart then into two at least two things when we talk about AI and defense? The first part is short-term tactical, you know, the parry, the dodge, the
SPEAKER_03yes, correct.
SPEAKER_02The hand comes in, we slap it, hand comes here, we slap it, right? So that's the that's the constant. There's a vulnerability that's it's open. Somebody discovers it, AI finds that they're trying to get into it, you know, takes care of that. And like the the perfect analogy I can think of is I'm watching, is I'm watching Daffy Duck try to plug the dam in as it's cracking. Like one cracks and puts his finger in, another one puts a toe in, puts the other toe in, sticks his beak in. Now he's like, Well crap, there's a new one now. What? So you gotta take away your finger from one of the cracks and put it in another. So like you're trying to fix the cracks in real time as as they're happening, or the attacks in real time as they're happening, which is your sock and your operational, operationalized defense.
SPEAKER_02But then there's the other part where I think AI might actually so so in that regard, I would call that advantage offense, but close to a draw, because if you can operate fast enough, AI can detect AI and you can at least minimize impact, okay? If we can agree on that.
SPEAKER_03Fair, yes, we can. Yes.
SPEAKER_02So I think it gets
SPEAKER_03but if you use if you're good at it. If you're if you if you're so that's the caveat,
SPEAKER_02right?
SPEAKER_03Yeah, you can get it.
SPEAKER_02Yeah, this isn't throwing like a a Claude at your sim and saying fine bad. Like that's that's not what that is.
SPEAKER_03Correct. No, no, no, not the sim. It's at the sim would be more intelligent. Pointing to draw logs. I think I think it's and not Claude, Claude is actually pretty smart. Pointing a random open weights model at your logs and saying, hey, give me, yeah, that I I've seen that. I've I've my recent blog is my lamentations about that, basically.
SPEAKER_02But I think that the the the second part of that is the non-plugging the holes in the damn part, which is how do we get is AI giving us an opportunity to potentially it's not getting ahead of, but like find the structural problems. There is getting ahead of there. No,
SPEAKER_03no, no. Don't don't undersell it. I mean, getting ahead is there. I mean, I know you and many other people are very much allergic to use AI to write bug free code, use AI to write vulnerability-free code. And uh and I've met people in my previous job who very much have ground, solid grounds for for actually succeeding at that. And uh the problem is that people on the outside usually point at their shitty code and say, Oh, this can never be written by AI. But people on the other hand look at their code, which is already well engineered, well architected, and say, look, we can use AI to make it essentially vulnerability-free.
SPEAKER_03And it is preempting the attacker. And if you have unlimited budgets in dollars and tokens, you can do this. At least that's my belief.
SPEAKER_02Okay, but hold on. Can we can we can we caveat this to reality then to the world that we live in, which is not unlimited?
SPEAKER_03But it is reality in some companies.
SPEAKER_02Really?
SPEAKER_03Yeah.
SPEAKER_02You've met people that have unlimited token budgets?
SPEAKER_03Yeah, I mean, look, last last the this is. Can you name the number of work
SPEAKER_02there?
SPEAKER_03Yeah, yeah, I can. I can. I mean, at the first unprompted event, there was a talk by one of the labs. I don't want to misattribute, it was either OpenAI or entropic. I'm thinking maybe OpenAI. And they said, we did this, this, this. Because tokens are essentially free, we can then do this, this, this, this. So they think of tokens as free. And that that that point triggered some people in the audience. But I I it's very clear to me that they the speakers meant well. They just described their reality.
SPEAKER_03And I was like, wait, hold on. Tokens are free? That's news to me. But it was just a side point in their presentation that uh it supported some arguments that they made. And
SPEAKER_02I mean it
SPEAKER_03I mean,
SPEAKER_02okay, but if you if you work at a look, even if you've got Anton, even if you've got an open weight model running in your own environment, I've seen the virtually free AI thing, which is bullshit because
SPEAKER_03you still need
SPEAKER_02the hardware, like you still need the support structure, like somebody still has to pay for you know the GPUs and the power and all that. Like, so it's free. It's just the cost has shifted from affordable to you to somebody else.
SPEAKER_03Okay, yeah, yeah. Affordable to you.
SPEAKER_02Maybe
SPEAKER_03that's
SPEAKER_02more so okay, fine. So let's let's take let me let me join your clan. Let me join the we can build, we can build error-free code clan. If I joined No, no, no,
SPEAKER_03no, no, no, vulnerability-free. Because vulnerability free people won't attack us hard. Yeah. Somebody I've made I I've quoted somebody from my previous job at at LinkedIn, and people kind of attacked not just the vulnerability free claim because I said you can't fix all the bugs. You don't even know what the bugs are. And it's like, no, no, no. It's the vulnerabilities, the classic known vowel types. It's not about fixing unknown vound types, it's not about not creating new vowel types, it's about code free of known vowel classes.
SPEAKER_02Okay.
SPEAKER_03I'm not even an abstract question. Why are you torturing me? Let's talk about soft.
SPEAKER_02I'm not. I'm just I'm just trying to look, I'm trying to understand the position because it's like this matter, semantics matter here because it's talked about vulnerability-free code. But like we I'm gonna make a stretch here, ready, James? Your head's gonna pop off your shoulders in a second. Like, we had the opportunity when we had reasonably good static and dynamic tools that we can say this code is vulnerability free according to this testing methodology, this tool, and as of right now, you put it in production and go, well, crap, that was 15 minutes ago.
SPEAKER_02Now it's vulnerable to this thing that just got released. Something we just discovered, right? So if the caveat is known vulnerabilities as of whatever time that is released, okay. I I think I think I can go for that ride. I'll buy the ticket, I'll get on the bus. Where is it taking me?
SPEAKER_03But vound classes aren't discovered every day, you know. The vounds are, but vound classes uh is is kind of a different story, right? And I think that again, I don't want to be the evangelist for my previous employer more than necessary. I still like Google, obviously. But the the point is that the eliminating whole classes of vulnerabilities was kind of a slogan at Google Security for decades at this point. And there were examples where no SSS, no XSS was found, no other classes, they were just gone, gone forever, gone and forgotten.
SPEAKER_03So that spark is real. And these using free AI tools.
SPEAKER_04I was gonna say that's not even that doesn't even require AI to do that. Yeah,
SPEAKER_03exactly. It was done before AI, yeah. Right. I think the
SPEAKER_04big difference though is that when you're sitting there talking whether you got unlimited tokens or whatever, it's still, I think back to your point, Anton, about well-engineered applications. And I think that's the limitation that people have. It's not whether they're using AI, it's that they're not their stuff's not well engineered, they don't know what well-engineered looks like.
SPEAKER_03No, that's zero chance.
SPEAKER_04They're used to putting together what we currently have. I mean, this is the same reason why we still have SQL injection and some of these other classes of vulnerabilities, because while some companies can go out there and completely eradicate it, you're still gonna have a whole swath of people that have no idea or are still building things based on the code examples they see in the book that was with you know written 10 years ago that are full of inline SQL and and bad examples.
SPEAKER_02Okay.
SPEAKER_03Yes, true, fair, yeah. It's all it's all it all exists at the very tippy tip of the pyramid with the ultra elite plus type of teams, budgets, and resources for sure.
SPEAKER_02So, Anton, if we take let's let let's be less academic about this then.
SPEAKER_03Okay, yes because
SPEAKER_02you just you just the way you just couch this is
SPEAKER_03that's true.
SPEAKER_02So let's be less academic. Let's talk about the reality of where let's you know hit the hit the 75%, right? The mid the middle the upper middle band of enterprises that are reasonably well funded, decent leadership, reasonably mature. What does AI what does AI get them? Not on the operational poking poking fingers into the dam, not fixing leaks,
SPEAKER_03okay,
SPEAKER_02but but broadly.
SPEAKER_03Okay, so this is uh I'm gonna hold, make a quick pause here and say that that question is really the question, the defining question of our time. What is the absolutely credible AI use case that dramatically helps defenders in that bucket that they can implement today? I I I wanna I want an answer. I do not have it. I'm sorry. I I my bias would lead me to SOC, but I also I seek a lot of counterarguments to that. So, how about we discuss it? What is the number one AI for defenders use case to practical today that does have an impact?
SPEAKER_02I I think if you take the uh the operational counterpunch piece out of it that takes sock out. So the non-SOC use case, where is the most practical application?
SPEAKER_03Yeah, if it concludes a sock, let's just say it's a sock. Yeah, it's fine.
SPEAKER_02But I I but I I think I think sock, I think if you if you bring security operations into it, I think it at best levels your playing field, right? Because that's the jo that's the goal of that department. It's not to be cyber defense, whatever we call it, like I think we've settled on the fact that it's a it's a defense, it's a countermeasure, it's a group of countermeasures. They're not the SOC doesn't look proactively at it, looks at what's happening now and how do I avoid it becoming bad, right? So if that's the case, then that's the wrong place to look to be proactive.
SPEAKER_02So if we're gonna be proactive, you've got policy, you've got compliance, you've got arguably software security, you've got uh don't throw anything at me, but I'm gonna say it, patching and vulnerability management.
SPEAKER_03Yeah, well management, yeah, for sure. It's gonna be on the list.
SPEAKER_02And but that's predicated on an a non-security thing, which is kind of the ITSM of asset management. And I would argue that the probably the greatest place that AI, so this is my position. AI, I think, will probably make the greatest impact to non-reactive cybersecurity in vulnerability management. Bugs in other people's code. That's not yours.
SPEAKER_03Not yours. Yeah, it's it's again. Yeah, yeah, I exactly. But I mean, if you're again in the and we are doing the upper middle class enterprises, we are not doing the quote unquote out of reach, ITRA reach, because you uh people can't point it at random up third party software and say And they have done it and it worked well.
SPEAKER_02Sure. But but that's that's not the reality for most.
SPEAKER_04What's the not reactive part of that though? Because I mean you're still reacting to what they're sending you. Like I when you say non-reactive, I'm thinking of something like okay, I'm gonna build I'm gonna design my systems in a way that are going to be protected by default. How am I using it? Yeah, so let me
SPEAKER_02let me let me be more clear. When I say operational, I specifically mean the functions of the SOC, which is like I'm being attacked. Yeah, react to an immediate
SPEAKER_03attack. React to
SPEAKER_02it in real time. So vulnerability management isn't that.
SPEAKER_03Yeah, yeah.
SPEAKER_02Vulnerability management, if it's if it's done in a and this may sound completely insane the way I'm saying about to say this, but if you're if it's done in a reasonable manner, I think in a when I say correct manner, it is actually proactive. It it's being proactive to something that you have to be react, you have to react, but it's not like you're you're patching it as somebody's attacking it. You're seeing a a shortcoming of vulnerability in a piece of software you own before it gets attacked.
SPEAKER_02You're thinking, do I need to patch that or not? Or do I need to patch this one, or that one, or that one, or that one, or this 7,000? Where in my enterprise do I make the most value of my team's time? And where is the growth? So this is the other part that I think I think AI is going to give us an advantage. In you guys remember like the one of the biggest detractors of corporate patching? I do, is patching something and praying to God it comes back.
SPEAKER_03Right?
SPEAKER_02It comes back up. I've been there, I've I've patched systems that you're like, this is simple, known side effects, none. You do it, you're like, okay, we're done, time to go home. And somebody says, uh, nothing works. You're like, what? But the system's up, the patch is applied, everything's great. They're like, nope, nothing works. What do you mean, nothing works? Like, I don't understand. How could this be? Right. So I think AI can help us with that. The predictive disaster that sometimes because you can it can look at a patch and say, This is what your system does, this is what will be closed off.
SPEAKER_02If you do this, this functionality goes away. It helps take the risk out of patching in a system that is not yours. Does that make sense, guys?
SPEAKER_03Yeah, yeah, it does. I mean, I think that to me, the intuitive definition of reactive versus proactive made sense, even though James is right. It's everything is it's sort of a response to an attacker, but response to an individual attacker right now versus response to an attacker having new capabilities is different. But yes, I want to use AI to design something really well and implement it for sure. Who would say no to that?
SPEAKER_04I mean, I I'm just thinking of it more so as I don't want to build an AI system that just is good at patching my stuff automatically for me. Like I want to use AI.
SPEAKER_03Oh, look, look, look, let's AA powered PHP patching. Let's sell this as a startup idea.
SPEAKER_04But I I'd rather have it help me design out an actual environment that patching happen, like there's testing for patching, there's all this stuff that doesn't matter what comes down through, the environment is self-healing to that. Whether that's it can spin up an environment, do the patching, test to verify that everything still works, that these patches are safe to deploy, and spits that back in. Like something to provide a solution, not just something that says, oh, well, I see PHP to shell and oh, well, this one's external versus internal.
SPEAKER_04We better go patch this external one first. Like, great. We sped up our typical patching process that we have, but it didn't really solve a patching problem.
SPEAKER_03Correct. Moreover, you haven't solved the security problem that underlies the patching problem. Maybe patching problem wouldn't exist if your infra is immutable and everything is pipelined and magical and unicorns, then maybe you won't have a patch problem. Okay,
SPEAKER_02but we're back to but we're back to the Okay, you're right.
SPEAKER_03Friend, good.
SPEAKER_04But you're closer to being able to design something like that, I think, with the use of AI, okay.
SPEAKER_02To
SPEAKER_04be able to say, hey, how can we design something out that would handle all of this versus taking three random people, throwing them a room and saying, hey, solve this, because we've had how many people in security for how long and we've never come to a solution. So you know, do you sit there and take the focus and put it there and say, I've only got this limited number of tokens? I want my tokens to go to this, not just so just automating the process we've been doing manually for the past 20 years.
SPEAKER_02Huh.
SPEAKER_04Maybe we need token sharing, you know, like all these like ISOCA groups and all that stuff. Like, let's get a group of people together and share their AI tokens so they can actually start solving some of these type of things. All
SPEAKER_02right, Mr. Kumbaya. Yeah. Anton, you wanted you said you wanted it, you wanted an answer to where we think that AI can best provide value in a non-reactive, non-real-time reactive aka stock capability. Would you agree that do you think I'm in the going in the right? Would you take would you take the ride with me if I said it was vulnerability management and PAP strategy?
SPEAKER_03Yes, but I would not know for sure is it because it's my bias, because my second area after the after sim where I spend most of the time is VM. So I I naturally gravitate to to this. So I'm I'm I I want to agree with you, but I I want to then second guess myself because maybe it is appsec. Maybe it is the fact that you know AI-powered SaaS beats regular SaaS by a factor of 10 or whatever the data was. So I don't know. I I I I agree with your VM take, but it may be very Anthony argument, not like average enterprise should do it.
SPEAKER_04I I feel like when you bring up AppSec, Anton, it's more of a a wider solution, right? Like VM, you're basically dealing with your own stuff within your own organization. But
SPEAKER_03if
SPEAKER_04we actually applied it to AppSec across the board, then there go half of your CVEs and most of your issues because most of the issues are in the security, you know, the security flaws of the applications you're using. If I don't have to patch anymore because we're actually writing vulnerability-free code.
SPEAKER_03Oh, okay. Yes. Then all that stuff,
SPEAKER_04like dominoes right out of the way. So while VM at it at your own organization is a bigger focus, at a broader scale, AppSec would have a wider effect on it, I think.
SPEAKER_03Yes, but it is also purely local AppSec, right? There's just the better SAST that helps you then create a patch to your own code. That to me is very legit, and I see a lot of people actually kind of succeed down that path in a very tactical manner. Uh it's just the SAST soft static analysis plus some kind of patch writer thing, right? Like that's still helpful, right?
SPEAKER_04Yeah. No, it is. I mean, it is, right? Because again, I mean, if if we're solving the software,
SPEAKER_03yeah,
SPEAKER_04then we're solving those downstream issues that are coming from the software.
SPEAKER_03Okay. This may James' point made me want to disagree with Raf about VM being the second after SOC. And here's why.
SPEAKER_02Okay.
SPEAKER_03I actually when you said likely it's vulnerability management, what specific use case of AI did you mean, Raf?
SPEAKER_02What do you mean?
SPEAKER_03To do what in VM? Look, here's the your unlimited talking vice VM. Like where?
SPEAKER_02Gotcha. I I think I think the application, so I think there's two places. One to help us understand risk better. Because which lead so understanding risk of if I apply this, what what is the risk of applying it? What is the risk of not applying it? Right? Like to help me understand the collective risk of doing it of an action or lack of. And then prioritization is the other piece of that. So I think it's a 50-50 split between the two.
SPEAKER_02So I think I I think if if I'm reading you know the the VM industry seems to be reinventing itself, which is cute. I I think that that you know that's like 20 years overdue, but that's just me. So if I'm if I'm if I'm following R Snake and Jeremiah and those guys correctly, then what we're doing is creating potentially a more intelligent model built around finance and actual true risk, rather than ah, everything's critical or everything's critical or whatever, right?
SPEAKER_02So more reality, less less, less vendor hype, and less in you know, formulas written by somebody who wanted to make sure the numbers distributed out evenly. So rationality to prioritization.
SPEAKER_03Okay, so I two-thirds disagree with that. And and but but but very gently. Uh I gently disagree with that. No, really, I sometimes I violently disagree here. I disagree gently. So the reason for my gentle disagreement is that I think we sort of know what to do in prioritization, we sort of knew what to do. And we but the reason why I'm gently disagreeing is because if you prioritize better with AI and you have 5,000 valves and you have magic AI engine that tells you out of this 5,000, 4,000 are supercritical, but your bandwidth is still the fixed 1,000.
SPEAKER_03AI has absolutely zero impact because your must-fix set is still bigger than your bandwidth. And AI thus missfires, right? It's
SPEAKER_02a good thing.
SPEAKER_03Okay, but the
SPEAKER_02but the whole yeah, okay, so hold on, but but the the promise and premise of this well call all right, let's call it magic AI, is that it's not 5,000 out of 4,000, it's like four or five.
SPEAKER_03No, no, no, but it's possible that 5,000 really are the critical ones. I've seen exactly that situation when the when they run the magic pre-AI algorithm and it tells them these 323 will kill you. Your business would be kapoot if they're exploited. And they say, sorry, we only have time for 100. Like that's a clash. And that prioritization, you can't prioritize yourself out of no having not having bandwidth. That's why I am not skeptical. I am mildly disagreeing because I I want better prior, but I don't think prior would help it.
SPEAKER_02Okay. So what's the other two-thirds?
SPEAKER_03The other two-thirds is yes, I want better prioritization. I I mean I really think that prioritization is helpful. And if it's more aware of your environment and likely it would give you a smaller set, but the set may still be large enough because of because of an apocalypse. Yes, that's my fear. That's why I feel like it's handy. Uh what about why not why why none of us none of us like I am, right? You like I am? You like I am?
SPEAKER_02I mean, look
SPEAKER_03at it.
SPEAKER_02I
SPEAKER_03don't like I am.
SPEAKER_02Listen, I think identity,
SPEAKER_04I hate it. It's not that I don't like it. I hate it.
SPEAKER_02I think identity is you know, people said identity is the next security frontier, and I and I completely agree. But what that means to me is it's the next thing we completely won't understand and how to do anything with for the next 20 years.
SPEAKER_03Yeah, yeah, but then maybe that's an AI, maybe that is a applic uh lever to apply AI.
SPEAKER_04But the quite the problem is that we gotta have people that actually understand it to apply it properly. And now, I mean, identity, identity's going crazy now with like agent identity. I mean, look, I mean, you got human identities, agent identities, service account identity, right? Like they've got all these different things, and it's all running super fast,
SPEAKER_03and
SPEAKER_04it nothing to really rationalize where everything's supposed to be. I mean, I don't know, maybe using AI early on with it will be able to help solve that better than us coming in late at the end of like VM. I mean, we already have posture management for every known crevice of security to be able to help with prioritization and and all that stuff, right? I mean, we we've got ways to do that, and of course, with AI, sure it can do faster, but if it has the right context, right?
SPEAKER_04There's a lot of ifs around AI being actually accurate and telling you this is the one you should focus on first, because if it doesn't have the right context, it could still send you down the wrong path and you're fixing the wrong things.
SPEAKER_03Yes. Actually, we are sort of, I guess I I it's not again, it's it's I am I'm the temporary host slash guest here, but what kind of conclusion can we give people at the end? Like defenders, here's where you apply AI and how. I almost want to arrive there in the next what seven minutes we have, right?
SPEAKER_02I I think so I would love to I would love to arrive at that, but I think that requires lots of trial and error.
SPEAKER_03Yes.
SPEAKER_02I think we need to I I think the vault market's gonna vet itself out over the next 12 months, which of course means 12 more months of utter sheer and shitty chaos.
SPEAKER_03Yep.
SPEAKER_02I think the IAM, the idea, I think the identity space, broadly speaking, whether it's malicious malice detection, whether it's you know accurate and provisioning and deprovisioning, whether it's cross-domain identity management, all that stuff, human, non-human identity, all that, right? I I think this space will mature and consolidate over the next five years. Without AI, I think AI can help it down to like a year, year and a half.
SPEAKER_02But I think everybody needs to get in on it. I don't think everybody will.
SPEAKER_04What's AI doing though? And that's that my question, because when you talk about identity, there's already so many controls in place. It's the fact that people don't implement them. I mean, you look at like authentication now. I mean, there's so many checks that go into when you authenticate into a good system that's saying, hey, are you coming from a same place that you've been before? Are you on the same machine you've been before? Right? Like all these things are being checked already. It's a matter of whether places are actually doing it or not.
SPEAKER_02Yeah, but that that's not, I think that now that we've uh addressed that problem, attackers are moving into the Yes, all those things are true and match your previous identity, but I've now I hijacked the AI bot on your machine to go do my evil nefarious work.
SPEAKER_03Yeah, that's fair. I like that. And maybe you haven't touched AI for security of AI because businesses are rapidly adopting AI, as we know. So AI for that is still a big bucket of fun, big ball of fun.
SPEAKER_02I don't know. We're uh like that. I think I would I would look at the application of AI tools. I realistically, I'm still putting my money in into the sock because that's the spot where you've got to you the the investment in humans is not keeping up with the threat, and the threat is real and clear and present, like it's there right now in your laws. I
SPEAKER_03I'll I'll uh it's the easiest thing in the world to convince me to agree with you completely, unquestionably, and basically say, look, if you wake me up at 3 a.m., where do I apply? I I know nothing about your company, about your enterprise. You have a sock, yes, in a sock. Sure. Sold.
SPEAKER_02Okay.
SPEAKER_03There are 123 startups doing this, right? Like there are 123, whatever the number is, according to LinkedIn influencers, but it's like in large double digits, if not low triple digits, right?
SPEAKER_02Yeah, I listen, I I saw the number somewhere around a thousand, but I I think that was the conversation we were a part of. But I think that that market can I want to make a bold prediction. I might be wrong, but I'm I'm prone to making these big loud predictions. I think the AI stock market collapses in on itself.
SPEAKER_03Define collapse. I don't think that disappears
SPEAKER_02I think it disappears. I think be here a couple of reasons. One, I think the investment in it has not vetted or has not netted the the the benefits because I I I look around and and uh haphazardly my day job. So full disclaimer, you guys know my my my day job, we we've gotten lumped into that accidentally and out and and without without wanting to. But so now I'm kind of like stuck looking at this space. And I look at and I see customers that are don't quite understand what they're buying, but are disappointed in what they've what they're trying, right?
SPEAKER_02They're they're like they they're not quite sure what the outcome should be, but they know that it's not it's not the one they're getting. So it's like it's it's like you know, taking your your children to to to a nice restaurant when they're young, and you go, What do you want? And you go, I don't know, but it's not this. You're like, well, that's not an answer. Like, I want something nice. You're like, great. What like here's your menu of something nice. I go, none of those things. Like, well, then what the hell? Right. So I think AI SOC has underdelivered, taken a ton of finance. And I think what you're seeing now, and where I I'm gonna have to go back and apologize to the guy on LinkedIn's I call him crazy, but I think I'm he may be actually correct.
SPEAKER_02And that is, I think you're gonna see the big, big companies, the Microsoft's, the Palos, the Googles, basically come back and say that's adorable that they tried that. Look, I can I can give you that for 10% uplift in my current platform, and it's darn good enough. It's not that it's that it's not that startup, but and I'm not I don't even think they have to buy these startups. I think so I think that market collapses. And I think the reason is because what we're talking about right now is how do you define value for the dollar invested in AI in security?
SPEAKER_02And if the answer is the operational day-to-day socks stuff, then you go, okay, what's the return I'm expecting to get? And it's not that the technology isn't marching forward at blistering paces. I don't think the buyer knows what they want.
SPEAKER_04Well, I think there's a there's there's also a little bit about there's got to be a question going on of why did I need AI to do what just gave me when we could have done that with simple automation? You know, I mean I think there's still a an imbalance going on there where we're using AI to do stuff that we could have done with scripts for years, right? That costs a whole bunch of money. Like, are we applying it correctly? But I think to your point, Raf, right? Like they've oversold what AI in the SOC is gonna do, and everybody gets it, and then they get it, and they're like, well, that was cute and all, but that's I mean, I've seen apps like somebody we were looking at an app, I mean, completely like, oh man, this is gonna be fantastic.
SPEAKER_04I'm like, I'm like, I mean, I could just go write the same query into Chat GPT, get the same result, right? Like, I don't need a special AI tool to do this. If I know the other information already, I can write my own queries into this stuff.
SPEAKER_03Right, but that's where people who just did the chatbot duct tape to a sim as a separate venture would would probably die first. I mean
SPEAKER_02self, yes.
SPEAKER_03Yeah, but it's it's like the the thing, the thought that sits in my mind is that imagine how you feel if you're a vendor number 97 by market share in AI stock. But but the biggest vendor maybe makes 10 million dollars, but you're number 97. You're not number one, two, three, four, five, six, seven.
SPEAKER_02You
SPEAKER_03are 97. So you're where are you in revenue? I don't want to even guess. You're probably, I don't know, but uh but the point is how does it feel to be a vendor number 97?
SPEAKER_04Just depends on what
SPEAKER_02we vendors making. It just depends on what seed
SPEAKER_04round you're in and where you are revenue, Anton. Doesn't matter. Like as long as you can get to like seed G or something like that, like screw it. You had a good year's run.
SPEAKER_03You're too cynical. You're too cynical.
SPEAKER_02Congratulations, James. You are now me. No, but that but but while too cynical, I don't think he's wrong.
SPEAKER_03I I I mean, again, I I am maybe I I I hate to be the guy who is biased in favor of large companies, but because I was always kind of like allergic to working at them. But as of today, again, I just see people who say, I'm just gonna wait for my preferred sim vendor to build it. I have no issue whatsoever waiting for two years. There's no pressure for me to adopt AI SOC today. Some said, I'm just gonna wait for Cisco, Palo, CrowdStrike, Microsoft, whoever. I'm just gonna go through a list, include Cisco so that my bosses here don't beat me up.
SPEAKER_03And sure, it's it's a perfectly legit legitimate tactic. And I don't know where 97 AI SOC vendors fit into this picture, except to acquisition targets for people who are laggards, for large vendors who are laggards.
SPEAKER_02I I think there's always gonna be those buyers that look to the startups first because they believe that that's where the innovation lives. And it's a hit and miss. Sometimes the answer is yes, and sometimes the answer is yes for a very short period of time, and sometimes the answer is no. I think we're somewhere between yes for a short period of time and no uh at the moment. And then the question becomes you know, is is that a good place to throw investment in right now? Does it get you the kind of value you want?
SPEAKER_02And if you define value, then and then how do you define value for for whatever it is that you're doing, right? So there's a there's a lot of precursors to this, but we're like way over time, so I and we still don't have an answer. Anton, what would you give the answer?
SPEAKER_03Look, I I honestly think it is in detection response, quote unquote, in this talk. I mean, I don't, but but I again I I I I always point out my bias. I'm not an EpSec guy. I don't I play one on TV apparently now in this episode, but but it is probably in DNR, and that that's why I almost feel compassion for the 100 startups because they see that and say, look, where do we apply AI without reengineering the enterprise? Well, that's probably in DNR in SOC.
SPEAKER_03Yeah, that's my answer too. Sorry, I'm boring. Despite my nirvana t shirt.
SPEAKER_02Is that an ironic nirvana t shirt? Are you actually a nirvana t shirt? That
SPEAKER_03is absolutely an ironic nirvana t shirt, yes. But it also reminds the connection to a nineties. In some connections.
SPEAKER_02I watch, I watch my kids. Friends wearing like Megadeth and Guns N' Roses, and I'm like, your parents aren't old enough to remember that band. I know you're not.
SPEAKER_04It's funny. My son wears all like the Metallica and all, but you can tell from behind me.
SPEAKER_03Yes, that they're he's into
SPEAKER_04that. So
SPEAKER_02my my my my son is into the talking heads, which I think is hilarious because like I don't know where he got that. I like him, but it's not something anyway. All right, so we we solved we solved the security AI question. Great. On the next episode, we'll solve some other life problems. I don't think we have. I I think I think I think much like everything else, the answer is more research and time needed. This will vet itself in one way or the other. I think we can anticipate that it's going to be rough between here and there.
SPEAKER_02Lots of people will spend money and see no value. Lots of people will wait and then something will happen, and then they'll say, Oh, if I would have spent money, I would have gotten it. I would have gotten. No, you wouldn't have. Stop fooling yourself. But I thought I think you know AI for sure provide provides value into the stock. I think we all agree on that. Beyond that, it's probably gonna be a coin toss, but it'll either be identity, application security, or vulnerability management, I think is where we're guessing.
SPEAKER_04Yeah, I'd be surprised if you I mean, with if you're not looking at application security for it, again, that's the global the solution, right? Like that's where you're gonna get the biggest bang for your buck if you can start getting those issues out of there, because there goes your CVEs, there goes all the stuff that you're patching. Like that is the root of a lot of this stuff, and we see it being solved by the I mean, Anton, I think to your point, right? Like, I mean, there are companies that are getting rid of full classes. I mean, you start switching to React as front ends, you start getting rid of cross-site scripting, you get rid of full classes of vulnerabilities.
SPEAKER_04If we can get that even further to where we're not seeing these, but I mean, look, we're half of them, right? They're still in the old apps that are still running like C, or they're in your browsers, they're in your Windows, they're in a lot of your other main providers aren't getting nailed with these type of vulnerabilities.
SPEAKER_02I think it's gonna be interesting.
SPEAKER_03It would be very fun to watch, yes. But I think that we all of us probably need to have in the back of our minds this really way to share the AI successes with defenders. I'm sorry if I sound like a cliche talking head on LinkedIn, but I really want the sharing of here's where I applied Gen AI, AI agents for my security, and this is what worked. I see some companies do it, it's great. Let's do more of that.
SPEAKER_04That's what I was saying at the beginning, Anton. Like when you talk about an attacker perspective, attack owns the media. They own everything. Like people want to go out, they do research on how to do attack, how I can use AI to break things, how I can get into this. Rarely do we ever see the people, even the groups that get together to be the defenders, they come up with some pledge that says, hey, you should use this, but they don't have nearly the details of like, here's how you could use this to do it, right? Like you can find plenty of details of exactly how to use AI to hack things.
SPEAKER_04But to actually find good write-ups and details of how to use it for defense, it that's what's missing. And we got to figure out how we get that. It's like getting open source developers, right, to want to fix their vulnerabilities. Like, look, I'm already doing this for free. Why am I gonna it? It's it's just a hard sell to get, but there's there's no glory in defense.
SPEAKER_03Uh yeah, you the maybe less media attention, but yes, uh, I think there's glory in defense, but uh yeah.
SPEAKER_02All right. I
SPEAKER_03mean, we're
SPEAKER_02all
SPEAKER_03doing defense. We must think there's glory in defense. It's not sell
SPEAKER_04it's not your marketing though for your startup that you're trying to sell business to everybody. Nobody's got the defense company that's actually putting this stuff out there saying, look how great we are, come use our stuff. It's look how great we did breaking this stuff. That I think is the difference, right?
SPEAKER_02Well, I think that's a good place to end. About, oh, I don't know, 15 minutes over time. Obviously, this is an easy topic to discuss, but I I'm curious on two things if you're listening to this, and if you've listened this far, uh the first one is uh ironic 80s or 90s t-shirts, in or out. Good, you know, like we gotta decide whether it's it's okay to wear a music uh t-shirt, a band t-shirt that uh you probably didn't actually like in the 90s, ironically. And and the question two is where is the best application of AI into non-operational, non, you know, non non spur of the moment defense.
SPEAKER_02Right? So if you take it, if you say not in the sim, not in the horrific bloodbath that is AI sock, where
SPEAKER_03love it. And we'll
SPEAKER_02leave it at that.
SPEAKER_03Yep.
SPEAKER_02Let's put it, we'll put it on LinkedIn, let's see what you guys think. If anybody has any any any thoughts or comments, because you guys are a quiet bunch when you listen. But anyway, hey, Anton, thanks for joining us, man. This has been fun.
SPEAKER_03Perfect. Yeah, let's do one more.
SPEAKER_02Uh well, okay. I'll twist my arm. I don't know what the next topic's gonna be, but we'll do we'll do another one. As long as you guys keep listening, we'll keep adding Anton in. We'll see what happens.
SPEAKER_03We should do one on agents and how they can be secured. This would be like painfully fun. So
SPEAKER_02these are two words I I I don't want to put together, painfully and fun. Because I've I've done things that are painfully fun. I don't I don't I don't want to do that. But agents, you're right. I'll you know what, just to be ironic, I'll have Muse write up a topic summary for that one.
SPEAKER_03Yeah, totally.
SPEAKER_02And email us all. All right, anyway, thanks for listening, you guys. This has been another episode of the Down the Security Rabbit Hole podcast with our special guest host, permanent, semi-permanent guest. I don't know. We'll see what we're
SPEAKER_03doing.
SPEAKER_02Rabbits.
SPEAKER_03Yeah,
SPEAKER_02yeah. We'll uh you next time you gotta show up wearing rabbit a year, so you can be one of us. And that is Anton. Yeah, see, there you go. And James, thanks. It's been great. Good having you aboard, great insights, and guys, we will catch you another time, another place on the next Dominic Security Rabbit Hole Podcast. We'll see you later. Bye-bye.
SPEAKER_05This is Bella. Thanks for listening. Don't forget to leave my Data Review and share this with your friends. Bye.
TL;DR: AI doesn't care about your firewall policy. We dig into the uncomfortable question many security teams are quietly asking: Are firewalls still relevant in the age of AI, agent swarms, and automation? Guest: Anton...
TL;DR: Stop vibe-coding your detections! This week's pod features Matt Coons at GitLab as we talk about detection engineering, but with AI agents. Guest Matt Coons Description Your detection backlog is endless, your...
TL;DR: Xavier Ashe was there when that first DDoS attack by MafiaBoy went down. He's got stories, and we're going to tell them. Recorded live at BSides Atlanta in 2026. Guest Xavier Ashe Description: A DDoS attack used...
TL;DR: This week's pod is a conversation about AI and "Open-Weight Models", and if you aren't intimately familiar with what these are this is an episode for you. Wes has a lot of insight, and this is a discussion rather...
TL;DR: Phile has been penetration testing for a long time. The difference now is that he's having a much easier time with it thanks to AI. It's an interesting world, and he's here to explain the offensive side of it....