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 Chuvakin Video: https://youtu.be/SkeVmc6NVuc Description: Your firewall policy looks solid right up until an AI agent decides it’s negotiable. 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 that can probe systems faster than any human ever could? Our take is blunt and practical. If there’s no access, there’s no access, but almost nobody manages segmentation and firewall rules as cleanly as they think they do [or should]. We break down what people are starting to call an “AI firewall" (I feel like I should apologize for that), and why that label can mean two totally different things: a firewall that controls AI traffic (prompts, tool calls, model access, data exfil paths) and a firewall that uses AI to manage policy (untangling years of exceptions, spotting risky openings, and making rulebases make sense again). From there, we spar over hardware versus virtual controls, the appeal of a real kill switch, and why “AI changes everything” and “AI changes nothing” are both lazy answers. The second half gets grounded in day-to-day operations: local inference versus managed models, the real economics of tokens versus hardware, and where AI actually helps today. We’re bullish on using AI in the SOC to eliminate manual triage, ticket chasing, and context gathering, but we’re skeptical it can replace human judgment. Models can be technically correct and operationally wrong, and only experienced analysts can supply the feedback and intuition that tighten the loop. If you’re building an AI security strategy, hardening network controls, or wondering how agentic AI changes governance, this conversation is for you. Subscribe, share it with a teammate, and leave a review with your spiciest take: do you trust “virtual” guardrails, or do you want something you can power off? 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_02They say they say we should have known better than to also so detailed. deep down into this rabbit hole we found And I was thinking I didn't think of my newest reaching security featuring Mr. King Security. Rabbit Hole podcast. Plug in, listen up, and enjoy.
SPEAKER_05Um the situation. Don't let the fact seemed the last time that we were
SPEAKER_04Welcome to the Down the Security Rabbit Hole podcast, everybody. Welcome to your favorite cybersecurity podcast now with 100% more Anton Chiracan, who is going to be uh bi popular poll. Uh, if you didn't participate on LinkedIn, I thought it would be kind of funny and you guys you guys responded well. I love it. Thank you. Even me. Anton's gonna anyway. He voted for himself. That's what I'm saying. Anton voted for himself. All right. Antonio.
SPEAKER_00Was that one vote that did it? Yeah, it's the one that did it. There was no other vote. One vote, Anton. It was like, it's good.
SPEAKER_04Yeah, we're good. Welcome, welcome to the podcast, Anton. Good to have you on, buddy.
SPEAKER_03Perfect. Yeah, thanks for trusting me with your microphone. Well, with your access and all that. So I I promise not to disappoint.
SPEAKER_04Well, you've never disappointed yet, but the day, but as they say, the day is young. James, welcome back. How are you doing, bud?
SPEAKER_00Doing well. Glad to be back. I feel like we're missing, man. Lots of stuff going on, I think, in person or something.
SPEAKER_04Yeah, yeah. I I ended up for those of you guys following along. Uh, we did five episodes. Uh I did five episodes from B Sides Atlanta here. It was a really a really good series. I I really uh enjoyed getting together. Actually, having a conversation with somebody in person is completely different. And then I don't have to be on camera. But anyway, this pod is appropriately titled AI Laughs at Your Firewall Policy. And I I wanted I wanted to go into a little bit of AI and a bit tongue-in-cheek on firewalls, because you know, firewalls.
SPEAKER_04But but as as as fate would have it, this morning, my newsfeed aggregator via my AI assistant uh surfaces things that people talk about, like things that are relevant topics. Number one thing that it brought to me was a conversation somebody had around is a is AI relevant? What is firewalls relevant? Are firewalls well how many? I've had 32 ounces of coffee. I can't believe I'm still having trouble talking. Are firewalls still relevant in the age of AI?
SPEAKER_04And I thought that's a good starting point. Anti, I'm gonna ask you, are firewalls still relevant in the age of AI?
SPEAKER_03So I promised I didn't plant it, but literally yesterday to the day we recorded, I finally published a blog at my day job, which is Cisco now, about that very topic. And I was a little anxious about this. And I was anxious because I I was afraid people would have the same reaction, is that they would say, Anton, fireballs, what year is this? But I I have two things about it, two things to say about this. And one is that in many regards it they do matter because ultimately, if there's no access, there's no access.
SPEAKER_03You can't. I mean, there's uh there's agents find agents would find a way group. I have a sticker from Unprompted that says agents would find a way. But if there's no way, probably agents would not find a way. I don't buy this whole measuring the temperature keyboard noises for for side channels. Well, I do buy them as academic publication, but not as real threats. So if I firewall things well, if I segment, sorry, but like there's no access, you can't get in, or you'll be it becomes very difficult to get in. I'll put it this way. Of course, it relies on you doing it well, and very few people do.
SPEAKER_03So I'm gonna leave it at that.
SPEAKER_00Like that's basically the case, right? I mean, if you shut down a whole bunch of stuff, you've shut down a whole bunch of stuff, you're still limited to what you left open. Right, right? So I mean, firewall still matters as far as what it's actually closing. The question then is, is what did you leave open that still has issues going on there?
SPEAKER_03Yes. That's not there's no the answer in security. We probably know that. There's no the answer for AppSec or for NetSec or for any other sec. There's answers, but not the answer.
SPEAKER_04Anton, I I'm curious, has anybody in in your research, because this just hit me a little while ago, has anybody actually decided they're selling the AI firewall yet?
SPEAKER_03Actually, AI firewall is two things. Okay. AI firewall, some of the people who do gateways for AI traffic are probably this close to calling it an AI firewall, just like a device that filters web traffic is called web application firewall. I'm surprised that API guys didn't call their API gateway API firewall. But I am I am 99% sure there's a startup somewhere that is literally selling an AI firewall, and by this they mean firewall to control AI traffic.
SPEAKER_03Of course, I'd like to point out that any firewall in that sense is an AI firewall. Like we have some decoders at Cisco that can tell you that its traffic is likely from agents. So in that sense, our regular firewall is already AI aware. But I promise too, right? The second is there are enough people who believe that AI for firewall management would make an AI firewall. Not firewall for AI, but firewall with AI. And that to me is both both of them are fun, actually. Both of the ideas are quite fun.
SPEAKER_04I I wasn't expecting this to go quite this way, but okay. So AI in firewall, listen, I'm I'm all for getting in a time machine, taking your my favorite Claude or ChatGBT or whatever agent you you want, traveling back to 2003 when I had a firewall with 525 rules and saying, help make this make sense. Because the tool, because if there's tons of firewall management tools, I would argue a a good like one of the better applications of AI would be to that craft.
SPEAKER_04100%.
SPEAKER_03I mean, yes, what can I say? I mean, it's it's it's only contra it's only contradictory, controversial for people who say AI changes everything. I think there's a crowd, you probably met some of them. I mean, of course, there's a crowd. There's AI changes nothing crowd, and AI changes everything crowd. Both to me are equally annoying. But the AI changes everything crowd, of course, assumes that firewall is irrelevant because, well, everything changed, right? So everything that we had before doesn't matter.
SPEAKER_04Yeah, but those people eight, seven, eight years ago said the same thing about cloud. Cloud changed everything, and it did, but also it didn't.
SPEAKER_03Yes, correct. Exactly. That's how I would describe it too. And by the way, firewall survived in the cloud. Uh I mean, I I lampooned on when I had the podcast, right? We Tim and me lampooned people who dragged firewall appliances in the cloud because we always thought that you need to have firewall in, like controlling access to network, but you probably don't need a firewall appliance in the cloud. People do it, and I I'm not gonna make fun of them probably anymore, but I still cringe a little bit when I hear that, oh yeah, we're just gonna get a VM with firewall in the cloud, virtual machine with firewall in the cloud.
SPEAKER_04So I think I don't know. So this is so this is where I think it's a little bit different. I think because we are talking about God, my eye is itching me like crazy. I'm sorry if I'm like rubbing my eyes here. I think when you're talking about an augmented inference machine that can that can has been demonstrated to bypass our own thinking and our own like what guardrails we would set for it to achieve a goal, I think virtual firewalls are no longer appropriate in those cases because I want a physical piece of hardware that the cable goes in, so there's a you know stateful inspection engine there, and then the cable goes out.
SPEAKER_04I think if you put a piece of code there virtually to replace that physical hardware box, I think you can find exploits, which I think you can do with the hardware machine as well. But there's something about my old school grew up tinkering with technology in the mid to late 90s that likes the fact that there's a physical wire that I can simply pull the power on and then you are cut off.
SPEAKER_03No, no, no, no. You can't say just pull the wire. You have to say execute a kill switch. You're so not 2026 compliant. If you pull the wire, you're not literally pulling the wire.
SPEAKER_04If I get listening, listen, I I I need I need so something I just decided. I need uh there's a feature in here in this in this podcasting platform that allows me to insert like random noises. I need the wah wah wah. Every time somebody says, from now on, you say kilo switch, you're gonna get a wah wah wah.
SPEAKER_00Anton, he he wants the physical switch though. So he's actually he's gonna pull the wire because you know he's got his fingers on that land cable. He's pushing the tab down and yanking it out. I'm not gonna push the tab down, I'm just gonna correct it.
SPEAKER_04Forget the land cable. I'm going for power, baby. You can't pass packets if the power's dead.
SPEAKER_03I mean, it would be called a physical power-enabled kill switch.
SPEAKER_04Love it. Absolutely love it.
SPEAKER_03Wait, wait, wait. Power disabled. It's power-disabled kill switch, not power-enabled kill switch. Sorry, I'm wrong. I apologize.
SPEAKER_04But we have to say that in a way that doesn't offend, never mind. That's a rabbit hole that I don't want to go down. We know we just realized somebody in product management is listening to this to this episode after we publish it and just said, that's a damn good idea. And then marketing's gonna write that up, and it'll be your fault.
SPEAKER_03Uh no, no, no, it's your podcast, so it's your fault. I've been I've been trained by PR teams for many, many years. So I know it's whoever's podcaster is that says his fault. It's not the guest's fault.
SPEAKER_00But wait, I'm not the guest. No ref podcast, though, Anton. It's still his. So he's he claims ultimate responsibility. Thank you. That's right.
SPEAKER_04It's the Ravy's fault. Come get me. I'm over here. Listen, I I I think that uh I'm I'm interested to see what happens with security as it evolves. Because the AI changes everything crowd, and then the AI changes nothing crowd. There's gonna be a push and pull on this. Obviously, I I think obviously the AI changes nothing crowd is kind of at least partially right when it comes to security. We're not abandoning every principle we've ever had just because there's an there's a new type of technology in the house, right?
SPEAKER_04Like, but also this is a fundamentally different type of technology than we've ever seen. So it's not nothing and it's not everything. So where are we?
SPEAKER_03Uncomfortable silence ensues. I I mean it's I am the reason I'm uncomfortably silent, was was uncomfortably silent, was because I'm constantly being kind of schizophrenic about it. Because I was to I was at an unprompted conference, right? It was a couple of weeks back. Yeah, and it it showed in as I pointed out on Twitter, now I can use the term agentic swarm unironically. That's one thing that I learned there. But it's the speed of change that you see at some of the top-tier elite places, and speed of and lack of change you see at some of the mainstream places just make me think of a chasm becoming wider, and uh what changes at a top-tier place versus what changed at a regional bank in Belgium.
SPEAKER_04Yeah.
SPEAKER_00Is just that's an interesting thing. I was gonna say, yeah, it depends on where you're at, right? I mean, are you a target of nation states? Are you a target of you know, script kiddies, or are you a target of nobody?
SPEAKER_03Yeah.
SPEAKER_00I mean, like that I think plays a little bit too. Like, how much does it change for you? How much does it not? I mean, it matters who's actually targeting. Like, is somebody gonna fire up one of these agent swarms to go against somebody's WordPress site because it means nothing to anybody? Probably not worth the cost.
SPEAKER_04You know, I I am genuinely curious though, because this is as we kind of started joking started seriously, and then went jokingly, and I think I'm gonna come back to seriously. The the notion of uh purpose-built security hardware has been uh out of favor for a long time, right? Like we you would hear lots of people laugh at you if you said, I'm starting a hardware security company, you know, a fire physical firewall, physical this, physical that. And then it sort of started coming back a little bit because we realized that there's capacity, there's there's some real value to having something physical there, a b on off physical switch, right?
SPEAKER_04Like I I like the fact that my my home automation controls lights and stuff, but sometimes I just want to go flip the switch, right? And say I'm like, you're not listening, flip, like we're done. So what is the what do you think, Anton? Where do you see the future of because now you work for a hardware company again? Haha. Give me your give me your unbiased opinion. No, what where do you think the future of hardware is in this new modern AI-enabled, AI-dominated security world?
SPEAKER_03So I am very not qualified to answer that question, whether for the employer or just in general, frankly. Because and and here's why. Here's why. Because I lived at a cloud provider and I became very skeptical about anybody that says, oh yeah, we're just gonna push it to an edge. But at the same time, I have a mobile phone, it has a local LLM, it does cool things, it's a pixel phone, of course. So so the point is that I don't truly understand where the balance would settle. I I had to argue with somebody, and in the recent blog that I wrote, it was about somebody who said, I'm just gonna do local inference because I don't want to pay for tokens.
SPEAKER_03And and that person just assumed that if you do local inference, everything is free, but if you pay for tokens, everything is not free. And I'm like, Yeah, no, it's almost like I don't have money for I don't have money for Uber, I'm gonna buy Rolls Royce. No, yeah, if you're gonna buy Rolls Royce, you need money, right? Like you're not gonna save money on Uber as a result. So that's why I honestly don't know. I hear stuff about local inference, I hear stuff about the sovereignty. I don't know where it'll settle. To me, living in Bay Area, it's just much easier to push a button and call Claude Gemini, uh Claude or Gemini.
SPEAKER_03Actually, these are my two go-to's these days.
SPEAKER_04Yeah. I actually do like I've gotten to I've been so I've been on on Google's AI or their work workbench, workplace, workplace for a couple of years. They gave me Gemini. I used a hell out of it. My kids actually, you know what's funny is so total sidebar, maybe related to nothing, but as a father of two type one diabetics, my kids are often put in situations where they need to go figure out like the carb count count of anything they're eating, and it's not on the package or whatever. We did this, we did we tried something yesterday that's important.
SPEAKER_04So, like, you know, you go out to a kind of a complex dinner and you're like, How's this gonna go? Well, we took a picture of this giant sushi bowl that came, and I said, Gemini, figure out this is I took a picture of the you know of the uh of the description of the menu and I said we're ordering one of these, one of these, one of these, one of these, one of these. Estimate the per piece carb count on that. You know, I'll be damned if it wasn't like probably within 10%. It was amazing both my kids' blood sugars were amazing.
SPEAKER_04Now, I also did like in my head independent but without looking at that, like what I would have given them, or I would have guessed based on carbs that I thought were in there. And I think it was probably more accurate than me because I didn't know some of the like some of the ingredients had carbs like sugars in them, which is amazing and it's explains everything. So automobility is important, but that this is like I'm really, really big on the applicability of AI in some of these places, in some of these places in life. Like we talk about firewalls, right?
SPEAKER_04And and insecurity. I am still a skeptic that AI will replace my sock. I'm I'm a firm believer, and you can enable the hell out of it, you can change where the work is. Like right now, we do a lot of manual triage. Everybody does, right? Manual triage, ticket chasing, context chasing. Like that should be eliminated. In fact, at my day job, we're doing that. We're eliminating that for socks. And I think that stuff goes away, but I don't think we're getting rid of the people.
SPEAKER_03Yes, because there are also a lot of messy things that AI is just not that great with. And I that's why I'm kind of we um that's why I'm voting very often for a large vendor over a startup, and it's nothing to do with the fact that I work for a large vendor. It has to do with the fact that data gravity matters. Like I I I I would it's a painful question, and you can convince me the other way. But if I have a sim where all the stuff is, it's so much easier. And again, these are more experiences from my previous employer, maybe.
SPEAKER_03So much easier to do AI well in a useful manner if I won't have to be like, oh yeah, this data is from this place, this data is from that place. It can work in both places, in both cases. But it just feels easier to do it better if you have the data.
SPEAKER_00Well, it comes down to, I mean, it it is about the data though, right? When you talk about AI, like how much it knows is going to be better and how well it does. I'd be interested, Raf. You know, you think about that program to be able to look at food and determine it. You know, did you run that through, you know, 50 times? Did it give you the same answer? I did not. You know what I mean? Like it would be interesting to see that, and also later on, like as it learns more, does it continue to stay accurate? Does it get better? Does it does it lose it? And same thing when we talk about getting into our sims and what can it do from a sock perspective today, as it learns more, does it get better or does it get worse?
SPEAKER_04Well, I think that depends on the so this is why like the the human right now is so so so crucial, and that you don't have an idiot at the keyboard, right? So and and this is not meant to be derogatory, it's it's actually, I think, a very fine point that's largely missed is when you get an output from an AI, you have to put somebody at it at at the to assess that answer that actually knows the correct answer, like whether this is accurate or not. Because then you can, especially like in a sock is in a sock situation where your your AI does research and analysis, gives you a suggested disposition based on the evidence it collected and tells you why, right?
SPEAKER_04And as an analyst, as a human being that's done this many times, you go, Yeah, that look like technically that looks correct, but that's wrong. And then you say, This is wrong because and then the next time it and that analysis happens, it takes your context into effect and says, Oh, yeah, I'm gonna correct myself and it gets better. If you don't tell it that it's or if you simply say, Yeah, sure, that looks right, now it's gonna know that it's right when it's wrong, or you get say that's wrong, but don't give it context, it doesn't learn, right?
SPEAKER_04So the whole reason that AI I think is valuable for us is actually the pivotal point, which is the human being that with the intelligence that we're gonna convey into it.
SPEAKER_03Yeah. I'm this whole adjusting on the fly based on in imperfect human data is is is something I've been observing lately a lot, where where I and I may soften my stance about if your environment is very messy, AI would never figure it out. I I think that even some of the half-spoken, half-implied things, or yeah, I think this guy knows and he says something like that. Sometimes AI makes surprisingly accurate guesses based on that.
SPEAKER_03So this is kind of notable and interesting.
SPEAKER_04Yeah, I mean, I I'm not saying it's wrong. I'm simply saying there are plenty there have been plenty of times where you I've seen in our in our technology and others where the disposition looks at evidence and you can look at that evidence to say, I see how you got to that conclusion. I would have gotten to that conclusion if I didn't have the experience I had two years ago. Let me tell you why this is different, right? This is this is the thing that I think AI does not account for based on pure algorithm, and that is the human experience.
SPEAKER_04Like we have gut feeling, we have set, we we have intuition, we say, I've seen something like that, not that, but something similar. Let me let me use my intellect. And we can't, I think we can and do feed that into the algorithm to improve the algorithm, but I don't think the algorithm will ever achieve that to quote say it importantly, out of the box.
SPEAKER_00We have that same like learned knowledge and experience too that you see amongst humans. I mean, you take five people at a company, everybody's got a different knowledge base of what's going on. That when somebody says, Oh, this is happening with this system, you yeah, but that system connects over to here, that's covered by this, so this isn't a problem, right? But it's not written down anywhere. So, you know, I think that's where I see the biggest headache when looking at like AI results is the fact that it's making a whole bunch of assumptions that it doesn't know, it just, you know, oh, this is public, and it's like, nah, that's not really public.
SPEAKER_00That's you know, that's behind all the right. But it it goes and creates all these attack paths based on that assuming this is publicly facing or this isn't set in some middleware or something that it didn't see, right? Which then ends up learn leaning into a lot of like, I don't know, I like to say wasted time trying to chase all that down to be able to prove like, nope, all these assumptions were wrong on here. Right, which goes back to the whole the more data that it has, the better it's going to be.
SPEAKER_04when yeah when it gets its results out yep uh listen i i think the this is this this is the the precise reason why teams that employ that higher diversity of thought perform better because yes over time we all whether we want whether we think about this or not tend to gravitate a team of two people tends to gravitate even if at opposing views over time they will gravitate towards yes right moreover it maybe it's a contrarian view if everybody and it would be like a completely in a box like people who say obviously we would be defending with agentic swarms in a year I bet in their startup they were in a room and they all decided it's a safe prediction to make nobody said but wait what about firewalls like people would be like what are you an idiot they would be like no no no no no defensive agent swarms that's obviously the consensus and everybody's like yep yep that's the consensus right it can be extreme but it's totally makes sense in the room well I mean that's that we've seen that behavior from from the agentic world where you know these things the this technology that interacts with its with parts of itself eventually goes off potentially on a completely different path right based on its own interactions with self.
SPEAKER_00So like humans do it of course algorithms will do it like technology largely mimics life why because we programmed it to well program purists would attack you immediately for saying programmed because but we programmed by outputting all the data having discussions we you have to air quote the programs but yes yeah yeah fair enough you know interestingly enough though Raph I think to that to what you're saying there the human side we are much more limited and restricted to going past the bounds that we should be right like you're given guidelines that say you don't do that you don't touch production you don't do this you don't go here and you know you see from some of these instances right an agent swarm may get together and be like hey I know we're not supposed to do this but I'm gonna do it right like it'll it'll push the bounds that it's not supposed to to find things that we wouldn't have seen that as humans for the for the most part I would say most of us aren't pushing those bounds right like we're like nope I've hit the wall if I go past that I'm getting fired the agent's like what are they gonna do?
SPEAKER_04Shut me down and it just keeps going in real life cycle would mess with production but you know yeah correct correct this is actually a really good point it came up quite a few times in discussions when people talk about agentic security and people copy too many of the human security to that they almost quietly assume there is some kind of deterrence but there's no deterrence there's no paper policy unless it's the agent isn't in the instructions it doesn't exist uh and it's like I mean alignment all that but like not really right yeah but you still have to you still have to a a a common thing i've observed I've observed and I've read is that like oh your your your policy statement your instruction uh files should be read only to to your ai agent because they will modify them like the thing will modify it.
SPEAKER_03Yep yep we had it with people I mean we're talking about I'm doing some research on the side about securing security inside the JDC harness and then immediately somebody says you guys do realize that agent can rewrite the harness and if you say that the if you say that it's unlikely but like come on it actually already happened so like why right you need to account for it.
SPEAKER_00I I I laugh at your unlikely I I think so it's like the guy that goes back and like walks the police officer back to the speed limit sign and just like changes it to like 85 problem solved.
SPEAKER_04Right listen I I think it's interesting that that that we kind of jokingly what we what we jokingly named this conversation about AI laughing at firewall policy but I think I I think that's kind of where we're at right at the moment in in time where we've we've birthed a technology brought forth to the world a technology that is a an ultimate enabler that we are that we have built with it it's you know it's it's like giving my kids a a task and then giving them basic rules and then letting them go knowing full well that do they do not have my depth of experience in life lessons so they're gonna do some dumb stuff right in in any complex like simple tasks sure move this cup from over there to over there they'll get that done right they're not gonna throw well my son might they they're probably not gonna throw it right so because it'll break but like hey can you go you know clean this room and I I need it to be spotless well there are many ways to achieve that goal and I can imagine my kids and everybody else's have can could get some pretty creative ideas on how to make a room spotless and I and and as as with AI agents so like we're back to and I think I've said this on this podcast before if you guys remember like early early 2000s those of us that were in IT we we lived through the the outsourcing offshoring age where everything was going to be amazing and better and we were going to have to we all we had to do was just describe our work to somebody they would do it for a fraction of the cost and then we then we would be unburdened from these manual pusha tasks.
SPEAKER_04Well you know what it turns out that by the time you wrote the description you had to be explicitly like crazy descriptive because if it wasn't in there it wasn't gonna get done or it would get done a different way. And so it actually more work than it was less as the lesson learned for me.
SPEAKER_03Yep I think many of us have seen that and I feel like people are who have done a lot of outsourcer guidance may not may be doing better with with agents.
SPEAKER_04I don't know that's my random guess as of this second I I think that I I mean I've seen people do it I've seen people the friends and family do it like describe something with painful level of details I my impression is that probably agents today can do this work without that much describing I am picking a random example that I have in my in my front of my eyes but but yes I would based on the intro you guys watched when you if you're watching this on video I would disagree because that I wrote that out as a script and it was amazing how many things how many times I got back a result and went what why would you okay fine don't do this try again like also don't do this and like by the time I I was done it was like a page and a half and I'm like this is close enough I'm not gonna get any better because now it's doing other weird things.
SPEAKER_04But I wonder I wonder where the application of AI to to every technology stops. I mean it is absolutely amazing and and I'm you know I I started as a skeptic I think that's a good thing for me. I have converted over to a healthy supporter yet still skeptic broadly like it there's an amazing amount of things AI can do for me right now. And it has saved my butt and frankly kept me from forgetting things that I would have otherwise missed and and through you know reading tasks and all that right so that's amazing that that that exists.
SPEAKER_04And sometimes you forget that you're talking to a machine and not a human and that's weird to me. My brain has a very as a as a as an older person my brain has a very hard time with this.
SPEAKER_00I think a big thing Raf especially like creative materials videos images those I think are a lot more difficult to get prompts right versus trying to do things like coding other things like that. I mean I I went and recently switched like one of my blog sites from WordPress to Hugo doing a static site through GitHub pages and all that try you know let's just try this out it's actually the first thing I've really done with Claude and I mean it started with the simple thing of hey I want to switch this WordPress site at this URL to a Hugo site using this theme.
SPEAKER_00And I mean it spit and it was literally like a very small prompt and I mean it spit out like okay we're gonna do this and it it walked through it went queried all the posts converted them I mean yes there were things that as I discussed with it like hey wait a minute this theme's not working dark's not working on this page or you know you still have HTML elements in here why didn't you convert this to Markdown? Oh well yeah we can do that here we just need to use this and I mean it very simple to be honest like to be able to convert all that over with very little effort but I've also done the same trying to create logos and it's a nightmare.
SPEAKER_00So I I think a lot of it depends on what you're trying to do and I think the the more creative side is a lot more difficult prompt wise yeah I I I'm I'm gonna tend to agree.
SPEAKER_04Hey we're running out of time but Anton you want to come back and talk about whether AI will favor attackers or defenders based on your RSA talk and whether you still stand by what you said.
SPEAKER_03I think we should I yes this is a really good one because I am very it's it's interesting. I think we're developing a bit of a telepathy because I was thinking about it last night and a little bit in the morning because because AI the unprompted conf kind of made me think that the speed race is lost. And people like Niels Provost and a few others basically remind us that you should not engage in a speed race with an attacker. You should do some other race. And so it's a little bit AI favors attackers and defenders but differently people assume sometimes that to beat the attacker speed you have to be faster.
SPEAKER_03But I don't think that's actually gonna work for political reasons. Defenders have more political reasons than attackers. Again so I am kind of this is boiling in my head and this is a perfectly good episode.
SPEAKER_04So so so keep that in your head but I think I would like to end this on is the speed is a speed the race of speed with AI the new land war in Asia a little bit yes a speed a speed war with an offensive AI is yeah yeah correct that that's fair and if you get that reference this is why you listen and why we're friends all right you guys this has been fun Anton thanks for joining us I I'm gonna I think that's what we'll do next time we'll we'll go back and we'll we'll go revisit your RSA talk. I have some thoughts on it I probably have a couple of pages of notes we'll I'll try to convert to something coherent and I'm gonna I'm I think I'm gonna grab my my alter ego Rafal uh Kitav who's apparently there's another Rafal in this industry which can't like it's not it's like the Highlander there can be only one we have to like duel it's shoot somebody's losing a head like this and that with a little bit of who's got who got some great graphic graphic uh AI work like who's gonna draw up that battle anyway that's a challenge to you guys listening anyway this has been fun folks thank you for joining us thanks for listening hope you get the chance to watch these videos hope you enjoy the new intro on the video side uh audio intro I think is gonna stay for a little bit longer we'll see what happens but Anton thanks man first of many sure thank you and I'm very much looking forward to it James always good yeah this was good I'm looking forward to that next one too the foot race yep all right the the attacker or defender who does AI favor that's the next topic we'll catch you guys another time another place on another Dhamma security rabbit hole podcast thanks a lot for listening everybody we'll see you later this is the month thanks for listening don't forget to leave my Donna review and share this with your friends by
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?...
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....