# What If AI Never Happened? The AppSec Reality Check

URL: https://securitytable.ai/what-if-ai-never-happened-the-appsec-reality-check/
Published: 2026-04-08
Duration: 2828 seconds
Season: 4
Episode: 7
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

**In this episode, we explore a simple but surprisingly deep question: what would application security look like if generative AI never existed? We break down how AppSec might still rely on deterministic, rule-based approaches, what we might gain in structure and rigor, and what we’d lose in speed, scale, and accessibility. Along the way, we debate whether AI is truly improving security or just accelerating existing problems, from “vibe coding” and false confidence in results to the growing gap between finding and fixing vulnerabilities.**

**We also get into the tension between human-driven security practices and AI-assisted workflows, and whether the biggest challenges in AppSec are actually technical at all or still rooted in people and process. Plus, things take a turn as we let AI weigh in…and roast us a bit in the process.**

**Per usual, it’s a mix of thoughtful discussion, strong opinions, and a little chaos.**

Follow AI Security Table:

➜ Home: [https://securitytable.ai/](https://securitytable.ai/)

➜ X: [https://x.com/AISecurityTable](https://x.com/AISecurityTable)

➜ LinkedIn: [https://www.linkedin.com/company/ai-security-table/](https://www.linkedin.com/company/ai-security-table/)

➜ YouTube: [https://www.youtube.com/@AISecurityTable](https://www.youtube.com/@AISecurityTable)

## Transcript

**Chris Romeo (0:11)**: Hey folks, welcome to another episode of The Security Table. This is Chris Romeo, joined as always by Isar Gemini. No, no, no. Isar Claude and Matt Gemini. They, they have, uh, pitched their allegiance to a particular, uh, AI. And, uh, for purposes of this conversation, I will be AI-less because I will try to represent the human side of the world. Uh, the, the challenge Matt's throwing before us is to ask and ponder the question: what if generative AI did not exist? Now, specifically in the domain of AppSec and cybersecurity, what would our world look like right now if this generative AI thing hadn't started running off with all of the future?

**Izar Tarandach (1:07)**: Oh God.

**Matt Coles (1:08)**: If it hadn't changed the future, right? Timeline-wise, uh, we're gonna call this one the man in the high AI. If things around 2023 had kept on going as it was versus split off into the AI world. The AI universe?

**Chris Romeo (1:22)**: It's a good question.

**Izar Tarandach (1:25)**: It's, uh, yeah.

**Chris Romeo (1:27)**: Would, is time travel required to answer this or no?

**Izar Tarandach (1:31)**: I'm still trying to decide which one of them, which one of them is going to be the bizarre award, the one with or the one without.

**Matt Coles (1:38)**: We could be predictive and generative.

**Chris Romeo (1:40)**: I just like the idea of time travel and I never have a reason to do it. So, I mean, it's, you never had a reason to do time travel?

**Izar Tarandach (1:49)**: No, I am too.

**Chris Romeo (1:50)**: Everything.

**Matt Coles (1:50)**: I mean, I mean, so we, we, if we're gonna get into that conversation, we need to figure out what ti— how you feel about timelines in time travel. Do you go back and can you change the past or can you never, yeah, I can never observe. You can only observe.

**Izar Tarandach (2:05)**: You can't change.

**Chris Romeo (2:05)**: Follow the Back to the Future rules.

**Izar Tarandach (2:10)**: So this weekend I tried to have a, and I didn't try, I actually managed to, I had a conversation with my AI. For you to explain to me the difference between the multi-world theory and the multi-universe theory. And that was an interesting conversation to have.

**Matt Coles (2:25)**: As opposed to say Star Trek, when they went back to the past, they didn't change the future, did they? Not in all cases.

**Izar Tarandach (2:33)**: Because there was a branch in there.

**Matt Coles (2:34)**: Right.

**Chris Romeo (2:35)**: Yeah. Proving Star Trek is better. On the other hand, Back to the Future.

**Izar Tarandach (2:41)**: Back to the Future does change the future back.

**Matt Coles (2:44)**: It does.

**Izar Tarandach (2:44)**: Because there's no branching.

**Chris Romeo (2:46)**: Let's see what other, what other, uh, although interestingly, when Dr.

**Matt Coles (2:51)**: Emmett went back to the 1800s, actually they had to fix, that's right. They had to fix the timeline in order to not break the future.

**Izar Tarandach (3:00)**: That's the timeline. Yeah.

**Matt Coles (3:01)**: Right. Yeah.

**Izar Tarandach (3:01)**: Mm-hmm. All right. So back to our AI, right?

**Matt Coles (3:05)**: So if, if AI never took us on the alternative path. What would, what would cybersecurity, what would AppSec look like today?

**Izar Tarandach (3:16)**: Well, for, for one thing, we would be getting much less spam because less people would be trying to sell us tools that they just AI slopped the heck out of it.

**Matt Coles (3:31)**: You mean, you mean you would, we'd be on the same path. We would have been on the same path of new tools coming out, trying to create new domains of security.

**Izar Tarandach (3:41)**: Yeah, but slower.

**Matt Coles (3:42)**: But slower.

**Izar Tarandach (3:44)**: And we wouldn't have a complete domain that has flourished with people having 20 years of experience on a 5-year field. And, uh, we would be treading to a lot less material that we have to decide, is this something that I have to pay attention to, or less. Mm-hmm.

**Chris Romeo (4:06)**: So you're saying there's been a glut of products and companies that wouldn't exist without AI because AI is being used to create them and to justify them. What do you mean by that?

**Matt Coles (4:21)**: You need, you need this because the AI explosion, and so you need this, right?

**Izar Tarandach (4:27)**: There is a whole lot of stuff that we have wanted for a long time. That has been instantia— instantiated by the promise of AI to be able to gather a lot of contextual data and make it reasonably easy to understand and process. All those wouldn't be around here if we didn't have the capabilities that we have today. Then there is all the classification stuff. Promising to put things into bins in ways that the regex couldn't, which also wouldn't be around.

**Matt Coles (5:07)**: Well, I mean, all right, so alternatively, machine learning, right, which is not GenAI— GenAI is an application of machine learning— machine learning could have advanced still, right? That's the classification problem. Could have solved the classification problem. And it did, but I think that it would've been a lot less accessible, right?

**Izar Tarandach (5:29)**: Exactly. The way that it came out of labs came because of that specific application of it that made it all of a sudden much more accessible. If you went back in time, uh, a couple of years and asked people what's PyTorch, they would be looking at you like, I don't know, something that, uh, Elon wrote to write a flamethrower. I don't know.

**Matt Coles (5:49)**: Right.

**Izar Tarandach (5:51)**: So, That there is that too. But I like the answers that we got from our AIs. Let's not say which AIs we're using and people can guess in the end from the answers.

**Chris Romeo (6:05)**: What do they win?

**Izar Tarandach (6:08)**: The satisfaction of being right or wrong.

**Chris Romeo (6:11)**: So nothing.

**Izar Tarandach (6:12)**: Okay.

**Chris Romeo (6:12)**: Nothing.

**Matt Coles (6:13)**: Nothing. Nothing. There's, it's free. It's free to enter and there's no prize.

**Izar Tarandach (6:17)**: Exactly. Exactly. Rules apply.

**Matt Coles (6:22)**: Well, you can be assigned points, but the points don't, you can be assigned points, but the points don't matter.

**Izar Tarandach (6:27)**: Exactly. Exactly.

**Chris Romeo (6:29)**: Fair.

**Izar Tarandach (6:30)**: Whose AI is it anyway?

**Chris Romeo (6:34)**: Okay.

**Izar Tarandach (6:34)**: So we asked it, what would it be? And Matt, what did you get?

**Matt Coles (6:40)**: Ah, well, uh, I got basically it would be manual. Reactive and bottlenecked by humans.

**Chris Romeo (6:53)**: Oh, how dare us humans to bottleneck something.

**Matt Coles (6:58)**: It would be an era of deterministic security as a bad— and then, and that's, and that, and the way that that's worded is as a bad thing. And if you think about that very carefully, deterministic security as a negative, versus the alternative of non-deterministic security as a positive. Think about that for a moment.

**Izar Tarandach (7:26)**: That's interesting.

**Chris Romeo (7:26)**: That, yeah, hold on. That's actually a positive. Cause when I think deterministic, I think predictable. Yes. Yes. That's the synonym I like to use for that just to make it make sense in my head.

**Izar Tarandach (7:39)**: Yes.

**Matt Coles (7:39)**: You run it, you run the same function multiple times and you get the same output.

**Chris Romeo (7:43)**: Right. Yeah. That sounds like a good thing.

**Izar Tarandach (7:46)**: It is. Not a bad thing.

**Chris Romeo (7:48)**: Yeah.

**Izar Tarandach (7:48)**: But consider this, who's answering and what their point of view is. You're asking something to tell you how the world would be if it didn't exist itself. So you would expect that it's going to say, well, it's going to be a bad thing because I'm not around. And one way to say I'm bad, bad thing, I'm not around if I am AI is, hold on, we're not talking, we're not talking AGI, we're talking generative AI. This is not AGI. This, this, this is just implicit bias of the thing.

**Matt Coles (8:15)**: Why would an algorithm—

**Izar Tarandach (8:17)**: Because it wants to sound helpful, right? So for it being helpful is it being helpful, something existing that is helpful. If that something is not there, then—

**Matt Coles (8:28)**: Probably we need to change the query then to say, if generative AI wasn't at use in AppSec, it still existed, but it was not available for AppSec.

**Izar Tarandach (8:41)**: So I got amused because I gave exactly word by word the same prompt to mine. And one of the things that it says is that the foundational tooling of application security, blah, blah, blah, blah, blah, existed before, blah, blah, blah, blah, blah. None of these are GenAI techniques. They are deterministic or heuristic-driven analysis engines. Then it says, they would exist exactly as they do and arguably with more investment and refinement if the industry hadn't been distracted by the shiny object.

**Matt Coles (9:13)**: So yours had no problem being existential.

**Chris Romeo (9:16)**: Nope.

**Izar Tarandach (9:16)**: Mine has absolutely no problem being self-reflective. And I think this thing is learning from me.

**Chris Romeo (9:23)**: I'm getting seriously, uh, so what, it's cynical?

**Izar Tarandach (9:29)**: You tell me from the next paragraph. Threat modeling, your domain specifically, is a structured reasoning exercise. PyTM doesn't use GenAI. Stride doesn't. The Threat Modeling Manifesto doesn't. The discipline would be exactly where it is because it is— its progress has been driven by methodology and practitioner adoption, not by token prediction. Yep.

**Chris Romeo (9:55)**: I don't know that I agree with that statement though, because you've at least seen in the threat modeling space generative AI be applied to various tools that help make threat modeling work, like DaVinci, to try to make the grunt work side of threat modeling easier. So it's, it's an, I, I think there has been some positive steps forward and I don't think it would truly be practitioner kind of driven.

**Izar Tarandach (10:32)**: And I agree with you, but as you bring DaVinci up, It was AI infused, so it was around the process, not around the threat modeling itself. So it made the process easier, but not the threat modeling more exact or more perfect or something else.

**Matt Coles (10:48)**: Is that the, is that the term we're using now is AI infused?

**Chris Romeo (10:51)**: Oh, I came up with this. It is.

**Izar Tarandach (10:53)**: Yeah.

**Matt Coles (10:53)**: Okay.

**Izar Tarandach (10:53)**: Yeah, exactly. Actually, I think I heard Chris Romey original.

**Chris Romeo (10:57)**: I think I heard someone else say it and I repeated it twice, 3 times as required and then made it my own.

**Matt Coles (11:02)**: As opposed to AI assisted. Anyhow, I updated my, I updated my query and, uh, I basically got the same, same results as before, but then, but then an updated verdict. Oh, in a world, in a world, in a world, yeah, without, without GenAI, AppSec would be more principled but less productive. We would see a shift towards formal verification and strong typing as the only way to mathemati— mathematically prove security rather than the best effort probabilistic security we often see today.

**Chris Romeo (11:42)**: Wrong.

**Matt Coles (11:44)**: Well, formal verification and strong typing would be mathematical ways of proving security and it would be structured and it would be principled.

**Chris Romeo (11:53)**: And they're so highly inefficient that you'll never release an application as opposed to using the probabilistic techniques that GenAI offer. Boy, I think I'm back on GenAI's team at this point. If formal verification is what we'd be stuck with without it, which is not the case because we've had that, it's been available to us since the '60s, right? This idea of, it's just that technology or applications became so complicated that to formally verify even an app that's running on one of our phones right now, would take 6 months.

**Matt Coles (12:32)**: Well, alternatively, you'd have to design things in a more structured way. And, you know, the benefit of Agile to be able to release quickly, but the downside of Agile to be able to release quickly. We've had this conversation before multiple times. Right.

**Chris Romeo (12:49)**: And in our sleep also.

**Matt Coles (12:51)**: Oh no, Agile.

**Chris Romeo (12:53)**: Yeah.

**Izar Tarandach (12:54)**: So apparently my AI is a bit more refined than Matt's, for those who are following the plot. And I am having—

**Matt Coles (13:03)**: you're using like the fancy, a fancy version.

**Chris Romeo (13:07)**: Yeah.

**Izar Tarandach (13:07)**: All right. And, and it's, it's tuned to, to an extent. And apparently it's been fed with the shit that I've been writing lately. So my prompt at the end of the answer, it gives a synthesis of the whole thing that it gave before.

**Chris Romeo (13:22)**: Aye, aye, aye, aye, aye.

**Izar Tarandach (13:24)**: The domain of application security would look substantially similar to what it looks like today. Because its core problems are structural, methodological, and cultural, not generative. The areas where GenAI has provided real value are primarily in acceleration—faster triage, faster drafting, faster pattern recognition—rather than in capability. The trade-off is that some of that acceleration has come with new risk surfaces and the dilution of rigor.

**Chris Romeo (13:56)**: The delusion of rigor.

**Matt Coles (13:58)**: So wait, delusion or disillusion? Delusion. Delusion.

**Izar Tarandach (14:01)**: Delusion. Okay. The challenges that would remain hardest without GenAI are the ones that are hardest with it. Scaling security expertise across large development organizations, making threat modeling a consistent practice rather than a checkbox, and closing the gap between we found the vulnerability and we fixed the vulnerability. Those are people and process problems. They were, they were there before LLMs and they still are. Hmm. Welcome to my AI TED Talk. It could have been mine.

**Chris Romeo (14:36)**: Now I, I do fear that your AI has been, that you're leading the witness here. Compromised. Because it seems like, it seems to be answering in a way that I would expect you to answer.

**Izar Tarandach (14:53)**: Yeah. Am I an AI?

**Chris Romeo (14:55)**: Maybe that's what we need to test. Uh, how do we, Matt, how do we reproduce the Turing test against these?

**Matt Coles (15:05)**: I, I don't know. Go, you ask, ask AI, I guess.

**Chris Romeo (15:09)**: We have to, well, that, that's just compromising this whole experiment, Matt. Um, let's see.

**Izar Tarandach (15:14)**: Wait, isn't that the premise of, uh, of, uh, Blade Runner? The reverse Turing test?

**Chris Romeo (15:21)**: Or is it the direct Turing test? I never thought of it as the reverse Turing test. Is it? I don't know.

**Izar Tarandach (15:30)**: It's, yeah, 'cause you're trying to, you're trying to prove that you're not a, a, a replicant. Yeah. So it's the reverse.

**Matt Coles (15:36)**: There is a, there is a, there's a site. I would not, I don't know if I would go to it, but Turing, TuringTest.live. Play the classic Turing test in a modern format featuring conversations with advanced LLMs like GPT-4 and Claude.

**Chris Romeo (15:54)**: Wait, GPT-4 is not advanced LLM. Get outta here. False advertising. Don't talk too loud.

**Izar Tarandach (16:04)**: Okay, so where do we land with the, the answers?

**Matt Coles (16:08)**: So, The common theme, I think, from mine is predictive and probabilistic today versus deterministic in the alternative world and rule-based, right? So everything being rule-based and obviously heavy human in the loop. So is that a good thing or a bad thing? I think we know that deterministic output is, is a positive in many cases, right? Especially with threat modeling and having humans in the loop. Obviously that's our need to have control. Right. But we can probably ignore that one, right? If we fully gave into the AI and let it do its thing, I mean everything from writing software to fixing software to fielding vulnerabilities and handling incidents, stuff would move along. Would it be perfect? No.

**Chris Romeo (17:22)**: How long before the world would end?

**Izar Tarandach (17:24)**: We'd be riding horses and how long before instead of MoatBook, there would be, What was the name of the other site? The one for programmers? Stack Overflow. Stack Overflow for agents. So it's funny you mentioned that.

**Matt Coles (17:43)**: So one of the output here was the Stack Overflow remediation loop.

**Izar Tarandach (17:49)**: Get out of my head, AI.

**Matt Coles (17:51)**: Without GenAI-powered coding assistance to suggest secure patches. I don't know if it does that, but Sure, the remediation cycle would be significantly slower. So when a developer is told they have a cross-site scripting vulnerability, they'd have to go look that up. So figure it out.

**Izar Tarandach (18:12)**: Put an asterisk on that one. I want to hear what Chris was going to say. And then there's something that my AI said.

**Chris Romeo (18:18)**: No, I'm afraid that you're actually an LLM. And given the fact that like you, you basically responded with, get out of my head. To the statement that Matt made, which made me think like you were already ahead of the answer here. And I'm really thinking we need to do that Turing test. I don't know what it's gonna tell me, but I'm slightly concerned.

**Matt Coles (18:41)**: I, I, I wanna finish the answer, but I'll wait for the response here first, Chris.

**Izar Tarandach (18:44)**: Uh, no, no, no, no, no. Go, go, go for it. There's no response here.

**Matt Coles (18:47)**: The tactic, the necessary tactic that would exist because, and, and again, this is being presented as I guess I'm reading it as a bad thing. Okay. Because of that conversation we had earlier about how AI are trying to be, they're trying to be helpful, but they're also, there's an existential discussion here to be had. The tactic, we would return to the search and paste era. Developers would spend hours on documentation or forums. Security teams would compensate by creating golden libraries, pre-vetted images, uh, pre-vetted internal SDKs that handle authentication and encryption. If you want to talk to a database, you must use the company's internal wrapper. It limits creativity, but ensures security by default.

**Chris Romeo (19:33)**: So wait, it's, it's arguing against doing those things? Or it's, hold on, I'm confused now.

**Matt Coles (19:39)**: It's, it's arguing that that's where we would be right now.

**Chris Romeo (19:42)**: That's where we would be, but that's a good thing. So like, so. We shouldn't have AI generate our each individual authentication. Let's just all have our own authentication. Whatever app you're working on, just have AI create its own authentication system for it. Sure, we'll have 500 different authentication systems. It's cool. AI generated it. It'll fix it.

**Izar Tarandach (20:04)**: That's dumb. So, so mine, because of the, because of the way that my, that my, uh, uh, system prompts are set. It gave me a contrarian view and it says what might actually be better without GenAI. So 3 things, less verbose coding, fewer insecure defaults. The volume of AI-generated code entering production without adequate review is a new attack surface that wouldn't exist. The just ask the AI to build it culture has lowered the barrier to writing code while doing nothing to raise the barrier of writing secure code. Okay, second thing, less false confidence in automated findings. There's a real risk that teams treat an LLM-discoded secure assessment as authoritative rather than, than it is probabilistic at best. Third thing, more investment in secure-by-design frameworks. If you can't patch the developer skill gap with an AI copilot, you invest more aggressively in frameworks, libraries, and platforms. Guardrails that eliminate bug classes by construction. I like my AI.

**Chris Romeo (21:13)**: That's kind of the opposite of what Matt's AI concluded.

**Matt Coles (21:17)**: Although it's interesting, why would you not?

**Izar Tarandach (21:18)**: But this is the contrarian view.

**Matt Coles (21:19)**: Yeah, but why would you not do that even in an AI-generated world?

**Izar Tarandach (21:24)**: It closes with, this is arguably the higher leverage approach anyway. So it agrees with you. Why wouldn't you? I like my AI. I like it.

**Chris Romeo (21:37)**: Just don't like it too much.

**Matt Coles (21:39)**: You know, you have a new, there's a new music video. It used to be I Like My MTV. Now it's I Like My AI.

**Chris Romeo (21:45)**: I think it's I Want My MTV.

**Matt Coles (21:48)**: Oh yes, it is. I Want My MTV.

**Izar Tarandach (21:51)**: I would like to see that clip.

**Chris Romeo (21:55)**: You can generate it with AI. Yeah, do it before Sora shut down yet.

**Izar Tarandach (22:01)**: And actually that clip brings us to the, brings us to the video also, but it brings us to the big thing of last week. Shout out to Artemis II. That was awesome. Mm-hmm.

**Chris Romeo (22:14)**: Yeah. This was the NASA rocket. The NASA rocket.

**Matt Coles (22:20)**: They're going, they're at the moon. I think they're at the moon now, going around the moon. They did have a they did have a big bang. They did have a big bang moment when the, uh, space toilet stopped functioning briefly.

**Chris Romeo (22:37)**: Indeed. That could be a big challenge cuz that stuff just floats around.

**Matt Coles (22:42)**: The, the Big Bang Theory, uh, did an episode on a space toilet that failed to function. Oh, uh, for those who don't remember, Wolowitz created it. Wolowitz, right? He's got to go to the ISS.

**Izar Tarandach (22:52)**: Yep. Ah, what a witch. I wish they had a follow-on just for him. Now instead we got Sheldon.

**Matt Coles (23:00)**: Young Sheldon. Well, and we got Sheldon's family spin-off again also.

**Chris Romeo (23:08)**: No, we haven't. Yeah, the somebody got married or something.

**Matt Coles (23:13)**: Somebody got married. Yeah, no.

**Izar Tarandach (23:15)**: Georgie and what's her name. No, no, it didn't. No. We didn't. It's a separate universe. It's a separate universe. Nope. It's a, no.

**Chris Romeo (23:27)**: That timeline was not us. Where do we go from here?

**Matt Coles (23:32)**: Back to our AI-driven lives?

**Izar Tarandach (23:35)**: Well, clearly my AI is more useful because it agrees with me.

**Matt Coles (23:40)**: Well, as we know, all models are wrong, but some are useful.

**Izar Tarandach (23:46)**: Never has that sentence made more sense than today. I don't think that Mr. Box, if it was actually him that said that, that there are questions about it, I don't think that he ever imagined that that sentence would be used in so many different occasions and it would apply so well.

**Matt Coles (24:10)**: And hopefully we don't have to owe him royalties.

**Izar Tarandach (24:16)**: True, true, very true. Oh yeah. Wait, wait, wait, wait. Okay. So let's, let's forget the AIs for a second and let's think it ourselves. If we were to, to answer the question and, uh, okay, we know that my AI agreed with me and I'm very curious to know how that happened. But, uh, apart from that. Where would AI be today if it wasn't for the shiny thing, GenAI? What do you guys think?

**Chris Romeo (24:46)**: Where would AI be? Sorry, where would AppSec be?

**Matt Coles (24:51)**: Okay, so I did update my question actually to ask if LLMs existed or generative AI existed in other domains, but not in AppSec. And I just have to read this one because it's— That's a good one. It's kind of funny. Perhaps the most dangerous aspect of this scenario is that attackers would still have GenAI. If LLMs exist in other fields, they exist for hackers to automate social engineering, rewrite malware, and synthesize exploits. So defenders would be fighting at human speed against machine-speed adversaries. The tactic would be a defense pivot towards extreme zero trust and runtime protection with microsegmentation and hardware rooted hardware roots of trust to ensure that even if AI-generated exploits break the app, it won't be able to pivot à la StrideLM.

**Chris Romeo (25:51)**: That sounds really good. That sounds like we should be doing that still, but I don't think it's actually true.

**Matt Coles (25:58)**: No, we should still be doing that.

**Chris Romeo (25:59)**: Think about it. Hardware, hardware root of trust and, and a solid hardware platform underneath and microsegmentation. You don't, but you don't, you don't use hardware to, to pivot between, you know, multiple sets of data on a given system. Right, right.

**Matt Coles (26:20)**: You would reduce the blast radius. You would loop reduce the blast radius. More broadly than just that database. Right. But, but yeah, you're right. We wouldn't use it for, well, unless you, unless you used it for identity purposes, right. That was if database access was tied to a physical identity, but that's not really, that's not really, no, it's also not really realistic because the system has to have access. right? And every container, like if you're talking about a, a containerized system, there's no physical, like all of those containers have access almost certainly.

**Chris Romeo (27:02)**: Yeah. Right. Yeah. So that, that almost your AI's response sounded really good, but when we start to unpack it, it's just a bunch of gobbledygook that doesn't really, doesn't really, it's not really a secure architecture, it's just a bunch of words that put together sound really good. And if we didn't stop and pick it apart, we could probably all confidently spout those words and get some people to believe us that this is a good thing. Right. But it would be them trusting our reputations, which is why I don't trust AI's reputation.

**Matt Coles (27:40)**: Is there a reputation system for AI? In my head, there is.

**Izar Tarandach (27:46)**: Yeah. So tangentially, uh, while you guys were going at that one, I, I asked my AI, uh, I, I haven't asked, I point out something. I said, you sounded a lot like a bunch of stuff that I had written and published in my, in its answer to me. That's a completely fair observation. And the honest explanation is straightforward. Your published work is almost certainly in my training data. Your book with Matt Coles, your conference talks, your podcast episodes on The Security Table, your writing about PyTM, the Threat Modeling Manifesto, your positions on integrating security into development rather than bolting it on separately, all of that would have been ingested during training. Blah, blah, blah, blah, blah, blah, blah. The practical implication for you, when you ask me questions that sit squarely in your area of published expertise, you should expect a high probability of getting your own thinking reflected back with a confident tone and no citation. That is useful as a sounding board.

**Matt Coles (28:44)**: Uh, that's useful.

**Izar Tarandach (28:46)**: Your, my, my AI is me. That is useful as a sounding board or for pressure testing articulation, but it is not useful as independent validation of your ideas. It would be worth keeping that distinction.

**Chris Romeo (28:58)**: So I need, I do need to know what AI this is because it's really pumping you up. Like I want to, AI is going to make me feel better about myself also.

**Matt Coles (29:07)**: It also, it also just admitted to copyright infringement.

**Izar Tarandach (29:11)**: I, the, the, the bit that I went yada yada yada over actually addresses that. That raises a genuine intellectual honesty problem. I presented it as, as though I was reasoning from first principles when I was substantially drawing on a specific body of work. I did not distinguish between things I can derive from general domain knowledge and things I absorbed from a particular practitioner's published perspective. And I could not make that distinction even if I tried, because I do not have clean provenance tracking on which training sources shaped which outputs.

**Matt Coles (29:48)**: So it knows you have, it knows it was trained on your data, but it doesn't know specifically its outputs are based on your data.

**Izar Tarandach (29:55)**: Yeah, it can't, it knows that it has some knowledge of a source. It can't trace output to source.

**Matt Coles (30:02)**: And of course not because of the way the neuro— I mean, the way the programming works, right? Underneath it, it can't. And you know what?

**Izar Tarandach (30:10)**: I'm fine with it.

**Chris Romeo (30:13)**: It's just because it's you. There's like another you now running around.

**Izar Tarandach (30:20)**: It could be worse. It could be trying to tell me that pasta is everything.

**Matt Coles (30:24)**: So the verdict on with GenAI, and I had, I turned off personalization. So my model has the ability to actually turn off personalization, meaning it doesn't use, it doesn't —it doesn't use prior threads to inform the current thread.

**Chris Romeo (30:40)**: I like picking on people's AI now. The AppSec domain— I'm so in love with it. They're like, my AI has this ability to turn this on.

**Izar Tarandach (30:47)**: Like, ooh, wow, so smooth, fancy.

**Matt Coles (30:52)**: Without GenAI, the AppSec domain would be much more bureaucratic. To stay safe, companies would have to be extremely restrictive about what technologies their developers can use. Innovation would be slower because new and unproven frameworks would be banned until the security team could manually write a 500-rule validator for them.

**Chris Romeo (31:16)**: That just seems mean. Yeah, we would— it would be maximum would be 400 rules that we would force somebody to write.

**Matt Coles (31:23)**: We would have solved the talent gap by reducing the need for talent through extreme standardization, turning security into an automated no rather than an assisted yes. That's a dark— that's a view. It's a horrible view.

**Chris Romeo (31:37)**: We would've solved the talent crisis because we'd have been said no all the time by just automating the answer no to people's questions about their security problems. Well, wait, we could have done that in 2002.

**Matt Coles (31:50)**: We, in 2022, we didn't. Why would we do that now?

**Chris Romeo (31:54)**: I mean, we should have. Wait, wait, wait. What, what could we do in 2022?

**Matt Coles (31:57)**: If we did, if we did, GenAI would never have taken off in a company.

**Chris Romeo (32:01)**: Yeah, because we would have been like, can we deploy this LLM? No, no, no. Feels like a toddler. That's the name of this system is Toddler.

**Izar Tarandach (32:13)**: That's how Toddler speaks to the world. Wasn't that fun? The time when security could say no.

**Matt Coles (32:20)**: As opposed to why? Why?

**Chris Romeo (32:23)**: I mean, if you remember, sometimes we used to say no just for the fun of it. Yeah, totally. We didn't even really care. We're like, nah, that's not gonna work. Too insecure. Tilted. I don't understand. Well, you need to go study because I mean, I have no security and—

**Izar Tarandach (32:39)**: even though, instead of saying no, I much prefer the silent look and nodding.

**Matt Coles (32:48)**: Or you're gonna say, how long, how am I, how long am I gonna nod before you walk away?

**Izar Tarandach (32:52)**: Yeah. Or the, are you sure that you just asked me that?

**Chris Romeo (32:58)**: I was thinking more about the disapproving look of just make them feel bad about themselves.

**Izar Tarandach (33:04)**: Disapproving, disapproving means that you are judging the merits of the thing that was just asked. Just staring at them.

**Chris Romeo (33:11)**: That's why I got into security. It was to judge the merit of things that people brought before me.

**Izar Tarandach (33:18)**: No, no, no, no, no, no, no, no, no. You're going to security to judge. People, not the merits of the things that you want to do.

**Chris Romeo (33:25)**: Oh, so now we're just judging people. I did not realize that was an option. Okay. Well, we don't need AI.

**Matt Coles (33:32)**: We don't need AI for that. You could, you could judge people.

**Chris Romeo (33:34)**: Let's have AI judge Yasar. Let's have AI judge you. How many times a day do you ask it that?

**Izar Tarandach (33:41)**: At least every morning. I have an agent for that. Hey, how am I doing this morning?

**Chris Romeo (33:47)**: You appear to be doing okay. Yeah, you're fine. You're fine. You're fine. It just pumps you up throughout the day through your ear.

**Izar Tarandach (33:56)**: Yeah. Calm down and carry on.

**Matt Coles (34:01)**: Let's see. Uh, who is Izar Tarn Das? Izar Tarn Das is a prominent figure in the science community.

**Izar Tarandach (34:06)**: Oh, once you know, please let me know because I keep asking my therapist and she has no idea.

**Matt Coles (34:14)**: Is Izar Taran Das, a, what do we want? A good security person?

**Chris Romeo (34:23)**: That's very subjective. Wait, wait, wait.

**Izar Tarandach (34:25)**: Now the judgment is going in a different direction. Are we saying good as in competent or good as in not being—

**Chris Romeo (34:32)**: It's morally good.

**Matt Coles (34:33)**: Morally good?

**Izar Tarandach (34:39)**: I come out as rogue.

**Matt Coles (34:45)**: Uh, I, I'm not, I'm gonna be afraid of what it finds.

**Izar Tarandach (34:50)**: No, not, uh, a common crit—

**Matt Coles (34:58)**: a focus on empowerment over gatekeeping. Say what?

**Izar Tarandach (35:02)**: No. Uh, focus on what? Okay.

**Chris Romeo (35:06)**: Here's what I asked. I said, I asked ChatGPT, is Izar Terran Dash morally good? And it answered, assessing whether someone is morally good isn't something you can determine objectively without a clear ethical framework. That said, here's a grounded fact-based review of Izar Terranova. What's known publicly— hold on, this gets better— he is well-known figure in application security and developer education. Two, he co-founded Security Journey, which focused on improving security coding practices.

**Izar Tarandach (35:36)**: Where's my check? Where's my freaking check? He co—

**Matt Coles (35:42)**: he co-founded Security Security Journey?

**Chris Romeo (35:44)**: Apparently, I don't remember you at the table, but here, I'll leave you—

**Izar Tarandach (35:47)**: Forget remembering, where's my check?

**Chris Romeo (35:53)**: I'll put this in the Signal just to show proof that I didn't make this up.

**Izar Tarandach (35:58)**: Among all the products that an AI could hallucinate that I co-founded, I would have been the most proud to have co-founded Security Journey. Oh, so that's cool. Kudos to the AI.

**Chris Romeo (36:13)**: That's a blatant, wow. I've not seen an error like that that I've caught. Chris, what have you been feeding your AI? I don't really talk to my AI. That's why it hallucinates. It's alone. I talk to the voices in my head primarily.

**Izar Tarandach (36:31)**: I mostly listen to them.

**Matt Coles (36:32)**: There are no recorded instances of ethical lapses, black hat activities, or unprofessional scandals associated with Isar.

**Chris Romeo (36:42)**: Scandal. Well, listen, when I think ISAR, I think scandal. Hashtag scandal. ISAR gate. ISAR gate. You just want ISAR gate. You're just asking for it. Come on. ISAR gate. Okay. So the, as I continue reading from a utilitarian lens, your work likely has a positive net impact?

**Izar Tarandach (37:11)**: Um, okay. This is starting to read as an obituary, so perhaps we should stop.

**Chris Romeo (37:17)**: Yeah, I'll have it, I'll have it write you an obituary too. Oh God, that I haven't tried yet. That's a bit morbid to write that about yourself. I like to do that on your own and your own AI. Oh, luckily for you it says I can't write an obituary for Isar because there's no indication that he's passed away. Oh, writing one for a living person would be misleading.

**Matt Coles (37:41)**: Okay, here we go. So I let the AI know that, that we're recording a, a podcast episode with him and it says, oh well, that changes, that changes things. Tell him, tell him. The AI has performed a sentiment analysis on his career and has some concerns.

**Izar Tarandach (38:05)**: Oh man. You bitch.

**Matt Coles (38:14)**: The manifesto purity tests. Is valuing individuals and your actions over processes and tools still applied when the individual is an LLM? It didn't ask that.

**Izar Tarandach (38:27)**: It did.

**Matt Coles (38:31)**: Is he ready to add a 13th principle to the manifesto? Thou shall not use ChatGPT to skip the hard work.

**Chris Romeo (38:39)**: No, he's not ready for that.

**Izar Tarandach (38:40)**: What, what, what, what, what, what?

**Matt Coles (38:41)**: Are you ready to add a 13th principle to the manifesto? Thou shall not, shall not use ChatGPT to skip the hard work. Because an LLM can generate a stride threat model in 3 seconds. It actually wrote ChatGPT? It actually wrote ChatGPT.

**Izar Tarandach (38:57)**: And your AI is not ChatGPT?

**Matt Coles (38:59)**: And it's not ChatGPT.

**Izar Tarandach (39:02)**: It's trolling you. Wow. It's trolling me.

**Chris Romeo (39:07)**: Even worse, it's trolling itself. You get Rickrolled here.

**Matt Coles (39:10)**: The PyTM paradox. He created a tool that forces developers to write Python code just to get a diagram. Ask him. Is this a roast?

**Chris Romeo (39:20)**: Did you tell us to roast him? I didn't. I didn't.

**Matt Coles (39:23)**: Ask him if this is actually a security tool or just a clever way to ensure that only people he likes are allowed to ship software.

**Izar Tarandach (39:33)**: Okay. Cut and paste that. Cut and paste that and send it to me, please. I want to read that very slowly. And then I'm going to go have words with somebody I know that works at Google.

**Matt Coles (39:47)**: Oh, here we go. On the Security Table Podcast, he's the voice of reason.

**Izar Tarandach (39:55)**: Ask you guys, I'm going home.

**Matt Coles (39:59)**: Ask him if that's a professional persona or is he just muted half the time because he's yelling at a clown?

**Chris Romeo (40:07)**: He's yelling at a clown in the background. A cloud or a clown?

**Matt Coles (40:13)**: A cloud. Okay.

**Chris Romeo (40:15)**: Sounded like it was calling us clowns, Matt. I thought it was referring to us as clowns. I was like, whoa, wait a minute.

**Matt Coles (40:21)**: He's lawful. I've categorized him as lawfully grumpy. He wants the world to be secure, but he's deeply suspicious of any solution that doesn't involve at least 3 hours of architectural debate and a reference to an O'Reilly book he wrote. Well, that's true. That's true.

**Izar Tarandach (40:35)**: That is true.

**Chris Romeo (40:36)**: I mean, you gotta get your reps in.

**Matt Coles (40:38)**: Pro tip for the roast: ask him if he's seen the latest GenAI threat modeling tool. When he starts his inevitable and justified rant about the hallucination of security, just tell him, 'Isar, it's okay. You can just admit you're jealous that the AI doesn't need a coffee break to find a spoofing threat.' Oh my God, that's hurtful. Damn. Wow. It's so hurtful. All right. So, Gen AI has earned its keep today.

**Izar Tarandach (41:13)**: So, you know what? My AI has words for you.

**Chris Romeo (41:17)**: Oh, wait. So now your AI wants to talk to his AI?

**Izar Tarandach (41:21)**: No, to him. Oh, okay. So Matt Coles, based on his security work. All right. I wrote. Here goes. Matt Coles, the security world's most dangerous stealth operative. The first thing you'll notice about Matt Cole's public security footprint is that you barely notice it. This is a man who co-authored the O'Reilly book on threat modeling, contributed to CWE, CVSS, and SafeCode, and has held senior product security roles at EMC, Analog Devices, Bose, and Albel Technologies. And yet, if you Google him, the internet basically responds with, "Did he moon the book?" His individual digital presence is so minimal it could itself be a case study in attack surface reduction. That's good. Oh man, that's good.

**Chris Romeo (42:09)**: That's good. That was well written. It's so digital. If you could be the example of a tax service minister.

**Izar Tarandach (42:21)**: Former colleagues would describe Matt as excelling at creating, establishing, and enforcing policies, methodologies, and procedures that are encompassing, clear, and concise. Read that sentence again. It is the most perfectly on-brand compliment possible for someone whose career has been spent inside large corporations making security programs work. It's an observation that Matt's contribution model is the kind that earns deep respect from the 30 people who are actually working with him. Wow.

**Chris Romeo (42:51)**: Yikes. Ouch.

**Izar Tarandach (42:52)**: Oh, God. He is doing the thankless, high-impact, unglamorous work of making enterprise product security actually function. Nobody writes breathless LinkedIn posts about that. It's basically saying that you are holding everything up, but nobody is going to write about it. Oh my God. Are you done yet?

**Chris Romeo (43:21)**: Hey, it's almost.

**Matt Coles (43:26)**: Oh, I wonder what it says about Chris.

**Izar Tarandach (43:29)**: I'm already nervous. Oh, wait, wait, wait. On the security table, you have 3 hosts. Chris Romeo is the polished, high-energy communicator. You, me, are the opinionated provocateur who starts philosophical fires. Matt is... the reasonable one.

**Chris Romeo (43:50)**: I believe that.

**Izar Tarandach (43:51)**: The one who says, well, actually, let's think about this from the other side. This is an absolute essential role. And also the one that nobody ever pulls. He does that a lot. Yeah. Yeah. He's the best player of the podcast. Without him, it would fall apart. Nobody buys a ticket to see the best player. That's mean. Oh, that is mean. That is mean. And he's probably going to read those and send you a single sentence text about it because that's exactly the kind of person that this roast describes. I like it. I like it.

**Matt Coles (44:33)**: And when I go postal, you'll wonder why.

**Izar Tarandach (44:36)**: Oh, wait, wait, wait, wait. You're gonna like this one. He is the threat model's threat model. Cruel, systematic, and quietly effective in ways that only become apparent when you look closely at the outputs. I, that, that's high praise. That's high praise. That's not a roast. I'm okay with that. No, no, no, no. This one is high praise. I, I would happily wear that one. Maybe that's a t-shirt. All right.

**Chris Romeo (45:03)**: Well, we're approaching the 45-minute mark, so it's time to end this experiment in something.

**Izar Tarandach (45:09)**: I don't even know what to call this experiment in.

**Chris Romeo (45:11)**: So we got roasted. We got confronted, you got roasted, and you got schmosted. I can't think of another word.

**Izar Tarandach (45:21)**: No, what, what, what was it? You got weighted, you got measured, the one from The Knight's Tale, and you were found wanting, something like that. I have to watch that movie again.

**Chris Romeo (45:33)**: You lost me. You lost me at—

**Izar Tarandach (45:37)**: you haven't watched The Knight's Tale?

**Chris Romeo (45:40)**: Probably a long time ago.

**Izar Tarandach (45:42)**: A movie that begins with a joust tournament with We Will Rock You in the background and you have not watched it.

**Chris Romeo (45:50)**: I'm sure I did. I mean, it's, it's an old movie. It is. So like, oh my God, it is. I just don't remember.

**Izar Tarandach (45:58)**: It is an old movie.

**Chris Romeo (46:00)**: So, all right. Well, folks, we will have to close it.

**Matt Coles (46:02)**: Chris, you didn't see your roast.

**Izar Tarandach (46:03)**: You didn't see your roast, did you?

**Chris Romeo (46:05)**: I'm, I'm, I, I watched, I read, I watched it on ChatGPT and it was actually very nice. So it was a video, it generated a video, bros, for me with Snoop Dogg was hosting with Elon Musk and some of my other friends. So, all right. But with that, we're outta time. We're at the 46-minute mark. It's been a long, strange trip. It's what most people write in their high school yearbook.

**Izar Tarandach (46:27)**: Trip, trip, trip, trip is the operative word here.

**Matt Coles (46:30)**: Yeah. We, we, we did the what if and, and fell off the cliff.

**Chris Romeo (46:34)**: Yeah. We fell down. Rabbit hole to land in the wormhole to, I don't know what other—

**Matt Coles (46:41)**: Some hole.

**Chris Romeo (46:43)**: I don't even know where else we go from there, but thanks for tuning in to this episode of the Security Table. Next week we'll be back with a more structured example of how you should have a podcast. Thanks.
