# The Walking Dead of Security: When AI Resurrects the Build vs. Buy Debate

URL: https://securitytable.ai/the-walking-dead-of-security-when-ai-resurrects-the-build-vs-buy-debate/
Published: 2026-01-28
Duration: 2422 seconds
Season: 4
Episode: 2
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

Are cybersecurity technologies really dead, or are reports of their demise greatly exaggerated? Today’s episode is a discussion on how AI is reshaping the classic build vs. buy debate, empowering non-engineers to create working prototypes and potentially reviving the DIY coding culture of pre-open-source days. We also talk about how developers trained on open source are now leveraging AI built from that same foundation, raising questions about innovation and originality in modern programming.

[Build vs Buy is Dead - AI Just Killed It](https://venturebeat.com/technology/build-vs-buy-is-dead-ai-just-killed-it) 

[Traditional Code Review is Dead](https://thenewstack.io/traditional-code-review-is-dead-what-comes-next/)

  

  
  

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 (0:10)**: Hey folks, welcome to another episode of The Security Table. This is Chris Romeo, joined as always by Matt Coles and Isar Terrandas. And it is late January 2026, and we are in the midst of snowpocalypse, is what we've been calling it here in the South. We have roughly 2 inches of some type of white substance on the ground with some ice on top of it. And it is, it's not a 1970s kind of club event.

**Matt Coles (0:43)**: So that's just icing on the cake.

**Chris Romeo (0:46)**: Yeah. Well, I actually haven't gone out to test it now that you make that kind of reference. I thought it was just snow and various precipitation, but do you have lawn furniture?

**Matt Coles (0:56)**: Because you could have a table cake.

**Chris Romeo (0:57)**: That's true. I do have some lawn furniture that I'm looking at out the back window here. I don't see a cake that's been provided though. So yeah, and you are—

**Matt Coles (1:08)**: you guys are in the United States North, uh, where the table and the cake have become one.

**Chris Romeo (1:16)**: Yeah. I didn't spend most of your night, uh, you didn't get one of those autonomous, uh, snowblowers I've been watching on social media.

**Izar Tarandach (1:24)**: I haven't. My wife has.

**Chris Romeo (1:26)**: Yeah. His name is Isar. No, they have these battery-powered, they're like a Roomba, but it's a snowblower.

**Matt Coles (1:35)**: Battery-powered Roomba.

**Chris Romeo (1:36)**: Yeah.

**Matt Coles (1:36)**: Snowblower.

**Chris Romeo (1:36)**: It's snowblower. But like, so I've been watching these videos of people, they seem to be relatively powerful. Like people are running them through this whole storm. They're just going, they go and run and then they go like the Roomba, they go back to their battery charging station and then they launch back out and keep, blowing out the snow. So I just figured, Matt, you, of everyone I know, you would be the one with an automated snowblower.

**Matt Coles (1:59)**: Oh no. Yeah. I'd have to mortgage my house to, to get one.

**Chris Romeo (2:03)**: But I don't have no, I don't have any idea because I live in the south. Like, I'm like, there's no way I would even be able to use, like, I would've been able to, I'd be able to use it once every 3 years. So I was like, I didn't even bother looking to see how much it was.

**Matt Coles (2:16)**: You know, if the batteries were interchangeable, you could use it for your, your leaf blower. Actually, the snow is light enough that somebody was— I saw somebody using a leaf blower to blow the snow.

**Izar Tarandach (2:26)**: I do that.

**Matt Coles (2:27)**: Nice.

**Izar Tarandach (2:27)**: On my deck, I blow it with the leaf blower. But Matt wouldn't be the one running the thing. He would be the one running after the thing with a laptop and an I2C connector. Wait, wait, I want to talk to you.

**Matt Coles (2:40)**: Come here.

**Chris Romeo (2:44)**: Well, this is the zombie episode of security table. It's kind of a montage to The Walking Dead. Brave.

**Matt Coles (2:54)**: Yeah.

**Chris Romeo (2:55)**: Of cybersecurity.

**Izar Tarandach (2:56)**: We, we just connect to the snowmageddon cuz, uh, we were expecting like to, to lose power and everything. Hey, hipster Conrad, no losing power.

**Chris Romeo (3:05)**: All of a sudden, bam, fades to black. Like, wait, what?

**Izar Tarandach (3:11)**: Oh no. Hard Mecca.

**Chris Romeo (3:14)**: Power outages ravaging the East Coast. All right, well, we got two different takes. And Yezar, I think you had some, uh, some apt analysis here that, uh, apparently everything is dead.

**Izar Tarandach (3:28)**: Yeah, the thing is, and even we have done that once or twice, but apparently everything is dead. AppSec is dead. Uh, I started hearing that even LLMs, a certain kind of LLMs is already getting dead because there's a new kind of LLMs coming on. So everything is dead. We live in a zombie world and I don't know.

**Matt Coles (3:53)**: Batteries. Remember Batteries Included?

**Izar Tarandach (3:55)**: Right. I think that the rumors of a lot of things' death have been greatly exaggerated, as they say.

**Chris Romeo (4:03)**: Who is that? This was, who's that quote from?

**Izar Tarandach (4:05)**: Uh, one of the presidents.

**Chris Romeo (4:07)**: Is it Hemingway or is it Mark Twain?

**Izar Tarandach (4:09)**: No, no.

**Matt Coles (4:10)**: Batteries Included?

**Izar Tarandach (4:11)**: No, the rumors.

**Matt Coles (4:13)**: Oh, the rumors of my death are greatly exaggerated.

**Chris Romeo (4:16)**: We have all this LLM power.

**Izar Tarandach (4:22)**: Mark Twain?

**Chris Romeo (4:23)**: It was Mark Twain.

**Izar Tarandach (4:24)**: All right.

**Chris Romeo (4:24)**: Yeah. I was right, right with my second guess. I don't know why I went with Hemingway.

**Matt Coles (4:28)**: Mark Twain. Yes.

**Izar Tarandach (4:29)**: No, wait, wait, wait. There's a foreign joke. Apparently that quote is dead.

**Matt Coles (4:33)**: It's greatly exaggerated.

**Chris Romeo (4:35)**: Yes. Everything's dead.

**Izar Tarandach (4:37)**: That quote is dead. Uh, that there's a lot of attribution. John Davis, Mark Twain. Yeah, no, we can't go down that rabbit hole right now. No.

**Matt Coles (4:51)**: Well, or raven hole. It's very dark.

**Izar Tarandach (4:54)**: So let's talk about this build versus buy is dead. AI just killed it. Okay.

**Matt Coles (5:01)**: Oh yes. Let's do that.

**Izar Tarandach (5:03)**: So it comes from VentureBeat.

**Chris Romeo (5:06)**: Well, that's people that are out there doing things.

**Izar Tarandach (5:09)**: Right?

**Chris Romeo (5:10)**: Well, already successful people.

**Izar Tarandach (5:11)**: Or pay, yeah, paying for other people to do things. So perhaps they're happy because now they're paying for AI agents to do things and they can pay less. Yeah.

**Chris Romeo (5:21)**: So I mean, their premise that I'm seeing here is that AI creates a world where non-engineers can generate working prototypes and code using natural language prompts. So not really anything new. Like we kind of knew that already, right? That people have been vi— this whole vibe coding we've talked about in the past. All right.

**Matt Coles (5:43)**: I am, I'm going to, I need to bring it up. What's old is new again. Remember before the world of open source, people used to build their own components.

**Izar Tarandach (5:52)**: Remember? Yeah.

**Chris Romeo (5:54)**: I don't remember that.

**Matt Coles (5:55)**: You are, we are all old enough to remember that.

**Chris Romeo (5:58)**: Speak for yourself.

**Izar Tarandach (5:59)**: And now that we are all, we are all old enough to have forgotten that there is a time.

**Chris Romeo (6:05)**: We can just say we are all old.

**Matt Coles (6:06)**: You may remember there, there is, there was a time. Where open source was considered bad.

**Chris Romeo (6:11)**: Yes.

**Matt Coles (6:12)**: And building components was good because you have control. And then open source came along and now reuse was good and reinventing the wheel was bad. And now about that, here we are back again with reinventing the wheel is good because the AI is reusing all of that old code.

**Chris Romeo (6:32)**: So it's like combined reuse I don't think that, I don't think the people that are, that are generating working prototypes using LLMs even know any of that that you just said.

**Matt Coles (6:46)**: No, I know.

**Chris Romeo (6:46)**: They just take what the, what the LLM spits out as—

**Matt Coles (6:50)**: wait, but Joe Surreality, that's what's happening. Back in the old day, we used, back in the old day, like 20 years ago, we used to invent our own code.

**Izar Tarandach (6:58)**: Back in my day.

**Matt Coles (7:00)**: And then open source came along and we used to reuse their code. And now we have AI reusing their code to reinvent the code.

**Izar Tarandach (7:07)**: You, you're just reminding me of like The Cathedral and the Bazaar. Mm-hmm. Remember the book and how if you use open source, you are basically a communist because it's soon as you put that code into your code, then your code becomes everybody's code.

**Chris Romeo (7:24)**: Yeah.

**Izar Tarandach (7:24)**: And all that stuff.

**Matt Coles (7:26)**: And that's different with, uh, AI that's been trained on public models, you know, public models that are trained on GitHub and whatnot.

**Izar Tarandach (7:36)**: So here's the interesting thing.

**Matt Coles (7:40)**: Oh, I don't want to go there.

**Izar Tarandach (7:42)**: It could be said that today's developers have also been trained on open source code, right? Because yeah, some learned coding at the university and whatnot and had the whole computer science thing. Explained to them very nicely. And some just, hey, made stuff work, right? And then you look at a piece of code that does sort of what you want to do and you learn from it and you move on. And where do you get that piece of code? Mostly from open source.

**Matt Coles (8:16)**: And, uh, oh, didn't you recently post something about this on LinkedIn? Is our—

**Izar Tarandach (8:21)**: what, the Stack Overflow? Yeah, I was just, I was just telling Chris about it. So the thing is, I was talking to a developer and we ended up, as one does, talking about coding agents. And he was like, yeah, remember when the solution was to go to Stack Overflow and get whatever code you needed from there? So he remembers when Stack Overflow was the solution. And I had to say, hey, you know what? I remember when Stack Overflow was a problem. And apparently not a lot of people on LinkedIn Connect to the fact that there is Stack Overflow, the place where you go to read code. And once upon a time, Stack Overflow, you're writing C or any other non-memory-managed language, was actually a problem. You could have a stack overflow and blow out of your recursion or whatever.

**Matt Coles (9:14)**: Yeah, it's where you went to read code and then occasionally you would just borrow that code and it would end up in your production stack.

**Izar Tarandach (9:19)**: Yes, exactly. Exactly. Right. And, um, uh, I don't know, it was, it was just a bit of a funny way to, to close the week to me. Cause I was like, dude, you know, Stack Overflow had its own problems and nobody knew if the code coming from there was actually good or not, but people were just cutting, cutting, pasting. And now we look at it and we say, oh, the good old time, good old days, right?

**Matt Coles (9:44)**: Yep.

**Chris Romeo (9:45)**: And now it gets used to train all of the models that are out there recommending code to people.

**Izar Tarandach (9:50)**: Exactly, exactly.

**Matt Coles (9:51)**: Yeah. So, but you know, I'm looking at the article and it's an interesting one. And I think there's a really important point about halfway down the article that they actually make. And I'll be modifying my sort of soapbox a moment. In the past, we used to simulate real-world systems to figure out how they function and whether it's something we want, right? We would try different concepts and then pick one of those concepts and move forward and see if it worked. What the article seems to be suggesting is not that, not that they build components, you know, bespoke components because AI can. The suggestion here is build to learn what to buy. And I'll extend that to build to learn what to integrate. Right. So, if that's the, if that's where it stops, meaning rather than building a component bespoke so that you don't have to buy anything, this is prototyping, right? This is concept engineering. And that's not horrible, right? As is our, I think you've mentioned AI can't brainstorm. AI can brainstorm, but AI can't invent, right? It can only build what was previously known or stitch together previous things, it may end up with something new, but it did it by accident. And so we're talking about conceptualizing ideas to inform humans on a path forward. And that's not a bad thing.

**Chris Romeo (11:26)**: I don't get how this works in the real world though. So like, okay, I'm thinking about buying a customer relationship management tool for my company. So I ask AI, build me Salesforce. So that I can, so that I can, uh, evaluate it and get real clarity about what I need from a feature perspective.

**Matt Coles (11:47)**: There's an open source version of Salesforce somewhere, I'm sure.

**Chris Romeo (11:49)**: AI just starts spitting out smoke, like it's like smoking out of the back of it. 'Cause Salesforce is probably 100 million lines of code or more, and it's a big thing, right? So like, I don't get how that's really practical advice to say, I'm gonna have AI build a knockoff of the solu— the problem, the solution that I'm evaluating, and then I'm somehow gonna be better informed. Like, if you don't understand the problem enough to choose an enterprise-grade solution, how are you gonna describe it to the AI in such a way that it gives you something of value?

**Izar Tarandach (12:24)**: So, all right, so let, let, let's take some things with a grain of salt here. The, the person who authored this, uh, piece, let's call it piece. Has a very non-veiled interest in, uh, in this scenario being true. But what I find— okay, I don't want to use that word. What I find a bit of a problem here is that, uh, on one side they infantilize everybody, and on the other, on the other hand, they, they, they make everybody a superhero. So they're saying you're not capable to know what is it that you need right now for your business, but you are really, really capable to go take an AI and build it in 2 hours and show 80%. No, that's exactly what he's saying here.

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

**Izar Tarandach (13:13)**: After 2 hours, they have 80% of the functionality. Dude, after 2 hours, I haven't found yet the door to the bathroom. I mean, there's no—

**Matt Coles (13:23)**: turn on the lights. There is our—

**Chris Romeo (13:25)**: if you could, if you could prototype, you could build something to 80% after 2 hours, you could invest another 10 hours and get to 100% in this logic and you wouldn't have to buy anything.

**Matt Coles (13:36)**: Exactly.

**Chris Romeo (13:37)**: You could just use your newfangled AI-generated solution. Now, you know what they didn't mention in this article?

**Izar Tarandach (13:43)**: But wait, wait, wait, wait, wait, wait. Before you go to, before you go to the, the thing that links these both things of infantilizing and superheroizing, whatever, is if I don't know what I need, how am I going to get to 80% of it?

**Matt Coles (13:58)**: What's 80%?

**Izar Tarandach (14:00)**: Like, who's going to decide what is this 80%? I, I, I, as you said, uh, uh, Chris, I just tell the AI, build Salesforce, and I have one agent that goes down and researches what is it the Salesforce does, and another one that comes and builds a plan, and another one that builds a to-do list, and another one that checks the, like, how in this person's mind the physics of this thing work? I don't know what I need. 2 hours later, I have 80% of what I didn't know that I needed. Now I can look at this 80%, understand what is it that I need, and decide if I want to buy or if I want to spend 10 more hours and build the other 20%. I mean, in which world does this make sense?

**Matt Coles (14:44)**: Well, okay, to be fair, they said built a version of this last night. That could have been 2 hours or that could have been 10 hours. Who knows how much they sleep?

**Izar Tarandach (14:51)**: No, no, no, no, no. He says here 2 hours. He says 2 hours.

**Matt Coles (14:54)**: Uh, oh, okay. Yes, you do. You're right.

**Chris Romeo (14:57)**: Yeah.

**Izar Tarandach (14:57)**: In the beginning.

**Matt Coles (14:58)**: Yeah.

**Izar Tarandach (14:58)**: At the end he says last night, but in the beginning he says 2 hours.

**Matt Coles (15:01)**: And then he says, shows them something built in 2 hours. Yep. Yeah.

**Izar Tarandach (15:06)**: But even if it was last night, 10 hours, really?

**Matt Coles (15:12)**: Like, where's the logical consistency here? Take a, take a smaller example, right? They didn't build Salesforce, but they built, um, Nessus. They reinvented Nessus. And I chose that specifically because I know Isara's background.

**Chris Romeo (15:31)**: I mean, it wouldn't take long.

**Izar Tarandach (15:34)**: Oh, I'm just thinking of the— I'm glitching here. I'm just thinking of the commercial where the ladies sit in front of the wall and they paste like Polaroids. And the lady is saying, this is my wall, this is my social media wall, and somebody says, this is not how those things work? This is not how those things work! I mean, I'm still trying for the logic consistency here. I understand we live in a world where logic went out of the door that I couldn't find a long time ago. But seriously, if you're going to step up and say X is dead, Don't you think that, uh, you owe X a nice way of killing it at least? And not this, this, this BS of I can do whatever in 2 hours with an AI agent. You can't.

**Chris Romeo (16:28)**: Well, apparently according to AI, declaring something dead is intentionally provocative to force reconsideration.

**Matt Coles (16:37)**: Is that the new Gen Z term?

**Chris Romeo (16:39)**: I don't know. I don't know what that means.

**Izar Tarandach (16:40)**: I don't do Gen Z.

**Chris Romeo (16:42)**: No.

**Matt Coles (16:44)**: SNL had a, SNL had a Gen Z translator on, uh, this weekend.

**Izar Tarandach (16:48)**: So you have a nephew? It's true.

**Chris Romeo (16:53)**: I just need to talk to, uh, I talk to my children sometimes to say, what does this mean? Because they say, I'm not, I'm not, I'm not joking though. Like I'll, they'll say something or I'll hear somebody else say something and I'm like, I don't know what that is. And I'll ask them, what does this mean?

**Matt Coles (17:08)**: I wonder if Google Translate will take care of that for you.

**Izar Tarandach (17:11)**: I just assume that it means nothing. You know, because again, saying that something is dead nowadays means nothing. You can kill and bury something, it's coming back.

**Chris Romeo (17:28)**: So you're saying that saying something is dead is dead.

**Izar Tarandach (17:33)**: Yeah, basically.

**Chris Romeo (17:35)**: Very meta.

**Izar Tarandach (17:36)**: No, wait, I'm saying that we should kill saying somebody is dead, something is dead.

**Chris Romeo (17:42)**: Hmm. Okay. So, I mean, at the end of the day, do you disagree with this?

**Izar Tarandach (17:49)**: Uh, completely and totally.

**Matt Coles (17:50)**: That build versus buy is dead?

**Chris Romeo (17:53)**: Yeah.

**Matt Coles (17:54)**: Well, okay. So I think that they bounced around that concept. In the first place.

**Izar Tarandach (18:00)**: So I think that somebody's trying to sell us something.

**Chris Romeo (18:05)**: So somebody's always trying to sell you something. That's how the world works.

**Matt Coles (18:08)**: I think, I think the important part of the VentureBeat article though, we're talking, oh, actually we are talking about the VentureBeat article still.

**Chris Romeo (18:13)**: Yeah. Yep.

**Matt Coles (18:13)**: So, so in that one, it's, it's kind of confusing because again, they're talking about being smarter about what they need to buy, which is different than implementing something to replace what they bought or to not buy something in the first place. Right? So if you're doing concept development and simulating and you can mock it up and say, hey, uh, company XYZ, I know you have your pitch deck, but here's what we actually want.

**Izar Tarandach (18:46)**: Right?

**Matt Coles (18:46)**: We mocked up what the interface we'd like it to look like or what functionality we want, and they can make a better purchase decision, right? Hey, can you do this for us or something like this? Or do you already have this feature, right? Versus we don't need to buy from you because we just invented it ourselves, right? Or we don't need to buy, take this open source component that's already been pre-running, pre-built and code reviewed for past decade and, you know, is functional. And we, we just created a new version of that ourselves and are using it cuz well, you know, you can put into production faster. That's two different things. Okay.

**Izar Tarandach (19:22)**: So as somebody who has done a startup or two, two things come to mind. First of all, it seems to me that marketing is dead because there's no interest for me. Everything is dead. That there's no interest for me to come into a boardroom and show somebody my deck. If the guy there in the back taking notes is just going to slip out to his coding agent and come back with 80% of what I'm showing in 2 hours, right? But then it comes to the, well, I'm a startup. I'm, I'm really small. I need some capital. I go to a VC and I show my deck. And now they're going to go to their team in the background and say, oh, look at this deck. Can you do this in 2 hours, 80%? Oh yeah, we can. Well, what about, what about the marketing? How do I, how do I, uh, I get that?

**Matt Coles (20:13)**: What about if the startup does capital? Well, what if the startup does that, that 2 hours of work a ahead of time and, and tries to predict what the VC or what the customer's gonna do. And so they go, hey, look, and by the way, if you wanna do some comparative shopping here, we did it for you already.

**Izar Tarandach (20:29)**: Oh, you just went into, I know that you know that they know that I know that they know that this is what I want to buy.

**Chris Romeo (20:35)**: And I know what you did last summer.

**Matt Coles (20:38)**: Well, you know, if you have, if you have infinite age, if you have infinite agents, why not just test the agents? Predict what this customer might want or predict what any customer might want and produce various versions of it.

**Chris Romeo (20:50)**: I was gonna say, are we getting to a future where every, like a company creates a million different versions of what their product is tailored to you? Like it's the Burger King, you can have it your way.

**Izar Tarandach (21:01)**: I'm going to write that and I'm going to call it an infinite army of monkeys. Right?

**Matt Coles (21:08)**: So just, just think about that, that concept though, with infinite compute power, why not?

**Chris Romeo (21:15)**: Of having, like, my, my, so the solution I buy from ACME Corporation and that they deliver to me is not the same one that you buy, even though, like, There's a way of doing where that's possible.

**Izar Tarandach (21:29)**: No, that's already true. If you go to any, any SaaS, okay, today, just to use a name that you already mentioned, two companies saying we use Salesforce. Can you say that they use exactly the same suite of products in exactly the same way with exactly the same kind of results?

**Chris Romeo (21:45)**: No, it's true. No.

**Matt Coles (21:46)**: The un— well, the underlying— wait, wait, wait. The underlying components are all the same and then you customize them, right?

**Chris Romeo (21:50)**: So.

**Izar Tarandach (21:51)**: Exactly. So that's the point.

**Chris Romeo (21:52)**: So you would have—

**Izar Tarandach (21:54)**: but I'm saying version. Of Salesforce, that's for you. That's the way it works for you.

**Chris Romeo (21:59)**: What if there's a world where like a startup is providing a product called, uh, ABC, but the— what they're providing to Matt and what they're providing to you are, are entirely different products that solve problems in different ways and have different features, but they're, they're, they're using AI to create derivatives of the core concept of their product that solves your particular need uniquely to you.

**Matt Coles (22:27)**: Oh, I have an even, I have an even more dystopian example, a dystopian thought on that. Why not have the startup just implement an MCP server and make it an API?

**Izar Tarandach (22:38)**: But wait, wait, wait, wait, wait, wait, wait. It's not that far ahead.

**Matt Coles (22:41)**: Look, look at this. 'Cause then you can tell, then you can tell the startup what you want and let them make something for you.

**Izar Tarandach (22:46)**: Go back in time. Go back in time to the depths of a year ago, 2 years ago, no code. Oh, we're going to give you blocks and you just connect them, blah, blah, blah, blah, blah, blah, blah, blah. Right? I don't know where that stuff went, but yeah, whatever.

**Matt Coles (23:01)**: That would have been awesome.

**Chris Romeo (23:02)**: It disappeared.

**Matt Coles (23:02)**: That would have been awesome to, by the way, I think that would have been amazing if it had worked.

**Izar Tarandach (23:07)**: It disappeared from Lord knows where, but if you go into the marketing and finance pits of hell that sometimes we see. They're alive and kicking in there. Joe from marketing loves to connect things with things and get things.

**Chris Romeo (23:20)**: True.

**Izar Tarandach (23:21)**: But nowadays you go and you look at, uh, things like, uh, uh, N8N, uh, what's the name of the other one?

**Matt Coles (23:29)**: Uh, LangChain?

**Izar Tarandach (23:31)**: No, not, not LangChain. It's, it's on top of, like, it, it, it's higher in the—

**Matt Coles (23:35)**: LangGraph?

**Chris Romeo (23:36)**: No, no, no. Landman?

**Izar Tarandach (23:39)**: So it, it's, it's—

**Matt Coles (23:41)**: Chris is dating himself with Landman.

**Izar Tarandach (23:44)**: It, it's basically old, old— It's basically no-node stuff that behind each node has a very clear interface to code and already has AI nodes inside. And I already has AI to create, uh, uh, sequences of nodes, right? So the AI is getting meta in here. The AI is telling the AI how, how to. So my, my point is, and already I forgot where I was going with this. My point is, if you use those things, you end up with different versions, customized versions of sequences of of, uh, uh, operations.

**Matt Coles (24:16)**: Was it Zapier you were looking for?

**Izar Tarandach (24:18)**: With the same blocks that at the end of the day is like a different product built of the same, same pieces.

**Chris Romeo (24:25)**: Yeah.

**Matt Coles (24:25)**: Yeah.

**Izar Tarandach (24:26)**: So we're, we're not that far from there. Yeah. But again, my, my whole point here and where I got, uh, uh, uh, where I got irked with, uh, Matt's example was that it's not like we're moving forward because You can't tell me that you built in 2 hours something that then you say, I used this to learn what I needed, because then it's not like you built a prototype. You, you, you built a, a wisp of an idea that you remotely had that you might need something. You didn't know what it was, but it had to be beige.

**Chris Romeo (24:59)**: Yeah. It's naive to say that just saying something's beige and that's gonna solve the problem.

**Izar Tarandach (25:04)**: Right. Is, I mean, there are processes in here. Organizations work in a certain way. Sure. We want to optimize that, but even to optimize something, you have to understand whatever is it that you are already doing. So I, I think that it walks back to my whole problem with this democratization of development or whatever. Some things should not be democratized.

**Chris Romeo (25:27)**: Some things should be left in the hands of people who actually know what they're doing. Cause I wanna, I'll argue with you on that cause I'll take the other side of that equation. But let's, let's wrap up our conversation on the fact that things are dead. We had another article, Traditional Code Review is Dead.

**Matt Coles (25:42)**: Oh my God. That's going to be—

**Izar Tarandach (25:43)**: Oh, definitely.

**Chris Romeo (25:44)**: What comes next?

**Matt Coles (25:45)**: That's a whole show.

**Chris Romeo (25:48)**: Yeah, but we have to, we have to do this in the context of the Is Dead show.

**Izar Tarandach (25:52)**: Yeah.

**Chris Romeo (25:53)**: Or the Walking Dead of security. So just to summarize, the article argues that traditional manual line-by-line code review is becoming obsolete in modern software development due to the rapid adoption of AI-assisted tools and automated quality gating practices. The idea that engineers manually inspect syntax and logic and pull requests is no longer the primary mechanism for ensuring software quality. I don't want to use computers anymore at this point, if that's the case.

**Matt Coles (26:28)**: Well, also, code review was not just a line checker, right? We had grep for that. We had static analysis tools for that.

**Chris Romeo (26:36)**: True.

**Matt Coles (26:37)**: Right. The, the reason why threat modeling, and sorry, there's a lot of paradigms between, or a commonality between threat modeling and line-by-line code review. It's so that individuals can share understanding about what they just are about to and just built. So just doing the scanning. So basically these, basically this argument is Everything is becoming a scanner, right? If you just scan, absolutely, right? AI, sure, have at it. Use a linter, you know, you call an MCP server to use a linter or whatever, or do your own linting, whatever it is. Line-by-line code review for correctness, sure. Line-by-line code review for Understanding of what we just built? No. Although the code, although the AI can, can, can describe it for you.

**Chris Romeo (27:39)**: But it gets worse than that. The, a little deeper in the article, it's talking about the idea that reviewers verify outcomes. For example, that automated tests pass and behavior in preview environments aligns with expectations. What I hear when I read that is we no longer understand what builds, how, how to measure quality of something that we've built. And without being able to understand that, we have to rely on the things the machine tells us, which means we're very quickly going to get away from building things of quality. Because if we don't understand how to get quality, how can we, well, adjust the program?

**Matt Coles (28:22)**: As somebody who used to do quality engineering, and I won't say it was a major part of my life, but it was a portion. Having developers become QA engineers and understanding how to think about quality as a problem or security or whatever. Uh, I mean, okay. Uh, right. So I'm a big fan of developers working in tech support for a period of time because they learn and understand customer behavior while developers working as QA engineers, but Not being dev and QA, right? Being QA, meaning shift roles, not dual role, because that's DevSecOps or DevOps or whatever, you know, DevStar, because they're taking on multiple responsibilities, doing equal each of them, not as well as they could specialize.

**Chris Romeo (29:16)**: It's a really good definition of that. The whole concept, doing everything poorly.

**Izar Tarandach (29:22)**: Yeah, here's my problem with this whole thing. We are asking a black box that we don't know how it works to transform our systems now in black boxes that we don't know how they were built. And to do that, we are going to do black box testing by putting some input and checking if we get the output that we expected.

**Chris Romeo (29:42)**: But even worse is testing that we don't understand because we probably didn't write the test. So we're probably using AI-generated tests.

**Izar Tarandach (29:50)**: This is like a dumbification of the whole process that it's not only criminal, it's dangerous.

**Chris Romeo (30:01)**: Is this the idiocracy of security?

**Izar Tarandach (30:03)**: This is the idiocracy of not only of security, of code quality. Because what happens now, when this whole thing, when this whole let's use LLMs to write PRs thing started, I wrote a very short piece. Where I said, okay, we have been suffering from alert fatigue for a long time. We just like stamp alerts because we get millions of them and we don't have the time and basically the mind to check all of them. Are we looking into the PR fatigue era now where we are just going to stamp, rubber stamp, whatever comes from the PR creation machine and let it go through because we don't understand that word that's overly used, context of the whole thing. And we are just trusting the LLM does it better. And here we are looking at that.

**Matt Coles (30:56)**: I mean, so you know what happened with, um, with, uh, curl, right? Recently curl got inundated by AI slop, uh, PRs and they shut down their bug bounty as a result.

**Izar Tarandach (31:12)**: So something that used to be a high signal quality input channel of things that we should really pay attention to, because Lord knows that curl moves the internet. Now it's closed because, you know, some people decided that just running an agent that reads code and burps PRs was a nice way to monetize this whole thing.

**Chris Romeo (31:38)**: Are we in that era right now where things that have a public interface that can result in you getting money are gonna be inundated with AI slop no matter what it is?

**Izar Tarandach (31:51)**: For sure. Look at Medium, look at Substack, uh, the amount of AI slop going in there, right? Cuz people get paid for the hits, I, I, I guess. So yeah, nobody reads, so why should I care if what I put in there is bad quality?

**Matt Coles (32:06)**: I mean, there, there is a, there are ways that AI can help make the code review process easier with a caveat because nobody's threat modeled this, I'm sure. But if the AI were doing code explanation, right? So if you, if you had the AI summarize what the code is doing, you could review the summary to understand the flow. And if you did formal verification, you could, where the AI could assist with the verification steps or sorry, producing the evidence for the verification steps, right? That's potentially one option. Lastly, what about test-driven development? I've seen this pop up in a couple of places. If you had test-driven development as a paradigm. Even if the AI wrote the tests, but, but importantly, you need the human in the loop to verify that the tests are the right tests.

**Izar Tarandach (33:08)**: But Matt, that, that's okay. So I saw something called spec-driven agent coding, which is what a human would look at probably as test-driven development, but basically it's write your spec in a markdown, feed it to the, the agent and watch magic happen. Agents.

**Matt Coles (33:29)**: Yeah, that's the Claude Code approach, isn't it?

**Izar Tarandach (33:32)**: I believe it's one of the ways that was suggested that this is the best way of working with one of those agents. But my problem is we can't ignore at some point that hallucinations happen and sometimes they happen badly. And now we are taking a process that, okay. We have created tools and we have reached a level of certainty in this process that enabled us to create this whole new organism, right? And now we are saying, okay, so I'm going to have this thing that, again, I don't quite understand, that may hallucinate at any given point. I'm going to let it read my spec, write the code. Write the tests, run the tests, and I am the human in the loop, and I'm doing scary quotes with my fingers, and because I am the human in the loop, everything is going to be fine because there is a human in the loop. The problem is that at the same time that we introduce these things more and more and more and more, the human in the loop that wasn't, uh, much of a human to begin with, right, is going to start losing touch with what's happening. They're going to start to adjust to all the moving parts.

**Chris Romeo (34:49)**: Are we heading towards a world where developers no longer know how to code? Because that's what I'm hearing when you describe that.

**Izar Tarandach (34:55)**: Yes, we are. Yes, we are. Even worse.

**Chris Romeo (35:03)**: Or classic. They don't know how to, they couldn't, they can describe to the AI how to build what they need, but they couldn't sit there and write it in Java.

**Matt Coles (35:11)**: Yeah, software engineering you're talking about.

**Izar Tarandach (35:12)**: No, but now we are expecting. This next generation to be able to sit down and read that code and critique what the AI did to a point where they say, no, this is wrong. This has to be, this has to be corrected. No, this is an edge case that I can identify, but this automatic mechanism couldn't.

**Matt Coles (35:32)**: So somebody needs a threat model, this new software engineering paradigm where AI, where, where AI is, is breaking the chain because they are causing, and the big story is right, that entry-level developers are are the ones who are going to suffer here, right? Because AI is basically the entry-level developer. So today you have senior engineers who can do the validation as a human in the loop. Tomorrow, 10, 15, 20, 30 years from now, there will be no senior developers because there weren't any junior developers. And who then is the human in the loop?

**Izar Tarandach (36:09)**: It's worse.

**Matt Coles (36:10)**: So it's a delayed reaction.

**Izar Tarandach (36:12)**: It's worse. The senior developer of tomorrow is the AI of today that got trained on everything that all the other AIs did. And because of the volume of things, because of the way that it was trained, because of the compute time that was put in it, because we write amazing transformers and whatnot, this yesterday's junior AI developer is going to be tomorrow's senior AI developer. And I don't know, are we going to end up with a cast of developers that actually know how to write AIs, or are we going to have AIs that write AIs? I don't know.

**Matt Coles (36:49)**: Are you going to retrain AI using AI-generated code?

**Izar Tarandach (36:54)**: I'm not going to do anything because I'm going to tell an AI to do it. But no, really, from what I've been reading, and please, I could be totally wrong here, and I'm not an AI expert by any means or whatever. I think that it's been understood that AI trained on AI output collapses at some point. I should go and do more reading about this.

**Chris Romeo (37:18)**: It seems like it's an entropy type of thing. It's going to get, it's going to get, it's going to, it's going to break down until there's nothing left.

**Izar Tarandach (37:24)**: Yeah. But in, in my head, I don't understand how the fact that something is created by AI ingests itself into the process of training to a point where the result collapses. So that's what I have to educate myself on. But, uh, I, I, I'm just very afraid of, uh, uh, uh, um, Middle Ages case where we know that back in antiquity we, we knew a lot of stuff and the Alexandria Library burned down and these and that and the other one. And we ended up with, I don't know how many years of Middle Ages where a lot of knowledge got lost.

**Matt Coles (38:04)**: About 500 years, I think.

**Izar Tarandach (38:05)**: Yeah. A lot of knowledge was like copied from one hand to another, but the emphasis was not in the knowledge. It was on the illustrations that went around the page. That, that's what gave them value, not what was written in them. Right. So that, that's my biggest fear today. We are heading towards a middle age and like any good catastrophic movie, it begins with everything being dead and becoming a zombie. So, you know, that to me closes the, the, the loop. We are killing everything, we are creating zombies, and now in 2 hours we are rewriting the OS for the world. Somebody already wrote a full operating system with, uh, uh, vibe coding, from what I understand.

**Matt Coles (38:49)**: If we could, if we could figure out how to make them zombie processes, that would be, uh, we'd be would be perfect, right? So just start them, kill the terminal, start them, kill the terminal, move away.

**Chris Romeo (38:59)**: Yeah. So, all right, well, I think what we've concluded is that in fact everything is dead, but also nothing is dead.

**Izar Tarandach (39:08)**: Exactly.

**Chris Romeo (39:10)**: And people who say things are dead in articles are really just trying to get people excited. Clickbait.

**Izar Tarandach (39:16)**: Yeah, but Dask, that's, that's dead.

**Chris Romeo (39:19)**: Dask is dead. I mean, I'll—

**Matt Coles (39:20)**: but that's just dead.

**Izar Tarandach (39:21)**: Like, that one is dead. That one is dead.

**Chris Romeo (39:22)**: Wait, wait, wait, what about if you wanted to have Dask but you knew Dask was Instead, you ask the AI to resurrect it and create a new, better version of it.

**Izar Tarandach (39:31)**: It's going to scan only 80% of the surface.

**Chris Romeo (39:35)**: That would make it probably better.

**Matt Coles (39:36)**: That might be better. That might be improvement.

**Chris Romeo (39:40)**: I'm getting ready to start trashing the whole industry again. I just got them all back in my good graces, and then I get dragged back in. Well, folks, thank you for listening to the original Security Table Podcast, not to be confused with other things called Security Table Podcast. Um, but for, uh, Isar and Matt, I'm Chris Romeo and, uh, get out there and say stuff's alive.

**Izar Tarandach (40:07)**: Please.

**Chris Romeo (40:09)**: Yep.

**Izar Tarandach (40:09)**: Make stuff alive.
