46 minSeason 3, episode 20

The Roller Coaster of Risk: A Threat Modeler's Perspective

With Chris Romeo, Izar Tarandach, and Matt Coles

Threat modeling & security architecture

What do roller coasters and threat modeling have in common? More than you'd think.

Episode title image: The Roller Coaster of Risk: A Threat Modeler's Perspective

Listen to the conversation

About this episode

What do roller coasters and threat modeling have in common? More than you'd think. In this episode, we explore how security professionals view risk differently than everyone else—and why that matters. From roller coaster anxiety to the ethics of identifying danger, we dive into the unique mindset that comes with being a threat modeler. Because once you learn to see threats everywhere, there's no going back.

Follow AI Security Table:

➜ Home: https://securitytable.ai/

➜ X: https://x.com/AISecurityTable

➜ LinkedIn: https://www.linkedin.com/company/ai-security-table/

➜ YouTube: https://www.youtube.com/@AISecurityTable

Transcript · 46 min conversation

0:10Chris Romeo Hey folks, welcome to a special episode of the Security Table coming from Matt's bedroom.

0:16Izar Tarandach Perfekt.

0:16Chris Romeo As you can see with the purple walls, if you're watching this thing on YouTube, Matt has a bit of a purple haze going on in his bedroom. We'll let you For the younger generations out there, we'll let you Google that, figure out what that means.

0:31Matt Coles Why am I hearing the Watchtower, dude?

0:34Izar Tarandach Why am I hearing Prince in the background? Why, Matt?

0:40Chris Romeo Have you ever dressed up as Prince for Halloween?

0:42Matt Coles No.

0:43Chris Romeo Oh God, he's already got a visual effect.

0:47Izar Tarandach Oh God, now I got a migraine.

0:50Chris Romeo I have not either. I've, I've never been a big Prince fan. I mean, I know Prince's anthology of music, just never been a giant fan that I would be like, hey, put some Prince on.

1:04Matt Coles Sure, it's, it's not that kind of music. It's not, he's got some kind of music, at least.

1:12Izar Tarandach I mean, anybody who changes their name for a symbol, I'll give them a—

1:16Chris Romeo but then changes it back, so it's like, come on.

1:19Izar Tarandach Oh, he changed it back?

1:20Matt Coles Madonna did the same thing. So, you know, it's—

1:22Chris Romeo she did.

1:23Izar Tarandach What was the symbol?

1:24Chris Romeo I don't remember that one.

1:27Izar Tarandach Two forward-pointing cones. What, what, what the—

1:31Chris Romeo I was having a conversation with somebody a couple days ago, and they're like— somehow we were talking about Madonna. They're like, Madonna's dead. I'm like, no, Madonna's not dead.

1:39Matt Coles Really?

1:40Chris Romeo And I had to double-check. No, no, she's not, she's not. My friend had bad information.

1:45Matt Coles Of course.

1:45Chris Romeo I'm like, what? No, no, not dead. Again, I think it happened in real life. It could have been back to that Purple Haze thing, kind of tying the whole conversation together. No, she's just a material girl living in a material world. Wasn't that Cyndi Lauper?

2:01Matt Coles Folk, folk, folk.

2:03Izar Tarandach Is it?

2:04Chris Romeo No, I just said it because it's so easy to confuse. There's an illustration about threat modeling there somewhere. It's so easy to confuse.

2:11Matt Coles We went from '60s to '80s. To '70s.

2:18Chris Romeo Let's hit the '90s.

2:18Izar Tarandach Wait, wait, wait, you looking for consistency here?

2:22Chris Romeo Yeah, you were looking at 100+ episodes, there's no consistency. What about Milli Vanilli?

2:28Izar Tarandach Ooh, that was a dig, archaeological dig.

2:33Matt Coles Well, at least it wasn't Ice Ice Baby.

2:36Chris Romeo Oh, come on, that's one of the best songs of the '90s. Name a better song of the '90s than Ice Ice Baby. It's more iconic.

2:42Izar Tarandach Uh, the '90s, the one— I don't think I was listening to songs in the '90s.

2:45Chris Romeo One with Marky Mark and the Funky Bunch? Uh, Good Times.

2:51Matt Coles Is that the name of it?

2:52Chris Romeo Yeah, of course it is. I'm a '90s anthology myself.

2:56Izar Tarandach I have to say, music stopped in the '80s, okay? Everything after that is a remix.

3:04Chris Romeo I mean, everything's a remix in life.

3:08Matt Coles All right, come on. So if you're going to talk about this, we have to talk about Run-DMC and Walk This Way. When they did it live with Aerosmith.

3:16Chris Romeo I mean, that was the first really crossover between two major genres of music that really had ever—

3:26Izar Tarandach No, a real, a real crossover would be the Monty Python Funny Walk sketch with Walk This Way in the background. Okay, that's a mix.

3:40Chris Romeo That's a very, very eclectic mix of—

3:43Izar Tarandach there you go—

3:43Chris Romeo comedy and music, which is a good thing. So yeah, I rode the Aerosmith roller coaster the last couple of weeks. Well, I didn't ride it for weeks, but I don't know if you're aware of this at Mickey World in Florida, there's this rock and roll coaster.

4:01Izar Tarandach Yep.

4:02Chris Romeo And it goes upside down twice. So, and so you'll You may get a chuckle out of this. So I'm standing in line waiting to go on this thinking, have I made a bad decision for my future, for my life? Yes.

4:16Matt Coles Yes, you have.

4:17Chris Romeo And then I look over and there are 7-year-old girls in the line. And so I'm thinking to myself, if I cannot stay in this line yet a 7-year-old girl can ride this ride, then that doesn't say a whole lot about me.

4:32Izar Tarandach On the other hand, 7-year-olds are durable.

4:35Matt Coles You're, you're, you're not.

4:36Izar Tarandach that's one.

4:38Chris Romeo But my pride was then—

4:39Izar Tarandach On the other hand, you got more experience in life than them, and you know what could go wrong.

4:46Chris Romeo So true, it could threat model the whole time, right?

4:48Izar Tarandach It actually goes to a very, very important point that I think is not made enough. As threat modelers, we are taking one for humanity constantly, right? Once you're a threat modeler, you never look at life the same way.

5:07Chris Romeo Like, that's true.

5:08Izar Tarandach We have— we avoid doing a, a whole number of things because we know what could go wrong so that you people can keep doing them. So you're welcome.

5:17Chris Romeo Well, I still rode the roller coaster. Nothing went wrong. I survived.

5:21Matt Coles Hold on, so what you're saying is you don't need domain expertise in anything if you have threat modeling experience?

5:29Chris Romeo Yes.

5:31Izar Tarandach And by the way, next Veterans Day, hug a truck modeler, because we are here on the line for you every single day. We know what could go wrong and we keep doing it.

5:42Chris Romeo So wait, no domain expertise needed. You can just BS your way through anything that you get put in front of.

5:50Izar Tarandach You can't point at anything and say that could go wrong. That's kind of the way it is. Basically just Muppet Show your, your, your, through your life.

6:00Matt Coles All right. So the big question, Chris, obviously you survived the roller coaster. Did the 7-year-olds?

6:05Chris Romeo Oh yeah. They were, they were, they were the ones with their arms in the air going like this. I'm holding on for dear life going, hmm, this could be it. This could be it. Oh, survived the next one. Come on. Keep that. It was a lot.

6:16Izar Tarandach Are you sure they were the same 7-year-olds that went in?

6:21Matt Coles They were doppelgangers. Came out.

6:23Izar Tarandach They are replaceable, you know.

6:26Chris Romeo No, I'm pretty sure, pretty sure they, uh, they survived and probably had more fun than I did.

6:33Izar Tarandach Wait, wait, wait, wait, no. Imagine yourself back then, you're going up the first climb, the things that are going through your head, right? You are questioning constantly, what did I do to get here?

6:46Matt Coles And what went wrong? Leading up to this.

6:49Chris Romeo Exactly.

6:50Izar Tarandach And, and, and then I think about that fella that created the— what was it? The roller— the suicide roller coaster. The one that actually will kill you when it goes down. Have you read this?

7:01Chris Romeo No.

7:02Izar Tarandach Some, some, some fellow actually like planned, projected one where the, the fall is like at such an angle and such a speed and whatnot that there's no way that you're going to reach the bottom alive.

7:14Chris Romeo Wow.

7:15Izar Tarandach And I keep thinking to myself, what goes through somebody's head as they're doing that climb?

7:23Chris Romeo Yeah, I got a 50-50 chance of making it.

7:26Matt Coles Yeah, it's a fit to purpose, you know. What could go wrong? Well, they could survive everything.

7:33Izar Tarandach Actually, no, because it's designed that way, so it would actually— if it works as designed, nothing would go wrong. So Works in their favor.

7:43Chris Romeo We're going to need to bring morality into the threat modeling, the fifth question.

7:46Matt Coles You mean ethics?

7:48Izar Tarandach Ethics and morality. That's a whole time. We have to get to that one at some point.

7:52Chris Romeo We brushed on it a few episodes ago.

7:56Izar Tarandach Indeed. Indeed.

7:57Chris Romeo Just don't remember what the conclusion was, but I could say that for a lot of episodes.

8:02Matt Coles We sort of go off track, which is something that could go wrong on a roller coaster.

8:07Izar Tarandach I'll put a call to the guy in Australia and ask what was it.

8:11Chris Romeo That's true. Shout out to our fan in Australia. Thanks for tuning in.

8:17Izar Tarandach How's it down under?

8:19Chris Romeo If we ever have a call-in line going out to Australia, if it happens, we know it got there. Takes another day, an extra day though, to get there. But okay, so I think we've kind of made our way through almost 10 minutes of an episode. Izar, you had what you described, you self-described as an epiphany.

8:48Matt Coles You had a Delphi moment.

8:50Izar Tarandach A shock.

8:50Chris Romeo Delphi, like the car company that makes car parts?

8:54Matt Coles Or the programming language?

8:57Chris Romeo Never used Delphi because it was Visual Pascal. I was a Visual Basic guy. I like Visual Basic better. I don't know why.

9:04Matt Coles Everyone does.

9:06Izar Tarandach Nobody does.

9:08Matt Coles Better than Pascal?

9:11Chris Romeo Or Visual Pascal, even better.

9:14Izar Tarandach I had to watch people doing Visual C and C++. Okay, so no.

9:20Matt Coles All right. As a C and C++ programmer, visual? It was all visual. If you couldn't see, how were you going to program?

9:33Izar Tarandach The only place that Visual has in C and C++ is by writing vi before the file name. That's it. That's all the vi that you need for C++.

9:45Matt Coles Not ed?

9:47Izar Tarandach No, no, no, no, no. cat. cat to the terminal, and then you correct with sed.

9:53Chris Romeo That's some real gangster command line right there. If you're using sed to edit concatenated file, and you're doing it all with like complex regexes on a 24x40 terminal.

10:13Matt Coles Green screen.

10:14Chris Romeo I'm blinking. It's flashing too the whole time.

10:18Izar Tarandach Yes, that's the only way.

10:20Chris Romeo Seeing every other sign of the characters. So, all right, well, Izar, what's your epiphany?

10:26Izar Tarandach So the epiphany, epiphany. I was reflecting as one does coming back from conferences, and that's that kind of stuff. And I think that threat modeling got to a state of— what was the word, Matt? I always get confused with ending E's.

10:48Matt Coles I'm not going to be able to help you on this. I don't know what your epiphany—

10:51Izar Tarandach you do, you do. The one from the shoes with the switch.

10:55Chris Romeo Oh, Nike.

10:57Izar Tarandach Nike. So I think that threat modeling got to a Nike moment. Just do it. Because I went to the Oracle. That's how we got to Delphi.

11:10Matt Coles If people are following this, just making sure you didn't go to Larry Ellison's Oracle, I hope.

11:16Izar Tarandach No, no, no, no, no, no, no. That one, we know how it ends.

11:19Chris Romeo You didn't travel to Delphi, right? Nope.

11:22Izar Tarandach Nope. Can't find it on the map.

11:24Chris Romeo That was a city, wasn't it? I think that's where the Oracle was. The Oracle of Delphi. Yeah. Okay.

11:29Izar Tarandach No, just checking.

11:29Chris Romeo I'm, I'm not a history, yeah, his flight got delayed, redirected, you know, and I freaking use ChatGPT.

11:36Izar Tarandach Continuing.

11:38Matt Coles Well, I asked, oh, pause, pause. Did you just call ChatGPT the Oracle?

11:46Izar Tarandach It's what everybody else seems to find their, their, their answers nowadays. And most of the time they come back as clues and enigmas and that kind of thing, and then people just impart whatever meaning they want to what comes back. So I did it to mine, and I asked it, at this stage of the state of the art in threat modeling, what are the 3 top challenges to be solved, and what are the probabilities that these are indeed each of the top 3? So what I got was When I say top 3, I mean if you ask a broad mix of practitioners, researchers, and tool vendors today, what are the 3 problems that most constrain the impact of threat modeling in the real world? Things that we deal with every day.

12:33Chris Romeo Okay.

12:34Izar Tarandach First one came: making threat modeling truly continuous and scalable in modern SDLC. Second one was usability, skills gap, and democratization. And the third one was reliable automation and AI assistance, especially with LLMs. The first two got more than 80% and the third one got 70%.

13:01Matt Coles Hmm.

13:04Izar Tarandach So why do I bring this back to Nikkei? Because in my gut, I think that The world has been given enough methodologies and guides and books, like, to cover the whole gamut of where you need to apply this. You can do threat modeling once in a lifetime, you can do threat modeling with, uh, Italian dishes, you can do it continuously with every story and every sprint. You can choose to do whatever you want, right? And on the usability, you can do it with catalogs, you can do it by knowing what you know, you can do by asking people, you can do by imagining stuff, you can do by different layers of cake that has nothing to do one flavor with the other. You can do it in so many ways that today to come and say, no, we have a problem with usability, is to come and say, okay, this whole body of knowledge here, this whole body of experience here, let's ignore it for a second and say, no, no, no, I still can't use this thing. I'm special. I'm a flower. I can't. I have to do it a different way. I have to do a special way. So at this point, I'm just saying, just do it.

14:23Chris Romeo I'm trying to figure out what this voice is like. There's two voices and I don't—

14:28Matt Coles uh, okay, he used Sora to, to set it to some lilting, uh, sound.

14:33Izar Tarandach Only two? Because in here there's way more. But anyway, So at this point I'm asking myself if these are the top challenges today.

14:44Matt Coles Well, so interestingly, I asked a different LLM the same question just now.

14:50Izar Tarandach My LLM will battle your LLM.

14:53Matt Coles Gemini will crush ChatGPT.

14:56Izar Tarandach Oh yeah. Very special LLM.

15:00Chris Romeo Wow, a lot of smack talk in here. This is like geek smack talking now.

15:03Matt Coles So, so I agree with two of yours. And the skills gap and the skillability problem.

15:11Izar Tarandach Wait, wait, wait, wait. Look at the top and see if the first sentence is, what did ChatGPT say?

15:16Chris Romeo Okay, go.

15:18Matt Coles I'll ask it what it thinks ChatGPT said. Just a moment. So, and actually if it asks ChatGPT for real, I'll be really— So, the skills gap at 95%. Scalability at 85% and conflicting with Agile and DevOps speed at 90%. Hmm. So roadblocks in the development process is causing roadblocks in threat modeling. Hmm.

15:53Izar Tarandach Hmm. No, that's usability again.

15:56Matt Coles Well, no, it's not.

15:58Izar Tarandach Why not?

15:59Matt Coles It's perception.

16:01Izar Tarandach Of what?

16:02Matt Coles Of this being a problem. Things that slow me down.

16:07Izar Tarandach That's like being in a completely out there example. It's like being in line for a roller coaster and seeing 7-year-olds and saying, this is a problem. Like, I mean, everybody's going to get to the roller coaster one way or another.

16:26Matt Coles You, you misunderstand the consternation that Chris was having, although Chris is being really quiet about it right now.

16:32Izar Tarandach He's reviving the moment.

16:34Chris Romeo I'm riding the roller coaster.

16:37Izar Tarandach He's asking Claude.

16:39Chris Romeo I don't ask LLMs. I am an LLM.

16:43Izar Tarandach Untrained, but an LLM. Yeah, 100%.

16:47Chris Romeo But I have a special power. I'm full of opinions.

16:50Matt Coles And is our, can you, uh, can you reveal your shirt to all there?

16:58Chris Romeo Just because you passed the Turing test doesn't mean you're smart.

17:06Izar Tarandach Okay, so what do we do with this? Like, the three of us together should be able to point at some other challenges that the LLMs didn't get, but I'm having a problem here.

17:20Chris Romeo Yeah, I mean, scalability is certainly the big one.

17:25Izar Tarandach But again, scalability, just do it. Just do it.

17:28Chris Romeo Just do it. Just do it doesn't work in the real world.

17:32Izar Tarandach Do it.

17:32Chris Romeo Tell people.

17:33Izar Tarandach Find what didn't work, what doesn't scale in your place, and change that. I know I'm being oversimplistic here, but if you don't do it, you don't figure out those things.

17:43Chris Romeo Yeah. I mean, you're discounting organizational friction. Which means you can't just go out and change it.

17:52Izar Tarandach HR told me to always discount organizational friction, but, but no, no, seriously, seriously, if you find that organizational friction is the thing holding you back, then you have to find what is it that you need to oil so that that friction disappears or goes down, right?

18:13Chris Romeo Sure.

18:13Izar Tarandach But if you don't do it and you don't identify that that's your problem, that that's the wheel that you have to oil.

18:17Chris Romeo I'm just saying, I'm just saying.

18:19Izar Tarandach You're not going anywhere.

18:20Chris Romeo The outcome is not as easy as you're making it sound.

18:24Izar Tarandach No, no, no, no, no. I'm not saying, I'm not saying that it is. No, I'm not saying that everything has been solved. What I'm saying is we have enough on the table, right? Think Apollo 13. We are in a room looking at the table saying what we have in the capsule that the astronauts can use. We have enough on the table that we can put together something that works. So to keep saying, I'm not going to do it because the perfect thing hasn't come yet. I'm not going to do it because the perfect tool hasn't come yet. I'm not going to do it because the perfect methodology hasn't come yet.

18:56Chris Romeo I mean, I haven't heard any of those excuses for why people wouldn't, can't do threat modeling.

19:01Izar Tarandach I have. I have heard enough of that will not work here because.

19:07Matt Coles I've heard that, I've heard that part, but not for the reasons you mentioned. Maybe consistency, like it's not going to work here because if we're not consistent, we're going to have different, you know, difference of different views of this thing, and it's not going to be practical for us, especially if people change roles, right? So if you have developers who are working on different projects, if you don't have consistency of practice within your organization, you're gonna get— you're gonna have challenges adopting threat modeling. I've heard that, but not that, uh, not— if somebody, if somebody's arguing that there isn't a perfect tool, they're not interested in doing threat modeling, right? That's a non—

19:53Chris Romeo it's an excuse, right?

19:55Matt Coles It's a non-argument.

19:56Chris Romeo Yeah, it's a way to make you go away.

19:59Izar Tarandach Actually, it's an excuse is a very, it's a very apt comment. I think that a lot of what's happening out there is people making excuses, especially because we have all these things in the table. So what I'm saying is we are at a point where you stop making excuses, you pick something up from the table and you do it and you see if it works or not. That's a big challenge now for me, getting people to pick stuff from the table and actually do it.

20:24Matt Coles So try and trial and error. Is becoming a problem, I think, right? We have people under a lot of pressure to get stuff out the door. There's a lot of pressure to use AI. There's a lot of pressure to just get— remove all the roadblocks. And so trial and error is itself a roadblock. If they have to trial and error something to figure out what works, they're engineering their process, not engineering their product. Yes, and I don't see a problem with engineering the process. That's— but that's our view, but that's not their view often, right? Developers don't want to spend time re-architecting their process. They want to spend time architecting a thing.

21:12Izar Tarandach Mm-hmm. That's the thing. So bringing back to a different t-shirt that I usually have. I have this one that says the process is bigger than outcome, right? Because without a process, you can reach an outcome, but you won't be able to reach it consistently because you don't have a process to tweak or a process to follow. Like, you can bake a cake from whatever you find in the cupboard, and it may even come out good, but tomorrow you won't be able to make the same cupcake because you don't have a process to get there. So the process is more important than the outcome, in my view at least, and in the T-shirt's view, and apparently the US Navy SEALs.

21:56Matt Coles But, but, uh, but if you, if you have a functioning process, why screw with it? Right? Why?

22:04Izar Tarandach Because no, if it's functioning, if it's functioning, yes. But if you're going to add something more like threat modeling to it, then you are by definition already screwing with it. So you might as well tweak, tweak it. Until it works again. And probably the outcome is going to be better than it was before. If your outcome is already perfect, hey, don't touch it. Go for it.

22:26Matt Coles Well, so here's another snag in all this. Roles and responsibilities. Whose responsibility is it to screw with the engineering process? And, sorry, to tweak the engineering process?

22:38Izar Tarandach Oh, sorry, because for a moment there I thought that you were talking about me, because usually I'm the one screwing the engineering process.

22:44Chris Romeo At least that's what people say at the meeting, at the conclusion meetings.

22:48Izar Tarandach So that's my responsibility.

22:49Matt Coles Retrospective. Isar screwed with the process. Who let him in here?

22:54Izar Tarandach Okay, that's a t-shirt. But no, the point is retrospectives are great. Okay. Because they usually, some of them, at least most of them. To me end up in, if only we had threat model that day. But, uh, then people come in and say, yeah, we could have, but it would have been, what's it that they say, big lift in the beginning.

23:25Matt Coles Heavy lift.

23:26Izar Tarandach A heavy lift in the beginning.

23:28Chris Romeo That's the new excuse.

23:31Izar Tarandach No, and you know what? Yes, we have this box here on the table. That says heavy lift in the beginning, but we have this other box here on the table that says a bit every time. So you don't like that box, take that one.

23:46Chris Romeo Yeah, back to your point that there's multiple options.

23:49Izar Tarandach Exactly.

23:50Chris Romeo To do this thing. Now here's the— let me ask you this. Let me ask you this question. See if this resonates with you guys. Is part of the problem with enterprise threat modeling adoption that we don't have enough Maverick-style leaders that are pushing the envelope and trying to change things?

24:13Izar Tarandach What's a Maverick? No, sorry, leader of what?

24:17Chris Romeo So somebody who's on the security team trying to press threat modeling forward, that's people pushing everybody's buttons to try to cause change.

24:27Izar Tarandach That is a horrible place.

24:29Matt Coles Yeah, go ahead. You first.

24:31Izar Tarandach No, I was going to say that that's a horrible place to be pushing the buttons because people in the security team know that that needs to happen.

24:42Chris Romeo No, I meant like there is somebody who's outwardly trying to move threat modeling forward across the entire organization.

24:48Izar Tarandach We are the guys who sit back and say, well, used to be, well, sometimes still, that sit back and say, if only they would do the things that I'm asking them to do. Okay, we've been telling people to do things, asking them to do things for X amount of years. Here we are. I would say, how come we don't have a maverick enough leader to look at the security team and say, those guys have been repeating the same thing for the last couple of years, perhaps it's time that I listen to them and do something about it?

25:18Chris Romeo Yeah, see, like, my approach in the enterprise was always to just start getting people doing it. I didn't go to their leadership structure and go, we have this great idea, we need you to— no, screw you, we're gonna, we're gonna do this from the grassroots until it gets to the point where it's like it starts to catch on and then it's got a life of its own and you can't really squash it because it's, it's got enough of a foundation underneath it.

25:41Izar Tarandach The problem nowadays I find is that there is substance in the universe that's more expensive than nobody unobnitonium, and that is developer time. Screw developer time and bad things happen to you. But now perhaps that everybody's using LLM, we'll get some— exactly. The black helicopters come by.

26:06Chris Romeo You gotta wear Teflon to work. That's all I'm saying.

26:09Izar Tarandach But the thing is, perhaps now with LLMs, now that everybody is vibe coding their way through life, developers will finally have some more time to do the stuff.

26:17Chris Romeo They got lots of time. What are we talking about? They got half a day free now.

26:24Matt Coles Well, so here's— so the alternative is we're talking to the wrong people. We should be talking to the agents, not to the developers.

26:31Izar Tarandach I started talking to the agents. I did. No, no, no, no, I did.

26:34Chris Romeo That's how we got here.

26:36Izar Tarandach No, no, no, no, no, no, no, that was the Oracle. But I started proposing to my developers that want to use agents that Could we just play with that agents.md file in your directory before we start doing stuff, before we start the planning stage and things? And some good stuff happens there, but it's not threat modeling per se, because what's going to happen there is going to be plan to code. So you're going to end up with code that is I'm not going to say marginally because I don't have numbers, but it might be more secure than code without the agents.md file. But still, it's not direct modeling. The, the thing, oh my God, it's even worse now that the design is emerging and it's not even emerging from somebody's head. It's emerging from somebody's agent's thing.

27:35Chris Romeo That's just a way of saying you don't have a design.

27:37Matt Coles Right, you're letting it evolve through the Agile process. Again, that's the— that goes back to the Agile development process and threat modeling are mostly incompatible, and not because of process, but because of the way that systems get implemented. As implementation is the design.

28:05Izar Tarandach On the other hand, because of the way these agents work, I'm finding that many times you end up with similar designs to different problems because they're taking their input from the body of knowledge out there, right?

28:20Matt Coles Yep.

28:20Izar Tarandach So what could be more standardized than, than that? And including the fact that they don't have creativity built in, you end up with cookie-cutter copies of systems where just the data fields names and the web UI fields names change.

28:37Matt Coles So what you're really saying is if you, if you poison, I mean modify the training data, you can get your pre-threat modeled components in and basically give them paved roads to work with that are already pre-approved and pre-secure.

28:54Izar Tarandach So the way that we were talking about it in 2000 and nothing, 2010, about the security toolbox where we had the predefined Legos that are secure. You're saying that now we should feed this thing not to the developers, but to the LLMs that the developers are using. And therefore the LLMs, okay, close everything down, we solve the problem.

29:17Matt Coles And then the developers don't have a choice. I want this thing. This thing comes out pre-baked. Now, obviously that's not going to happen.

29:25Izar Tarandach Why?

29:26Matt Coles Well, because somebody will screw it.

29:28Izar Tarandach Somebody will screw it up. Yeah, no, that's for sure.

29:30Matt Coles They'll put it in different orders and it won't work the way it's supposed to.

29:34Izar Tarandach Might even be me. But in theory, there's no reason why it shouldn't work. Like, take, build enough of a corpus of well-designed systems and use that to fine-tune an LLM. In theory, that should ground the thing enough so that you get good designs out of it.

29:56Chris Romeo Yeah. But you know what the biggest catch is? Where you, where are you gonna find well-designed systems from?

30:02Izar Tarandach That's, that's the one. And the second is, again, you're limiting creativity because you are basically pointing at a catalog of these are the choices that you have.

30:11Chris Romeo Well, I mean, LLMs have never had an original thought.

30:14Izar Tarandach Yep.

30:15Matt Coles And with the, and with the push towards, you know, validated components and things like that. We're going to that catalog eventually. The problem is the individual building blocks may be secure. It's the integrations where the new security problems enter, right? It's how you take these building blocks and stitch them together in these various ways that you're going to get the problems that you don't know about yet.

30:42Izar Tarandach So perhaps we have to go the other way around. Instead of training them to feed us answers that they have seen before and know are secure, we should train them to recognize solutions that they haven't seen before and therefore might not be secure. And that's where the person steps up and says, oh, interesting problem, let's look at it.

31:07Chris Romeo Hmm, may have solved something. I don't really understand what it is, but we may have solved something.

31:13Matt Coles We also, we also aren't doing threat modeling at this point. We're now, we're now doing solution design.

31:19Chris Romeo Yeah.

31:19Izar Tarandach Let's see. This is not bad at all. Plus I heard it pays much more.

31:25Chris Romeo These are just changed careers live on the air.

31:30Izar Tarandach No, threat modeling.

31:32Matt Coles Threat modeling. Just do it while the threat modelers are like, go do something else.

31:39Izar Tarandach No, but wait, back in the manifesto, we said do threat modeling over talking about it, right? So I think that I'm going back to that and saying, stop talking about it and pick the pieces from the table and do it.

31:53Chris Romeo Sounds so simple.

31:55Izar Tarandach It sounds so simple, but if you don't do it, then how do you know how complicated it is? That's why I keep— because, you know, I was thinking about back about TMC and the presentation that we saw there. We saw really good stuff, really nice stuff.

32:12Matt Coles ThreatModCon.

32:14Izar Tarandach Yeah, but did we see anything new? Did we see anything that changes significantly the way things are done?

32:22Matt Coles Yes.

32:23Izar Tarandach Or that puts a new problem? What?

32:25Matt Coles Okay, so we saw a new— we saw a method of gamification that helps, that is a way for folks to get more— to be— for threat modeling to be more approachable.

32:35Izar Tarandach Yeah. And I loved it. I really thought that it was really well thought, it was super creative. It was super fun to watch. And I go back to my college days and I think, if I couldn't get 5 geeks back then to play Dungeons& Dragons, what makes me think that now I'm going to manage to get 20 developers, which have the unobtainium developer time in front of their faces, to stop what they're doing to play Dungeons& Dragons so that eventually they may end up thinking that threat modeling is more approachable.

33:11Matt Coles You can't solve everything, but you asked, you asked if there was anything innovative and that was—

33:17Izar Tarandach Yeah, yeah, yeah. No, no, that is innovative by itself, but in the scope of the things that ChatGPT spewed out as the top 3 problems.

33:30Matt Coles By the way, can you refresh my memory? What were the top 3? Because I did ask Gemini what it thought. Scalability, usability.

33:35Chris Romeo Scalability, usability.

33:38Izar Tarandach Security and the use of LLMs. The third one is just trying to look for a job.

33:45Matt Coles So Gemini said, I bet, I bet that, I'm willing to bet that the challenge you mentioned was something like lack of organizational buy-in.

33:56Izar Tarandach Wait, wait, wait, let me see if I have it here.

33:58Matt Coles I'm going to tell it it was a self-flag for using LLMs.

34:02Izar Tarandach Wait, let me see if I have it here. Well, it did say about democratization, right? It says, um, poor onboarding. New engineers and contributors, especially in OSS, rarely get a simple actionable way to participate in TM. Okay, let's talk threat modeling in open source for a second, right? How many open source projects we know that the developers approach with a threat model in mind? And the reason why I'm bringing it out is Open source people usually do it for love. They, they, they love the projects that they are working on. They love the things that they are solving. So these people are doing it with a different motivator and they definitely have the motivation to make it as secure as they can.

34:49Matt Coles Mm-hmm.

34:49Chris Romeo Mm-hmm.

34:49Izar Tarandach How many projects have we seen that start with a good, okay, people, let's sit down and do a threat model?

34:55Chris Romeo Uh, only the ones where a security company sponsored the threat model and performed it for them. Trailer.

35:04Izar Tarandach And even those, even those, I'm going to go ahead and say they weren't there at the beginning of the project.

35:10Chris Romeo All right, you guys want me to put you on the spot completely, potentially embarrass you on a live podcast?

35:16Matt Coles You're going to do that, aren't you? Using our own dog food. Yes.

35:19Chris Romeo So when you started working on PyTM, did you first sit down and create a design and create a threat model for it?

35:29Izar Tarandach If you go back to the first README, of PyTM, one of the first things that you're going to find in there was, yes, we use eval. Yes, we know it's bad. And yes, here it is why we decided to use it. So there.

35:46Chris Romeo Now, is that a threat model or is that just one potential issue?

35:50Izar Tarandach It was the one threat that was identified back then. It's a script, it's a library. It's not reaching out to anywhere. It's not storing anything.

36:00Matt Coles Sensitive.

36:01Izar Tarandach That was the one problem that we found. We drew the vector, we said this is how this could be attacked, this is why we don't think that it's an actual issue. Screw you.

36:14Chris Romeo Mr. Coles, do you agree with the testimony from Mr. Terendosch?

36:19Izar Tarandach I have no recollection.

36:20Matt Coles I, I, I, uh, I'm debating between that and we paid attention to the Python guide on using eval and copied its text a bit.

36:31Izar Tarandach There you go.

36:32Matt Coles That's okay. Yes, yes, we threat modeled the hell out of it.

36:38Chris Romeo You know you're under oath, Mr. Kohls.

36:43Izar Tarandach But wait, wait, I'm gonna go back and I'm going to say that later on, later on, we may have made a mistake in there in sort of ignoring that when we Okay, so for those who are not familiar with PyTM, so you have a library of threats. Each one of those has a condition, and these live in a JSON file, and the condition gets evaluated in the Python code by an eval, and that creates a true or false about the threat existing or not existing. So later on in life, we said, oh, you know what? People want to have their own threat libraries and be able to load those. And we developed the, the capability, and that basically moves the, the conditions from the protected place where only root can write to that file where it's installed publicly for everybody's use to people's own home directories where they can define their own JSON files and load those. But we never checked the code in. So, the capability exists, but it's theoretical.

37:59Chris Romeo So, yeah, I mean, you were, but you were, before I started this line of questioning, you were going down the road of OSS and threat modeling.

38:08Matt Coles Yep. Right.

38:09Chris Romeo And so, do you think they, so we're agreeing that most people don't, this is not something that's embedded in the OSS mindset.

38:16Matt Coles No, it depends on their view. You, right? So I know from looking at crypto libraries and some communication libraries, you know, that they do heavily rely on threat modeling.

38:31Izar Tarandach SSL.

38:32Matt Coles I'm thinking curl, for instance, right?

38:34Izar Tarandach So, oh yeah, yeah, they're strong.

38:38Matt Coles So certain projects already have a security mindset. And almost certainly if they're not documenting their threat model, they certainly know what their threat model is and they rely on that information when they produce, when they, when they respond to or find issues, right? So if somebody submits a PR, there's a basis for making that decision about whether it's a problem or not. Yep. But the lion's share of open source projects almost certainly do not know how they're going to be— they do not know how they would use it or how others would use it such that they can build a threat model, and/or they don't know how to do threat modeling in the first place.

39:23Izar Tarandach So perhaps something we could do at the OWASP Threat Modeling Project is to offer open source projects a pool of threat modelers that they could engage to create a threat model of their systems at any point. That could be the true threat modeling as a service thing.

39:38Matt Coles Talking about a scaling problem, how many open source projects are there?

39:43Izar Tarandach How many do you think are running to ask for it?

39:46Chris Romeo Probably none.

39:48Matt Coles So combine it, combine it with the OpenSSF project that has a bootstrap for security for open source project maintainers where you can maybe drop threat modeling as a requirement into that Oh, I didn't know they had that.

40:02Izar Tarandach That's nice. Wow, that's nice. Tell us more about that.

40:09Matt Coles Let me look that up for you and we'll get that in the link.

40:11Izar Tarandach Please. Because I had no idea that that existed. And it's funny that we are discussing the usability and the scaling and we completely ignore the third one, the reliable automation and AI assistance.

40:30Chris Romeo I don't see that as a challenge, though.

40:32Izar Tarandach Right?

40:34Chris Romeo That's a benefit. Using AI to make the threat modeling person's journey easier, more efficient, more optimized is really the focus.

40:50Izar Tarandach But isn't that a shortcut to threat libraries?

40:57Chris Romeo Yeah, it's a threat library that you don't have to have as much knowledge of and context of, right, as the human. It's the machine taking some of the context, which is one of our challenges, right? There's a challenge there that the LLMs didn't mention. Developers don't know enough about security and privacy.

41:17Izar Tarandach Oh, they do have the skills gap here. Where's the skills gap? Too expert-centric. Most methods assume access to people who understand both security and the system deeply.

41:29Chris Romeo I agree. I mean, that's one of the ways you can use AI, though, in threat modeling, is to break down that wall where the AI provides the context so that you don't need a Matt Coles sitting next to you at your desk to be able to unpack these threats.

41:45Izar Tarandach Speaking of which, have you guys seen the— I think it's Google— the latest thing that they offered It's called CodeWiki. Look into that. It's like somebody takes NotebookLM, apply it to a codebase, and generate not only documentation, but a deep dive into the architecture. And it even creates videos and diagrams and whatnot.

42:12Chris Romeo It's wow.

42:13Izar Tarandach It's nice.

42:15Matt Coles So the two— there's two OpenSSF projects, just a little tight loop on that. There's OpenSSL Project Security Baseline, and there's—

42:22Izar Tarandach Oh, you said OpenSSL! I understood OpenSSL. Sorry, my ear. My bad, my bad.

42:30Matt Coles And the OpenSSL Scorecard. Those two projects drive security behavior for maintainers.

42:37Izar Tarandach So PyTM, for example, has an OpenSSL scorecard.

42:42Matt Coles Yeah, right. Now, if— but if PyTM was part of the baseline, it would be a natural driver for current models.

42:49Izar Tarandach Yep, yep, true. Hmm, OpenSSF is doing some good stuff, eh? All right, so where do we land with this?

43:04Chris Romeo Uh, do we land anywhere?

43:06Matt Coles Some models are— all models are wrong, but some are useful.

43:11Izar Tarandach LLMs just do it.

43:15Matt Coles LLMs get sued by Nike for—

43:16Chris Romeo Yeah, I was gonna say this could be a, this could be the first episode to get shut down by a large corporation.

43:23Izar Tarandach And we don't even have background music. Well, we had some in the beginning, but it wasn't actually music. It was just us talking about music. That's meta. Also meta can sue us now.

43:33Chris Romeo Yeah. Thanks. Thanks, César. Second lawsuit. Why don't you hold up another product? Oh, like a monster?

43:41Izar Tarandach Yeah.

43:41Chris Romeo No, hold up another product like a monster so we can add that to the video.

43:44Izar Tarandach I have no idea what you're talking about.

43:47Chris Romeo People that are going to see, like, line up people, the security table. If we have one thing, we got deep pockets. Not really. Oh yeah.

43:55Izar Tarandach No, if you count the holes in mine, they are very, very deep.

43:58Matt Coles I got some lint in mine if you want.

44:01Chris Romeo That's about all we'll offer in settlement.

44:04Izar Tarandach Sorry, we are modern here. We don't lint. We use black.

44:07Chris Romeo That's true. We use black.

44:10Izar Tarandach We ask Snick.

44:11Chris Romeo We're kind of off the linters of the world, but they do provide value.

44:16Izar Tarandach So they do.

44:17Matt Coles They do.

44:17Chris Romeo Leave them be.

44:17Izar Tarandach Don't be like a good linter.

44:19Chris Romeo Leave them be. All right. Well, I think we're just going to leave this episode with no conclusions.

44:23Izar Tarandach Leave you as it is. The conclusion might be that I have to stop asking LLM stupid questions.

44:28Chris Romeo But yeah, that could be a conclusion.

44:30Matt Coles That could be a conclusion.

44:30Izar Tarandach That could be a conclusion. Don't let me touch LLMs.

44:33Chris Romeo Yeah, I don't think we're going to go that far because—

44:36Izar Tarandach but just do it. No, seriously, seriously, people, just do it. Do what the Threat Modeling Manifesto does, says, and stop talking about it and just go and do it.

44:45Matt Coles Let's make this a non-get, don't-get-sued approach.

44:50Izar Tarandach Just threat model.

44:51Matt Coles Yes, just threat model.

44:53Izar Tarandach Just threat model. Let's end on a positive note. Positively, just threat model.

44:57Chris Romeo I just trademarked the word just. So now you owe me money. Just in.

45:04Izar Tarandach This just in. Justin? No. All right.

45:08Chris Romeo We'll leave the world a better place by saying just Threat Model. We're signing off. Have a good time doing your threat modeling.

Transcript supplied by AssemblyAI. It may contain transcription errors.

Originally published as The Security Table. Part of the AI Security Table archive.