# Don't Bury the Model T: Why STRIDE Still Drives in an AI World

URL: https://securitytable.ai/don-t-bury-the-model-t-why-stride-still-drives-in-an-ai-world/
Published: 2026-07-01
Duration: 3543 seconds
Season: 4
Episode: 14
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

Do AI systems require a new threat-modeling method, or are we abandoning useful tools before understanding their limits? Chris Romeo, Izar Tarandach, and Matt Coles first examine npm's move toward safer install defaults and the risk that agents trained on older behavior will simply re-enable dangerous options. Then they debate the claim that STRIDE belongs to a world that no longer exists. The discussion separates threat categories from modeling techniques, non-deterministic model behavior from deterministic controls, and unfamiliar failures from familiar security responsibilities. The Model T analogy gives both sides something to argue with: an older framework may need adaptation, but declaring it obsolete does not explain how the replacement improves the analysis.

Mentioned in this Episode:  
➜ [npm install documentation](https://docs.npmjs.com/cli/install/)  
➜ [Microsoft Threat Modeling Tool: STRIDE threats](https://learn.microsoft.com/en-us/azure/security/develop/threat-modeling-tool-threats)

Chapters:  
00:00:00 - Intro: deterministic fans  
00:01:32 - npm and secure install defaults  
00:03:22 - Will AI agents re-enable risky behavior?  
00:05:35 - Other package managers and safe defaults  
00:10:02 - Checking packages and internal repositories  
00:15:26 - Let's Encrypt and code-signing certificates  
00:20:04 - Install scripts and compatibility  
00:25:05 - The search for the wheat color  
00:27:34 - Was STRIDE built for a vanished world?  
00:30:04 - Threat categories and modeling methods  
00:35:19 - Non-deterministic AI and deterministic controls  
00:40:15 - Hallucination and cross-context failures  
00:45:04 - STRIDE and the Model T analogy  
00:50:34 - Tracing code versus explaining model behavior  
00:55:46 - Closing assessment

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

**Izar Tarandach (0:09)**: We're on Jeannie.

**Chris Romeo (0:10)**: It's true. We've gone from Jeannie to Dano, whatever that means.

**Matt Coles (0:14)**: Book 'em down.

**Chris Romeo (0:15)**: That'll be the name of my, my 7th book, From Jeannie to Dano.

**Izar Tarandach (0:21)**: History of bookings.

**Chris Romeo (0:24)**: History of bookings. Well, it appears once again on the Security Table, you have entered a conversation that was already happening, but we want to welcome you, our trusted deterministic fans from, from around the— or would they be non-deterministic?

**Matt Coles (0:43)**: Uh, they're almost certainly non-deterministic.

**Izar Tarandach (0:45)**: No, they're deterministic because we can trust them.

**Chris Romeo (0:48)**: True. I like that. Put that on a t-shirt.

**Matt Coles (0:50)**: Wait, wait, wait, wait. That Can you mathematically prove that?

**Izar Tarandach (0:55)**: Uh, I could call on people who actually know math to do it.

**Chris Romeo (0:58)**: Yeah. But I, I got the proof right here. 1+ 1= 2. Done.

**Izar Tarandach (1:03)**: Uh, on which base?

**Chris Romeo (1:04)**: Proven. Base 7.

**Izar Tarandach (1:08)**: It, it, it could be 0 and carry the 1.

**Matt Coles (1:10)**: True.

**Chris Romeo (1:13)**: I forgot about that. It's always the one that gets— well, hey folks, I'm Chris Romeo joined by Izar Terrandash, Matt Coles. We are the knights of the security table. Actually, we're more like the knights that say knee than we are the knights of the round table. But okay, let's dive in here. Apparently NPM has made some changes to stuff and junk.

**Matt Coles (1:39)**: About darn time.

**Izar Tarandach (1:42)**: About darn time is right.

**Chris Romeo (1:45)**: So, well, actually it's going to be this month apparently that these changes will be happening. Okay, about almost darn time. Basically, they're taking out some bad behavior. If I had to summarize this, they're taking out some bad behavior that npm could do. They're not removing it. They're just making it something that you have to enable.

**Matt Coles (2:06)**: In other words, by defaulting it, enabling security by default.

**Chris Romeo (2:10)**: Yes.

**Matt Coles (2:11)**: Finally.

**Chris Romeo (2:12)**: So just quickly to summarize what I see as the changes. Well, who am I kidding? What my AI says it sees as the changes. Npm install will not run pre-install, install, or post-install scripts from dependencies unless explicitly approved, which is that secure by default thing again. npm install will no longer fetch dependencies from Git repos, direct or transitive, unless explicitly permitted, removing code execution path where git dependencies— .npmrc file could alter which Git executable is used. And then third, dependencies installed from remote URLs such as HTTP. HTTPS tarballs would no longer be resolved unless explicitly permitted. So once again, does it sound like they're not, they're not taking anything out?

**Matt Coles (2:57)**: Nope. Because they don't want to be adding, they don't want to be breaking massive backward compatibility problems, right?

**Chris Romeo (3:02)**: Which I think is the right thing to do for something with the scale of npm.

**Izar Tarandach (3:06)**: They are gating the bad behavior, right?

**Matt Coles (3:09)**: Well, you have to make a decision. You as a developer, well, or you as a developer or you as the AI who is taking the place of the developer need to make a decision. Now, interesting, actually interesting. Sorry. Now that brings up an interesting thought. By leaving the behavior in, but enabled, but disabled by default. Does that mean almost certainly any AI that has been trained on npm prior to v12 will just re-enable it and we're going to be You're going to be broken again or potentially broken again.

**Izar Tarandach (3:43)**: It will try to do it the wrong way. It will see that it doesn't work. It will try 30 other ways, including erasing your whole file system.

**Chris Romeo (3:55)**: No, but it will work.

**Izar Tarandach (3:58)**: And then you are going to say, but wait, this changed. And it's going to go back to you and say, you are absolutely right. I should have checked which version of npm I was running before I destroyed the file system.

**Matt Coles (4:13)**: It's too late now.

**Chris Romeo (4:16)**: I mean, so who needs a file system?

**Matt Coles (4:18)**: Well, cause right. Cause V12, when V12 comes out, I mean, this won't be a, this will be a problem in the first like couple of weeks, but as soon as you get new models that are trained on V12, they'll know V12 has different behavior.

**Izar Tarandach (4:31)**: But to me, the really interesting thing here. Is that npm has taken this high profile as a package distribution system, as a package manager, right? And it's not a very new thing. For a while, it was limited more to people who dealt in a certain part of the space, but now it's a—

**Matt Coles (5:01)**: People who write JavaScript.

**Izar Tarandach (5:03)**: See, I didn't want to go there, but then you went there. Yeah.

**Matt Coles (5:06)**: So, uh, Ben, what else are you supposed to go there?

**Izar Tarandach (5:09)**: It's npm.

**Matt Coles (5:09)**: No, no, no.

**Izar Tarandach (5:09)**: I mean, it's a bad—

**Matt Coles (5:11)**: it's Node package manager. Node.js package manager.

**Izar Tarandach (5:14)**: It's just that kind of person, people that go, oh, that's just for front ends. No, no, no, no. Let's write the backend because that's a great idea. Yeah.

**Matt Coles (5:20)**: Cause somebody decided to write JavaScript to write backend servers.

**Izar Tarandach (5:25)**: That's for a different episode. But the, the, the thing now is—

**Matt Coles (5:28)**: I think we've had this episode before.

**Chris Romeo (5:30)**: We did?

**Izar Tarandach (5:31)**: Deja vu. So, uh, what I'm interested here is when will the other 30,000 different package, uh, package formats and package managers out there actually do the same thing, right? Because this is just the right thing to do. We should be making difficult for people to do bad things, even without knowing that it's a bad thing. Right. So if you really trust the package that you're pulling, then sure, by all means, allow dangerous, whatever the flag is. And if you don't, then at least stop and think before you do it and consider the consequences of the thing that you're asking your shell to do for you.

**Chris Romeo (6:19)**: So I just made the connection here that these npm changes would have prevented, if they had been in existence, the Shai-Halud worm would have been not allowed, wouldn't have been able to operate. It was using a post-install script.

**Matt Coles (6:34)**: It would have had to follow a different vector, right? It would have had to find a different package manager that had bad behavior. It wouldn't have been through npm. It would have been through something else.

**Izar Tarandach (6:43)**: It would have been slowed down because a lot of people would just look at the output and say, oh, you just have to add a flag. Let's do it.

**Matt Coles (6:50)**: Well, actually, that's an interesting question. Is there a bypass here? So if you disable but keep all the functionality. Could a package re-enable it? Like as a two-step process?

**Izar Tarandach (7:05)**: Any, anything that can call another thing will be able to re-enable it, right? So if you have a bad script that calls a good script with bad flags, you're going to get a bad outcome. But the thing here is that at the end of the day, who is this going to actually serve? The people who already had a minimum of security awareness to know that this might be a problem. So unless it comes out with a big honking banner saying you're trying to do this thing and this thing may damage your— then I don't know if a lot of people are going to extract a lot of content from this.

**Matt Coles (7:47)**: Well, I mean, honestly, right now with everybody doing, you know, or not everybody doing vibe coding, or spec-driven development, developers will never see it.

**Izar Tarandach (7:59)**: Yeah, but this is going to break a lot of pipelines down there, right?

**Matt Coles (8:03)**: Because they rely on pre and post-install functionality?

**Izar Tarandach (8:06)**: They rely on installing a lot of these packages, and some of them do the post-install thing. I'm not saying all of them, but some of them I'm sure do. So there may be some breakage in the horizon if this becomes the default.

**Matt Coles (8:20)**: Well, why wouldn't you just ask your AI to fix it? Which, as you said, may result in the deletion of your entire application, but at least initially.

**Izar Tarandach (8:35)**: Which perhaps carries us into the next one.

**Matt Coles (8:40)**: Well, before we get there, how did, how did RPM solve this?

**Izar Tarandach (8:43)**: Right.

**Matt Coles (8:43)**: So you, I think we were talking earlier, RPM.

**Izar Tarandach (8:46)**: Well, on RPM, you can say explicitly which stages of the install you want to run, right? If you don't say anything, then it's going to run the whole thing, the pre, the install, and the post. Or you could say just run this one or just run that one. The thing is that RPMs, perhaps because of their age, I think, they do have a bit more of a security veneer around them. So you have they are signed. I'm not familiar with npm to the point of saying that those packages can be signed. I think that it's more of a provenance thing. Could be wrong. But RPM does have the whole signature verification ecosystem around.

**Matt Coles (9:31)**: But signing doesn't eliminate bad behavior.

**Izar Tarandach (9:34)**: No, it doesn't. But it gives you at least a Again, that's why I said the security veneer. It gives you the provenance. And if your sources are trusted, then by the transitive property, you're saying, okay, I trust them to run the pre-post and install scripts the same way that you run the, that you trust them to run their binaries. Right.

**Matt Coles (9:58)**: Well, I guess that's today it's implicit trust, right? Because nobody's checking the packages. I mean, if they did, do you know anybody who— I mean, do you know of anybody who goes and looks at all of their JavaScript packages? Like explicitly, I want this one, I want this one, I want this one. Okay, besides Chris.

**Izar Tarandach (10:21)**: No, I know people who will bring stuff in, scan stuff, put them in their own repository and direct all their developers to only use that. That repository, right? Because at the scale and speed that these things go, be honest with you, I'm not expecting anybody to do that verification by hand. It's like, oh, new shiny thing.

**Chris Romeo (10:49)**: Yay. And there's tools to do that as well. You don't have to build that type of internal repository from scratch.

**Izar Tarandach (10:55)**: Yeah, true. But hey, you know what? All in all, a good step forward.

**Matt Coles (11:04)**: Well, yeah, it is.

**Chris Romeo (11:05)**: It's about time. But there are some bypasses to it. That's what I was looking at right now. Your, your AI can bypass it is actually true. Um, if developers just run npm approve scripts --all to unlock their build, the protection collapses entirely. And it says AI coding tools adding dependencies at speed may outpace human review, risking the allowlist becoming a mere formality.

**Matt Coles (11:28)**: Mm-hmm. Well, so the alternative is they disable the functionality entirely, make it backwards breaking change, you know, making a breaking change. And then, and then what? So now you need to have explicit trust of packages, right? That become— that defeats the purpose of npm perhaps, or has it become like PyPI?

**Chris Romeo (11:48)**: No, I mean, npm's goal should be the inherent trust of all packages that you get from it. Now, that's an easy thing to say, not as easy of a thing to implement. But isn't that why they exist for the future, at least, is to provide trusted packages? Aren't they a trusted repository of packages?

**Izar Tarandach (12:15)**: But are they trusted because they said that they are to be trusted, or are they trusted because by use people decided to trust them.

**Matt Coles (12:24)**: I mean, people trust PyPI, right? Or not.

**Izar Tarandach (12:32)**: No, I mean, did they wake up in the morning and said, we are going to build a trusted environment? Or did they just say, we are going to be the place where people will come to, to get their packages?

**Chris Romeo (12:44)**: I think it's option 2.

**Izar Tarandach (12:45)**: I don't know.

**Matt Coles (12:46)**: Yeah. Speed over convenience over anything else.

**Chris Romeo (12:49)**: Right.

**Izar Tarandach (12:50)**: Yeah.

**Chris Romeo (12:51)**: But in the second article that we're going to talk about here, there was mention made of OpenSSL's, maybe it wasn't in the second article, but apparently open, no, they have a, they have a signing project for packages.

**Izar Tarandach (13:07)**: Yeah. The cosign. And, uh, I forget that it's actually a very good. Good project with a lot of very good people in there.

**Matt Coles (13:15)**: Is it? It's like, um, Let's Encrypt, which—

**Chris Romeo (13:21)**: that's what I started saying about 5 years ago is like, somebody needs to build the Let's Encrypt certificate version for packages. Because think about, do you not even think about TLS certificates anymore for public sites? It's just a commodity. That you don't even pay for. Like they just get generated by this, this service. If we could get to that point with, excuse me, with packages like that would be to be able to have that provenance, true provenance and cryptographically proven provenance of a package. It's a step.

**Matt Coles (14:00)**: Interestingly though, so let's, getting a little sidetracked, but let's encrypt verifies that you have ownership of the site before it gives you a certificate. How would it do that for software packages?

**Chris Romeo (14:17)**: You have to be a repo. You have to be a repo thing, or you do the same exact thing in a repo. You have to put a file in a repo like you would do in a DNS info record or something to prove ownership.

**Izar Tarandach (14:30)**: Hmm.

**Chris Romeo (14:35)**: There, I solved it. Somebody can go build it now.

**Matt Coles (14:37)**: There you go. Somebody write a spec. Somebody write a spec and get Claude to write it for you.

**Chris Romeo (14:42)**: Yeah. Somebody vibe code it today and let's get this thing deployed tomorrow.

**Matt Coles (14:47)**: What's the cost of that? Actually, who pays the cost of the certificates?

**Chris Romeo (14:52)**: I think it's a not-for-profit. That's, uh, charity, a charitable organization. Last I heard, I don't know if some big companies— I feel like some big companies seeded it with money, but I'm double-checking just because with everything else in the world, I don't know anything anymore. I just type words into an AI bot and it tells me, because, because AI has become a has become a search engine. Yeah, there were 5, 5 original sponsors, Mozilla, EFF, Cisco, Akamai, and IdenTrust.

**Matt Coles (15:32)**: Mm.

**Chris Romeo (15:33)**: They kicked in a bunch of money. Yes. But now the Gates Foundation's on board.

**Matt Coles (15:38)**: So there you go. Well, so maybe we can get them to, uh, to make code signing certs. Call to action.

**Chris Romeo (15:45)**: Issued more than 7— Let's Encrypt has issued more than 7 billion certificates since launch.

**Matt Coles (15:51)**: And how many of those sites have served malware?

**Chris Romeo (15:55)**: I mean, some, 30, 40%.

**Matt Coles (16:01)**: What's in, what's in your thread bottle?

**Izar Tarandach (16:04)**: We have to do that commercial one of these days.

**Chris Romeo (16:10)**: All right. Well, let's talk about this other thing. I think we've, we've talked about the npm. Yeah.

**Matt Coles (16:15)**: It was a good change. By the way, when did GitHub own npm?

**Izar Tarandach (16:21)**: When what?

**Chris Romeo (16:22)**: GitHub bought npm?

**Matt Coles (16:23)**: Well, okay. It's GitHub announces there's an npm change. Why did GitHub announce?

**Chris Romeo (16:29)**: That's weird. I didn't think of that.

**Izar Tarandach (16:32)**: Because they're the ones who are going to break the most.

**Matt Coles (16:37)**: It's, it's, it's a GitHub. Yeah. GitHub blog. About npm, our next npm major version.

**Izar Tarandach (16:50)**: Huh.

**Matt Coles (16:50)**: When did GitHub own npm?

**Chris Romeo (16:57)**: I'm going to click on the actual link up there.

**Izar Tarandach (17:00)**: Yeah, I did. And it has a label on top, retired.

**Matt Coles (17:03)**: What's that?

**Chris Romeo (17:03)**: Yes, that's weird.

**Matt Coles (17:05)**: I just noticed that too.

**Chris Romeo (17:07)**: What is that?

**Matt Coles (17:11)**: What's retired? NPM? The announcement?

**Izar Tarandach (17:15)**: The announcement? What the? Okay. Now I'm confused.

**Matt Coles (17:21)**: What?

**Chris Romeo (17:22)**: Maybe we're not in favor of this now. Now we're confused.

**Izar Tarandach (17:25)**: I'm confused. GitHub Enterprise. Okay. I need something to drink.

**Matt Coles (17:31)**: Oh, and then, oh, and then the previous, the previous one is Dependabot version updates now support the Deno ecosystem. GitHub acquired NPM in 2020. 2020. Okay, we're a little bit behind the news.

**Chris Romeo (17:45)**: What?

**Matt Coles (17:48)**: Get up.

**Chris Romeo (17:49)**: My mind is just—

**Izar Tarandach (17:50)**: what?

**Chris Romeo (17:51)**: Is there— is there— so upset about this, he just left the studio.

**Izar Tarandach (17:55)**: Left?

**Chris Romeo (17:55)**: He just left. He just got up and walked away and said, screw this, I'm out.

**Matt Coles (18:00)**: He needed a drink. He's going to get something stronger.

**Chris Romeo (18:02)**: O'clock in the morning.

**Matt Coles (18:03)**: He's getting a Monster.

**Izar Tarandach (18:05)**: I went to check if, uh, GitHub by mistake had acquired itself.

**Matt Coles (18:10)**: Well, apparently GitHub has acquired npm in 2020. Where were we?

**Izar Tarandach (18:17)**: But wait, wait, wait, wait, was npm a thing that could be acquired? I didn't know that. That's how out of npm I am. Actually, since, since we, we mentioned anything, uh, uh, vibecoded. I want to turn to our listeners and ask for a question, ask for help. So npm, if you do npm audit, it will look into something called a package lock into the directory that you are, go through all those packages and their dependencies and see if there is any outstanding security alerts for them or whatever. You can see these, I don't know. That's, that's all good and nice for the packages in the, the project that you're working on and their dependencies. But there's all kinds of tools that you, you install using npm, which could be -g global or, or private. I haven't found yet a way to scan those. For security issues. So if anybody knows of something like that, drop us a line, give us a comment, poke us on LinkedIn or something. Because if there isn't one, my question is why? That's no less important, right? I mean, even our AI agents are written in, distributed through npm. So it shouldn't be a matter of like, going and say, oh, I remember that once I installed this thing, let me see if I can update it. There should be a—

**Matt Coles (19:58)**: Will this change break AI installs now?

**Chris Romeo (20:04)**: I don't know.

**Matt Coles (20:05)**: Probably they rely on these pre and post install scripts to deploy. In fact, they probably do because you deploy the bits and then you run the post install script to actually do the install.

**Izar Tarandach (20:17)**: Yep. I guess.

**Matt Coles (20:19)**: So that's an interesting, that's an interesting thing because if they make this the default behavior, but it breaks all the installs for convenience, they'll probably be disabled by default soon.

**Chris Romeo (20:29)**: It's set in the last, in the last version of 11, npm 11, they made them all warnings so that people would be getting some visibility that things are about to break.

**Izar Tarandach (20:40)**: There's a bunch of stuff that you can add with npm. Like Skills and this and that and the other one that are not exactly binaries, but they have post-install scripts because they ask all kinds of questions on the install.

**Chris Romeo (20:54)**: According to my research, only 2% of npm packages need install scripts to function.

**Izar Tarandach (21:01)**: I had like, yeah, but the question is which 2%. I mean, that makes sense because a lot of them, most of them, I'd say, are probably libraries that come as dependencies, right? So you need them around in order to be able to run something. But exactly what Matt is saying, that these packages that now are part of our AI— I'm sorry, I can't call them agents— our AI clients, right? Those things do seem to need Postgres. So I remember in Over the weekend, I installed something and it asked a bunch of questions afterwards. I'm guessing that that was a post-install. If it wasn't, then there's another thing that we have to turn off. But, uh, the, the, the derivative here for, for me is that if it becomes a learned behavior that you have to run it with dash dash whatever to get those things to work, then what have we achieved?

**Matt Coles (22:00)**: It's the security theater, the button, single button click. Where's your, where's your single button keyboard there? Is that, cause that's basically what it is all over again.

**Izar Tarandach (22:09)**: Thanks for remember.

**Chris Romeo (22:13)**: Yeah.

**Matt Coles (22:14)**: So I'm looking at the discussion thread on this. So the functionality exists, but it doesn't exist in a question is one, one, some user to their discussion, will Node 26 ship with npm 12? And it's an ongoing discussion about which version of, of Node will ship with this thing by default.

**Izar Tarandach (22:37)**: I think we should invite somebody to talk to us about the npm universe and make us more educated about it, me at least, because I don't feel like I know enough of the details there.

**Matt Coles (22:48)**: I mean, that'd be interesting. That'd be an interesting episode anyway to talk about not just, as you mentioned, there's other things, right? That we RPM, but also Go, right? There's other sort of language with built-in package manager discussions to be had.

**Chris Romeo (23:02)**: Right.

**Izar Tarandach (23:03)**: Because now they are environments. They're not languages. Once upon a time you had VI and a compiler and life was good. And now you have—

**Chris Romeo (23:12)**: Life was not good.

**Izar Tarandach (23:13)**: Hey, life was great.

**Matt Coles (23:15)**: Okay. There's no IDE. It was better than having Ed. You remember Ed?

**Chris Romeo (23:24)**: What was the line-by-line one?

**Izar Tarandach (23:26)**: Ed. That was Ed.

**Chris Romeo (23:27)**: Was it Ed?

**Matt Coles (23:28)**: Yeah.

**Izar Tarandach (23:28)**: Yeah. It was VI.

**Chris Romeo (23:31)**: VI was an improvement over Emacs, which is Yzhar's go-to.

**Izar Tarandach (23:36)**: Hey, hey, hey, no, nobody says a word about Emacs.

**Chris Romeo (23:40)**: See, I told you. This is his.

**Izar Tarandach (23:43)**: I wonder if somebody already wrote a Lisp package that runs Claude inside of Emacs. I'm sure they have.

**Matt Coles (23:49)**: If they haven't, you will have it by next week.

**Izar Tarandach (23:53)**: Ah, no, I don't touch Emacs anymore.

**Matt Coles (23:55)**: You don't have to.

**Izar Tarandach (23:56)**: I'm sorry to say.

**Matt Coles (23:57)**: You don't have to, Claude will.

**Izar Tarandach (23:59)**: Yeah.

**Matt Coles (24:00)**: Have your friend Claude do it.

**Izar Tarandach (24:01)**: Those were good times.

**Matt Coles (24:03)**: Those were good times.

**Chris Romeo (24:04)**: Tell it to use Emacs, it'll be like, no, I can't do that.

**Izar Tarandach (24:08)**: No, no, no, no, seriously, seriously. I remember Eliza. Do you guys remember Eliza?

**Matt Coles (24:13)**: Eliza, yes, a little bit.

**Izar Tarandach (24:14)**: Yeah. Eliza running inside of Emacs. That was one of like the first things that I actually discovered on Emacs and it led me And that led me through a very, very deep rabbit hole of learning Lisp and Prolog.

**Matt Coles (24:31)**: When did you get colorization in your editor?

**Izar Tarandach (24:35)**: It was around '91, '92. There was the, what was the name of the package? Transmogrifier, something like that. That would give you syntax. Color in Emacs. And the fun thing was it took me a long time to get to a color scheme that I could work with.

**Matt Coles (24:59)**: And there was no accessibility at that point. I mean, no.

**Izar Tarandach (25:03)**: And just last week, just last week, I was trying to like find a color for each different agent that I was— sorry, client that I was running. And I was looking for wheat. Wheat was like the default of XEmux when it came out. And I just couldn't find the right RGB for wheat. And it drove me nuts.

**Matt Coles (25:28)**: You know, there's a limited set, right? You could test them all. And then, I mean, wheat is mostly white, right? So it's yellowish.

**Izar Tarandach (25:36)**: No, that looks like work. We don't do work anymore.

**Matt Coles (25:40)**: You don't have to. That's what your clients slash agents are for.

**Izar Tarandach (25:44)**: Okay. So what did I do?

**Matt Coles (25:45)**: I found an agent swarm, test each of these, look for wheat.

**Izar Tarandach (25:49)**: No, no, no. What you do is I went up in the attic. I found an old laptop that I had that ran Linux circa, I don't remember when. And it had in it like my .files from way back when, and it had .emacs and .x, uh, uh, .x, what was it? .xattributes or something like that.

**Chris Romeo (26:11)**: .rc?

**Izar Tarandach (26:12)**: No, no, no, no, no, no, no. The .x files.

**Matt Coles (26:15)**: All right. I have a surprise for you, Azhar.

**Izar Tarandach (26:17)**: And in one of them, I had .foreground equals the RGB for wheat. So I was able to find that.

**Matt Coles (26:24)**: I have something for you.

**Izar Tarandach (26:25)**: No, don't tell me that Gemini found it for you.

**Matt Coles (26:28)**: I did. Not Gemini, but there is an Emacs themes site. With GitHub Pages for the wheat.

**Izar Tarandach (26:38)**: Yeah. You know what? I had much more fun looking at my old laptop.

**Matt Coles (26:43)**: Including, including the RGB color, color blocks. So, you know, exactly in CSS, so you can apply it to your HTML pages.

**Chris Romeo (26:54)**: Yeah. I heard wheat's making a comeback. It's supposed to be the color of the year for 2027.

**Izar Tarandach (26:59)**: It's easier on the eyes.

**Matt Coles (27:01)**: It is.

**Chris Romeo (27:03)**: Everybody's gonna be using it. It's gonna be, houses are gonna be painted wheat.

**Izar Tarandach (27:08)**: Okay. So wheat is the new dark mode.

**Chris Romeo (27:11)**: Yeah, it is. I heard that too.

**Izar Tarandach (27:14)**: T-shirt.

**Chris Romeo (27:15)**: From people in, people in Silicon Valley. I heard that.

**Izar Tarandach (27:19)**: T-shirt. Okay, cool.

**Chris Romeo (27:20)**: Wheat. I mean, they were on the TV show.

**Matt Coles (27:22)**: Wheat on black.

**Izar Tarandach (27:25)**: What?

**Matt Coles (27:26)**: Wheat on black.

**Izar Tarandach (27:27)**: So, oh, we turned black. So we're going to jump into the STL is broken or we're going to keep it for next week.

**Chris Romeo (27:34)**: I think we should keep it. Oh no, let's talk about the Stride built world longer exists. I kind of said that wrong.

**Matt Coles (27:41)**: Stride was built for a world that no longer exists is the topic, yes.

**Izar Tarandach (27:46)**: In a world that doesn't exist anymore, a methodology rose to transform the world.

**Chris Romeo (27:51)**: You really need that guy. Once again, you need that guy from the movie previews. In a world that doesn't exist, Stride is missing.

**Izar Tarandach (28:04)**: Hey, we are continuously searching for it.

**Matt Coles (28:08)**: It's a stride out of time. All right.

**Chris Romeo (28:13)**: So let's do this.

**Izar Tarandach (28:16)**: Has anybody come up with a StrideBear agent anymore already?

**Matt Coles (28:21)**: Probably. I think you just invented it.

**Izar Tarandach (28:23)**: Yeah, I think I'm going to call this. I'm going to put a stake on the ground for that one.

**Matt Coles (28:28)**: Stride for client?

**Izar Tarandach (28:29)**: Stride for client. No, that, that's a consulting gig.

**Matt Coles (28:33)**: Ah, there you go. So, uh, so we're interesting.

**Chris Romeo (28:41)**: I mean, do we want to be here? Do we want to be just normal mean levels or do we want to like—

**Izar Tarandach (28:45)**: No, no, no. Wait, wait, wait, wait, wait. I think I'm in the, in the wrong place. I'm in the STL is broken and that's actually good news.

**Matt Coles (28:53)**: You're, you're the wrong article.

**Chris Romeo (28:55)**: We're in Stride Was Built for a World That No Longer Exists.

**Matt Coles (28:57)**: Or move backwards because it was posted previously.

**Izar Tarandach (28:59)**: Oh, backwards, backwards. Okay. Okay. Let me go look for it. Okay. You guys start. I'll, I'll, I'll drink it.

**Matt Coles (29:05)**: It's in chat, by the way. Just watch the Isar just catching up to, because he was still, he's still stuck on his wheat.

**Chris Romeo (29:13)**: He's still, he's still just looking for the color wheat. So he might tell us.

**Matt Coles (29:17)**: No, he's playing, he's playing Settlers of Catan and he's looking for wheat.

**Izar Tarandach (29:21)**: Just, just assume that I'm stuck. Like it's, it's, it's the safe choice.

**Chris Romeo (29:26)**: Just, he's playing, uh, what's the, what's the game that Oregon Trail?

**Matt Coles (29:34)**: Oregon Trail.

**Chris Romeo (29:34)**: He's playing Oregon Trail over there.

**Izar Tarandach (29:36)**: Oh, you, you just died of, uh, what was it?

**Chris Romeo (29:39)**: Dysentery.

**Izar Tarandach (29:40)**: Dysentery. Dysentery.

**Chris Romeo (29:41)**: I don't know. Izar died of dysentery again and again. All right.

**Izar Tarandach (29:48)**: So, hey, I played the thing on an Apple II. Yeah.

**Matt Coles (29:51)**: I didn't play it on Apple II. Apple II Plus. Yeah.

**Chris Romeo (29:54)**: In, in green.

**Matt Coles (29:55)**: Uh, yep. Anyway.

**Chris Romeo (29:58)**: All right. So, yeah. How did I end up in central? It's gonna be normal here when we respond to this.

**Matt Coles (30:04)**: So Stride. So two things about Stride. Stride is both a threat categorization and a methodology.

**Izar Tarandach (30:13)**: Yeah.

**Matt Coles (30:14)**: Right. So you have STRIDE, the spoofing, tampering, repudiation, et cetera. And you have per element and per interaction.

**Izar Tarandach (30:22)**: Okay. Agreed.

**Chris Romeo (30:23)**: Correct.

**Izar Tarandach (30:24)**: So far so good.

**Matt Coles (30:26)**: Both of those, I would argue, still apply in an AI world. And in fact, it applies, I think, even more so in an AI world because what's an AI? An AI is a process. Uh, AI is a behavior, has behaviors, and Stride, the Stride threat categories are all about behaviors.

**Izar Tarandach (31:00)**: Oh God, I'm gonna be so mean here.

**Chris Romeo (31:04)**: To who?

**Matt Coles (31:05)**: To who are you going to be mean?

**Chris Romeo (31:06)**: To whom?

**Izar Tarandach (31:08)**: To a world that no longer exists. I mean, why do people feel this, this, this need to put down something that has been serving us quite well, I might say. Not exactly my choice of where to go, but has its uses, definitely.

**Chris Romeo (31:33)**: Right? Sure.

**Izar Tarandach (31:34)**: And, and then start saying, oh no, no, no, doesn't work anymore.

**Chris Romeo (31:37)**: No, no, no, no, no.

**Izar Tarandach (31:38)**: The world changed. Guys, for, for, for, for the love of Tulu, it's bytes. They go into CPU, they come out. We're not quantum yet. It's not like all possible things happening and will happen everywhere at the same time. No, it's bytes come from memory, go into CPU, an opcode gets executed, a result comes out, right? It's, it's the same thing as people saying, uh, you know what, we, we, we can't have maps anymore. No, no, cities can't have maps anymore because hey, you know what, cars can take any way they want, right? You never know which way a car is going to take. Hey, it can drive this way, it can drive that way, it can drive the other way. So let's not make maps anymore.

**Matt Coles (32:19)**: If you're in Boston, like one-way streets don't matter, right?

**Izar Tarandach (32:22)**: So, well, that, that's a different thing. Jerusalem, same thing. I mean, you can come in a different day and the, the, the, It's going completely different way. But my point is, we are all running around saying, oh, all these things are non-deterministic. Yeah, sure, fine. Yeah, yeah, yeah. You don't know exactly what's the next step, but the whole thing should be bound. Only a certain number of things can happen. You just don't know, okay, how it got there, the order, but—

**Chris Romeo (32:48)**: That's an important issue that, but I've even got something we can do before that. Can you, can you go to the top of this article? And this example that he has here about the mid-market healthcare company integrated ChatGPT, they ran a formal STRIDE threat model and then the world fell apart.

**Izar Tarandach (33:07)**: Wait, I haven't seen the world falling apart yet.

**Chris Romeo (33:10)**: Okay. Well, so then week 3, a patient called their medical records appear in the ChatGPT response, not their records, someone else's. Can't we tie each of these instances, like summarize, and then the next one was summarize my previous chat history. Isn't tampering in a refactored AI world, isn't this just tampering? And maybe Stride needs to be, some additional explanations need to be made for how Stride applies to AI. I keep seeing this.

**Matt Coles (33:40)**: I was ready to go explode over there.

**Izar Tarandach (33:43)**: Go ahead. So the thing here is that But what's not being said here, right, is that it's the same thing as, okay, coding agents. We already agreed that we should relate to our code coming out from a coding agent as code from a junior developer, right? Check it, read it, review it, whatever. Just don't accept it at face value. On the other hand. When we're talking about chatbots that could have like a query functionality, people are willing to accept those things as if they are like master librarians that know exactly where to go and get only the right thing and bring it right in front of you and put it in front of you. You know how you don't get things that can do what's being described here? By binding the amount of agency that they can have by using controls beyond the AI that will say, you know what, mine that record, but not that one. Databases have per-role controls and have had them for what, 30 years? Because that stuff works. But just to say, hey, AI, whenever somebody asks for a record, only show their records, nobody else's, okay? That's not non-determinism, that's ridiculous optimism. That's not how things work. So yeah, the behavior of the AI may be non-deterministic, but the controls that you put in it are deterministic. And those controls, they tie to specific scenarios of failure. Those scenarios of failure tie to impacts. Stride likes impacts. You probably, as you said, could bind those things back to Stride. And it's not like all of a sudden there is this new class of stuff that Stride never saw before and oh my God, we need another acronym. We need layers, we need layers.

**Matt Coles (35:49)**: All right, so I'll play devil's advocate a little bit here. Why did StrideLM exist? It's a rhetorical question. You don't have to answer.

**Izar Tarandach (36:02)**: Wait, wait, wait.

**Matt Coles (36:03)**: Do we want to tell people who are not familiar with Uh, Stride LM, right? Stride plus lateral movement. And that's the plus lateral movement that is the important part. Somebody decided that stride was necessary, but it was insufficient or it was lacking something. And so they added lateral movement, right? And so you had the base and then you extended it to fit a particular use case.

**Izar Tarandach (36:26)**: But that's the, that, that's the thing. What did you extend it with? You didn't extend it with one more impact. You stand to, you, you extended it with one more way of using the previous 5 impacts to get to where you want.

**Chris Romeo (36:41)**: True.

**Matt Coles (36:42)**: Right.

**Izar Tarandach (36:42)**: Right.

**Matt Coles (36:43)**: Right. So utilize tampering to get a pivot.

**Izar Tarandach (36:46)**: Exactly. So it's the same thing as prompt injection, right? I, I, the, the, the folks at OWASP who are doing the, the, the whole LLM, uh, top 10 again. I had the chance to go into the channel that's discussing prompt injection itself as a threat and put in there a lot of what I've been putting here on the podcast for a long time, that it's not a threat. Prompt injection is not a threat by itself, right? And I was expecting some pushback or I was expecting to learn something that would completely change my view. And I did learn a lot of stuff for sure, but nothing that disagreed with me. And they were extremely gracious in, in allowing the newbie to step in into something that clearly had been discussed ad nauseam over there. Right. And they said, yeah, it's not really a threat. It isn't. It's a vector. It's a, it's a way to get there. It's something that looks bad. But it's not by itself a trap because it's not doing things. You need other stuff on the tail end to actually have an impact on, on something. Right? So the LM in here, the lateral movement by itself, it, it wouldn't be a vector. It wouldn't be an impact. It, it just would be something that takes you further. It would be even a result. You, you want to get that foothold so that you can get lateral movement so that you can go somewhere else and have more impact.

**Matt Coles (38:17)**: And you're touching upon something that's very important. Actually, it's mentioned in the next paragraph. It said the threat model said this was unlikely. This was the hallucination part. And, but the attacker made it repeatable. Stride doesn't decide likelihood. Stride is a category of threats.

**Chris Romeo (38:38)**: True.

**Matt Coles (38:38)**: Or Stride is a way of identifying threats based on a DFD. It doesn't decide likelihood. Something else decides likelihood. Traditionally, that's the humans who decide likelihood. They look at a particular threat and say, how is it likely to occur? Is it likely to occur? Is it something we should do something about? Right? To some extent, or actually in reality, you're actually asking one of the four questions of the four-question framework when you're applying STRIDE. What could go wrong and/or what are we going to do about it? And one of those answers for what are we going to do about it may be nothing because it's unlikely to occur. So it's not Stride, it's a threat modeling process and the decisions that are made that decide that that threat may be unlikely. And so if the threat model decides something's unlikely because of past behavior, but now you have repeatable behavior that says otherwise, update your threat model. Don't kill Stride away.

**Chris Romeo (39:35)**: Yeah, I'm looking, I'm looking at this going, the people that should be fired are the people that didn't do a proper threat model considering all of their, all of their threats. Like they ran a formal Stride threat model is what it says at the top. Why wouldn't you, why wouldn't you consider AI in that threat model? Like this is kind of a strange example.

**Matt Coles (39:56)**: Maybe if you don't know, if you don't know that hallucinations can occur with cross, with cross-conversation chat text.

**Chris Romeo (40:00)**: Well, then you shouldn't have deployed a ChatGPT-connected API into an application.

**Izar Tarandach (40:03)**: That's one.

**Chris Romeo (40:05)**: I didn't have the knowledge to understand how it worked.

**Izar Tarandach (40:08)**: Should I ask it how it worked?

**Matt Coles (40:12)**: Can you hallucinate? No, of course I can't hallucinate.

**Izar Tarandach (40:15)**: Wait, wait, wait, wait. When somebody says something like an hallucination occurred because of a past conversation.

**Matt Coles (40:25)**: Cross-context.

**Izar Tarandach (40:27)**: Cross-context.

**Matt Coles (40:28)**: Which we know. So maybe we know that now. Maybe we didn't know that. No, no, no, no, no, no, no, no, no, let's go back there.

**Izar Tarandach (40:34)**: Let's go back there. Let's forget the hallucination for a second. Let's talk about the cross-context here. How long have we known about the need to clear your buffers before you start a new session of anything that runs in a loop? How long have we known about mining swap files for passwords? How long have we known about looking for data structures in process memory to figure out what was happening? It's not a new thing people!

**Matt Coles (41:08)**: How many people understand that npm automatically does things when you install packages?

**Chris Romeo (41:14)**: But th—

**Matt Coles (41:14)**: That's the thing... We're talking no-no-no-we've-known-about-it-for-decades-nobody-is-paying-attention-to-it-that-doesn't-make-Stryd invalid.

**Izar Tarandach (41:25)**: You know what? No, no, no. That's for sure.

**Chris Romeo (41:27)**: That's for sure.

**Izar Tarandach (41:28)**: No, that's not what I'm saying.

**Matt Coles (41:29)**: That's not what I'm saying.

**Chris Romeo (41:30)**: I know, I know.

**Matt Coles (41:31)**: What I'm saying again is—

**Izar Tarandach (41:31)**: I'll just bring it back to— I agree with you completely. What I'm saying again is we are making names for this esoteric things that happen inside AIs, like, I don't know, cross-context contamination. Oh my God. CCC. Oh, it's a new acronym. Yeah, yeah, yeah, yeah. It looks good. But damn it, it's just applied stupidity. I'm starting a new session, I clear the memory.

**Chris Romeo (42:02)**: Seems to be basic practice and good hygiene.

**Matt Coles (42:06)**: Well, actually, it is good hygiene, but doesn't it happen for you all the time? And I'm not defending it, by the way. I'm just saying, if you start a new application, you make an assumption. We've talked about assumptions a lot in threat modeling, right? That you're making an assumption that memory gets cleared for you by the compiler, you know, compiler setup, you know, prefix code to take care of it for you.

**Izar Tarandach (42:31)**: That would be what professionals call a wrong assumption.

**Matt Coles (42:36)**: As a computer scientist, as a computer scientist, what do you say?

**Chris Romeo (42:41)**: Wait, are we professionals?

**Izar Tarandach (42:42)**: No, no, no, no, no, no. Sorry. Sorry. As a computer scientist, I would expect somebody to make that wrong assumption. As somebody who actually wrote software for a living, I would say, no.

**Chris Romeo (42:53)**: Are we in the professional category though?

**Matt Coles (42:56)**: No, I don't know.

**Chris Romeo (42:56)**: 3 of us?

**Izar Tarandach (42:57)**: Wait, wait, wait. What profession is it?

**Chris Romeo (42:59)**: I mean, what profession?

**Izar Tarandach (43:02)**: No, I mean—

**Chris Romeo (43:04)**: Casino gambler?

**Izar Tarandach (43:08)**: No, that's when we're talking about the probability. Crypto. The thing here is, the thing here is we, we, we, when I say we, it's not even the royal we because I don't think that we are because I like to think that we are looking at these things in a different light.

**Matt Coles (43:26)**: We versus we.

**Izar Tarandach (43:28)**: Yeah, right. The vertical we as opposed to the horizontal we. There is this rush to create this whole—

**Chris Romeo (43:39)**: I'm really confused right now. I'm not a professional.

**Matt Coles (43:42)**: A little Steve Martin there, were you?

**Chris Romeo (43:45)**: Yeah.

**Izar Tarandach (43:45)**: There is this whole rush to create a new horizon of thing, a new taxonomy, a new ontology, a new something of things, right? And—

**Chris Romeo (43:59)**: Oh wait, hold on. I got to steal third. He's on first. Who's on first? Who's on third?

**Izar Tarandach (44:08)**: That there's this whole, uh, let's invent this whole thing. And, and I think that's coming and saying that something like stride all of a sudden lost its meaning or it's not relevant anymore is— it's such a leap in trying to coin a new revolution where there isn't one.

**Chris Romeo (44:29)**: No, let me give you an illustration. Let me give you an illustration that might, might help to wrap this, this thing up and put this article to bed. So remember the Model T, right? Ford created this very early model car. I can look at the Model T and say, I can't believe it doesn't have airbags. How do they, you know, it's, it's not a car or seatbelts. It's not a car because it doesn't have airbags and seatbelts. Yes, it's still a car. It's still function. You can still drive it down the road. Have, have other things, have cars progressed? In over the years since the Model T? Yes, of course.

**Matt Coles (45:10)**: Are they still cars?

**Chris Romeo (45:10)**: That's what we're talking about here. There are newer things. Yes. Is Stride the latest and greatest thing? No, there are things, there are, there are new method, not methodologies, but, um, mnemonics and things that have been created for threat modeling of AI.

**Matt Coles (45:26)**: Yep.

**Chris Romeo (45:27)**: And so why would I try to drive a Model T across the country and expect it to be as safe as a modern-day car?

**Izar Tarandach (45:36)**: I wouldn't.

**Chris Romeo (45:37)**: I would look for a modern-day car and I would leave my Model T in the museum. So I just suggested Stripe needs to go in a museum. Dang it.

**Matt Coles (45:46)**: No, no, no, no, no, no, no.

**Izar Tarandach (45:47)**: Look, look, look, you got something here. You got something here.

**Matt Coles (45:51)**: He had something, but he completely went off the rails.

**Izar Tarandach (45:55)**: He tried to go through, through the whole country, but The Ford T is going to take you from point A to point B. Okay. Same way that a Ferrari is going to take you. Probably it's going to be slower. Depends on the angle of the—

**Chris Romeo (46:10)**: Oh no, it is going to be slower.

**Matt Coles (46:11)**: It's going to be slower.

**Izar Tarandach (46:13)**: You, you, you, you have to remember that for all the lack of airbags, for the fact that the brakes were made of wood, for the— they were actually. My family had a Ford T for a while. But, uh, 1929, not a 1910. But, um, for all the things that today we look at them and go, what? You know what, the brakes still stop the car, the engine still pulls the thing, the wheel still turns the, turns the car around. You are probably going to have to drive differently. And so I'm going to say, you know what, instead of a Ferrari, let's talk about, of a Model T and a Tesla. The Tesla is going to do a lot of stuff for you, but it's going to do the same things that the Model T did. Because what you're looking for is the base functionality. You want to get from point A to point B.

**Matt Coles (47:09)**: I'm gonna say the car analogy is the wrong analogy here.

**Chris Romeo (47:12)**: Probably.

**Matt Coles (47:12)**: It's the motor vehicle analogy.

**Chris Romeo (47:13)**: We just spent 5 minutes on it.

**Matt Coles (47:15)**: You did, but I think you're going the wrong way.

**Chris Romeo (47:17)**: I think you're nice and well defined. Oh, we didn't see the sign.

**Izar Tarandach (47:20)**: We didn't see the sign.

**Chris Romeo (47:21)**: Yeah.

**Izar Tarandach (47:21)**: Yeah.

**Chris Romeo (47:22)**: Well, road signs. Wrong way.

**Matt Coles (47:23)**: Road signs or motor vehicle driving laws and regulations and conventions are probably the better example. And I'm touching upon that because as we look further down the article, I know we're running short on time on this episode, but it talks later about how the threat model doesn't have a control to address prompt injection. Well, that's great because that's not STRIDE's job, right? STRIDE's job is to find the threat. STRIDE's job is not to define a control. The control comes from you identify a threat and then you figure out what you're going to do about it.

**Chris Romeo (48:08)**: Yeah, true.

**Matt Coles (48:08)**: And so you have conventions and you have standards and you have guidelines about how do you deal with certain threats. STRIDE doesn't tell you that. That's, that's one of my, that's one of my gripes about Stride is it expects an understanding of a lot of knowledge around the threats to be able to address them.

**Izar Tarandach (48:28)**: That's the whole think like a hacker thing that we—

**Matt Coles (48:30)**: and if you, and if you don't understand, but also think like a computer scientist or think like a developer or think like somebody who knows how to read API standards, uh, API specifications, you have to understand the technology you're using in order to address a threat. Against that technology. So if the AI model has unexpected output and you identified that in your threat model, but you don't know how to fix that because you don't understand how the technology works, that's a different problem than Stride is broken. That's you as a technology deployer don't understand the technology you're deploying.

**Izar Tarandach (49:13)**: The, the, the fact is that I think that we, we are fast approaching the rise of, uh, if it hasn't yet and I just don't know about it, the rise of a new, uh, uh, profession, which is the, the AI psychologist, right? Because no, seriously, what, what Matt just, just put down that, you know, the LLM could give you, uh, I won't say bad. I won't say inexact. I would say a different kind of output than you were actually expecting. Now, between you and what generated that output, there are a lot of boxes. Some of them are gray, some of them are white, some of them are black. The black ones, at some point, somebody's going to have to rise up and say, I understand what's happening there, and I can explain to you what you have to do in order to get less of that. It's going to be an AI psychologist, you know.

**Matt Coles (50:13)**: Is the psychologist for the AI or for the person using the AI?

**Izar Tarandach (50:16)**: I think that's for the person using it. We are on the psychiatrist step already.

**Chris Romeo (50:20)**: But, uh, so it's for the AI.

**Izar Tarandach (50:24)**: It's for the AI. I mean, interesting. It's what Matt's saying. You know that something can go wrong, but you don't know what to do about it. And all of a sudden, when you had code to look at and you could exactly trace, and now I'm going to talk a bit against myself, when you had the code to trace and you could figure out what the path was, then you could go back on it and say, oh, here's the bug, I'm going to fix it. Now that you have the LLM, and I'm not talking about what happens outside with the result of the LLM, later on. That part I have established my position quite forcibly. But inside the black box, the non-explicability of what's happened inside to get to that result, right? That's the bit where you can't go back and find, oh, oh, oh, this weight is the one that caused the distance between the thing and the other thing. And that's what caused the output. You don't have that that level of explicability.

**Matt Coles (51:28)**: Right, we don't have that, right? We don't know how to do that yet.

**Izar Tarandach (51:32)**: Exactly.

**Matt Coles (51:33)**: Because it's non-deterministic also.

**Izar Tarandach (51:37)**: And it's complicated math.

**Chris Romeo (51:39)**: But—

**Matt Coles (51:39)**: Yeah, we don't know how many dimensions there were involved and number of parameters and all that sort of stuff, right?

**Izar Tarandach (51:44)**: Exactly, exactly. And when we do, we forget because it's a lot, right? But what I'm trying to say is that In between that, what could go wrong and what can we do about it? If you're not able to trace back to the root cause of what went wrong, it's going to be very hard to say, this is what I can do about it. So people come up with all kinds of fences and guardrails and whatnot that go around that, that black box bit and hope for good things.

**Matt Coles (52:12)**: Yeah. I have a thought exercise for, for, for listeners and for you guys as well, because you may have forgotten it. When you use C or C++, actually there's other languages. No, no, no, no, no.

**Izar Tarandach (52:26)**: Don't go there. C, C, keep it C. Okay.

**Matt Coles (52:28)**: All right. I'll keep it C. There are, there are things in C that have undefined behaviors.

**Chris Romeo (52:36)**: Yes.

**Matt Coles (52:36)**: Did you throw away using C because of undefined behaviors?

**Izar Tarandach (52:41)**: No, we defined the standard.

**Matt Coles (52:44)**: And when you had an application that exhibited an undefined, or sorry, when you had an application that used functionality that may have fallen into the bucket of may exhibit undefined behavior, what'd you do?

**Izar Tarandach (53:00)**: I called it elegant.

**Chris Romeo (53:05)**: Elegance.

**Matt Coles (53:05)**: But this is, this is the exact same. This is a similar problem, right?

**Izar Tarandach (53:08)**: This is exactly the same thing.

**Matt Coles (53:09)**: You, you have, you have undefined behavior in the technology and you account for it.

**Chris Romeo (53:15)**: Right.

**Matt Coles (53:15)**: That's Stride. May, Stride may point you in that direction, but Stride will not tell you how to address that undefined behavior. Right. You just have to deal with it. And as an application developer in the past, you would've said, well, for X number of boundary cases, something bad is going to happen and we're going to, we're going to 6/7 crash, right? We're just going to, we're just going to bail because there's nothing else we can do.

**Chris Romeo (53:39)**: I thought you said 6/7 crash for a second. But it was SIGSEV.

**Matt Coles (53:45)**: SIGSEV.

**Izar Tarandach (53:46)**: SIGSEV, yeah. You know, now you're talking about levels of behavior, right? Sorry, levels of bad behavior. And one thing is to have a memory exception and end up in a SIGSEV. And another one is having your model say that all of a sudden, I don't know, Something ridiculous that has been said lately. So many options. So many options.

**Matt Coles (54:16)**: Mathematically proven, something complicated.

**Izar Tarandach (54:19)**: Exactly. No, something as simple as, yes, I don't know, that group of people are responsible for all the problems in the world, right? So what went wrong here? For one, in one side is, hey, you try to write in a memory area that you don't have a bit that says that you can write. So I'm gonna raise my hands up and, and bail out. The other one is, well, you know what, the thing got trained with a lot of texts and a lot of input from different social media, and it got the impression that this was a fact, but we're not quite sure because, you know, that there was just this amount and that amount, but then it went through reinforcement learning and the person who was reading the thing decided, hey, you know what, this is a fact. So yeah, yeah, yeah, sure. Why not? Go ahead, that's a good answer. And it got into the model somehow. So the, the, the, the factors that could influence that, uh, bad behavior, they are so dispersed and there's, there's such a pipeline of them. And then there's the unexplainability and, and undeciblity. I can't with this.

**Chris Romeo (55:28)**: Wow.

**Matt Coles (55:29)**: All right. You should probably end it there before Izar goes completely.

**Chris Romeo (55:33)**: Apoplectic.

**Izar Tarandach (55:35)**: I can't say that one too.

**Chris Romeo (55:37)**: Oh yeah, that's one of the words I can say, so I just go with it when I work.

**Izar Tarandach (55:41)**: It always comes apoplectic.

**Matt Coles (55:43)**: If you could, if you could spell it, that would be even better.

**Izar Tarandach (55:45)**: But yeah, so I, bottom line, I think that, you know, to go back and declare that, that the old ways of looking at what's out there are wrong No, it's dead. The old way, the old way of looking at new things is dead. It's a bit premature. It may show a bit of lack of understanding of the old thing itself. And the, the race to create new things just to, well, let's put it as ulterior motives to pop on social media.

**Chris Romeo (56:24)**: Yes.

**Matt Coles (56:24)**: There's more to talk about this article, but let's leave it there.

**Chris Romeo (56:27)**: What, what? No, no, no, no, no.

**Izar Tarandach (56:27)**: Talk about this article. No, no, no. I'm not talking about this art school. No, definitely not. Definitely not. No, no, no, no, no, no. Please, please, please don't misunderstand me. But other things that we have seen that run to, you know, create these new paradigms and new approaches. And at the end, they end up doing the same thing.

**Chris Romeo (56:48)**: Well, people are trying to get famous. Yeah, but you don't get famous by just, you know, reciting an old thing.

**Izar Tarandach (56:53)**: True. But you can get famous by pointing out that you know, you can look at something in a different way without dismissing everything that came before.

**Chris Romeo (57:08)**: That's, that's a whole other episode right there. That is a bit of a phenomena. You've just, you've just hit on a phenomena of that, that seems like it's in the last 5 years or so that to create something new, you have to break something else down. Is like a pathway that people are approaching fame. And not the fame in the 1980s NBC television program. Not that.

**Izar Tarandach (57:37)**: I'm gonna leave forever.

**Chris Romeo (57:39)**: I knew it was in your head, so I had to draw it out.

**Izar Tarandach (57:41)**: I can't, I can't, I can't.

**Chris Romeo (57:43)**: All right, with that, we can no longer be here because we don't— we've already used our one 1980s television theme song. That we could sing. And if we let another one sneak out, the FCC will shut us down. So we have to—

**Matt Coles (57:58)**: it's fair use at this point.

**Chris Romeo (58:01)**: I like it.

**Izar Tarandach (58:02)**: The way that you are saying means that anything automatic listening to this thing can actually recognize the song that I am singing, which I'm going to take as a compliment.

**Matt Coles (58:10)**: Yeah.

**Chris Romeo (58:11)**: No, I think we knew what it was. It worked.

**Izar Tarandach (58:13)**: Oh, no, no, no. You baited that one. Come on.

**Chris Romeo (58:17)**: Of course I did. That's what I do. Folks, thanks for hanging around talking about npm changes, strides forward in the future and whatnot. And when it comes to AI and what—

**Matt Coles (58:32)**: and Model Ts.

**Chris Romeo (58:33)**: And Model Ts. Yes.

**Izar Tarandach (58:34)**: Where are we going? We don't need stride. Yeah.

**Matt Coles (58:40)**: Or brakes.

**Izar Tarandach (58:42)**: Wooden or not.

**Chris Romeo (58:44)**: We do need to refactor the flux capacitor though. And I'll leave leave you with that. As we say, adieu. Adieu. Yes, we do.
