42 minSeason 3, episode 12

Vibe Startups, AI Problems, and Matt’s Precious Computer

With Chris Romeo, Izar Tarandach, and Matt Coles

AI security & autonomous agents

We’re talking about the rise of "vibe startups" - entrepreneurs hunting for problems to solve rather than building solutions from personal experience. We chat about AI security challenges, questioning whether these are truly new problems or just old security concepts repackaged for the AI era.

Episode title image: Vibe Startups, AI Problems, and Matt’s Precious Computer

Listen to the conversation

About this episode

We’re talking about the rise of "vibe startups" - entrepreneurs hunting for problems to solve rather than building solutions from personal experience. We chat about AI security challenges, questioning whether these are truly new problems or just old security concepts repackaged for the AI era. From prompt injection and guardrails to the scary reality of AI agents acting as humans, we examine whether the industry's obsession with AI is leaving traditional security gaps exposed.

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 · 42 min conversation

0:09Chris Romeo No, you're just good.

0:15Izar Tarandach Go, go, go.

0:18Chris Romeo Hey folks, welcome to another episode of the Security Table. You got to see a little bit of behind the scenes there as Isaro was preparing himself. And, uh, was having a potential bad hair day, turned it into a good hair day.

0:31Izar Tarandach Monday is always a bad hair day.

0:33Matt Coles Well, wait, look in profile. Is there a turn profile?

0:37Izar Tarandach I can't look on profile. Why don't bend that way?

0:40Matt Coles He's got a little top going up.

0:43Izar Tarandach Ah, it's a fashion. It's my team team. It's my team team.

0:48Chris Romeo Yeah, he's, uh, so it's— you as the audience could have heard the the nerd fanboy session that just happened here about Matt's new computer. These are literally— get into a Gollum impression from The Lord of the Rings or The Hobbit, wherever you want to pull a Gollum from, uh, for Smeagol. And, uh, literally kept referring to Matt's computer, new computer, as his precious and how it needed to be given to him now.

1:19Izar Tarandach It's got a nice graphics card.

1:21Izar Tarandach It's got memory.

1:23Matt Coles I can play with it.

1:25Chris Romeo There was a lot of nerd envy happening here just a few short moments ago.

1:30Izar Tarandach I can play with it.

1:34Chris Romeo Probably gonna happen for the entire episode, so we apologize in advance. Every 5 minutes that, uh, he starts—

1:42Matt Coles if he brings out the flight joystick again, we, you know, we're— it's gonna be a fun episode.

1:47Izar Tarandach It can do an infinite loop in no time flat. Yes, it can.

1:50Matt Coles Oh, I bet that's— It can probably melt the loading now. It will heat your office too.

2:00Izar Tarandach I have actually, I have to fan, to fan Boaz on something that I found this weekend and I cannot believe how I didn't see that before. And it's just a recommendation.

2:13Izar Tarandach Wait. Oh my goodness. Of course, I'm—

2:18Izar Tarandach oh, here it is. Here it is.

2:19Matt Coles Here it is.

2:20Izar Tarandach So it's a YouTube channel by a guy called Kai Lintic, right? And he's got a series of interviews with. So it has an interview with a DevSecOps person in 2025 that's going to have you like, rolling, and from there it just gets better. So yeah, so it's Kai Lentet. I don't know how I never found this before.

2:49Matt Coles You know, the DevSecOps jokes just—

2:51Izar Tarandach No, that's the thing.

2:54Izar Tarandach That's the thing.

2:55Izar Tarandach It's not—

2:57Chris Romeo you got to watch it. Like, yeah, but our future late night talk show is going to get canceled as a result of you making fun of DevSecOps people.

3:08Izar Tarandach It's not that it's not making fun. It's just like, did you fact check?

3:14Izar Tarandach No, it's observational.

3:16Izar Tarandach It fact checks itself. There's a pipeline for this. Somebody wrote a GitHub action that actually does the fact checking for itself and it declared itself fact checked.

3:27Matt Coles We can't, we, we, all we have to do is, all we have to do is follow the breaches.

3:31Izar Tarandach Exactly.

3:32Chris Romeo All the bridges, come on. The year is 2025.

3:37Izar Tarandach Once more into the breach.

3:39Chris Romeo Into the breach. All right, Izar, set this up for us. You had a topic that you wanted to discuss.

3:45Izar Tarandach Yes, yes, yes, yes, yes. So an interesting phenomenon has been happening lately with me, and I wanted to share with everybody and hear your voices about it. So It used to be that I would get hit every once in a while by someone on LinkedIn saying, hey, I've been looking at your work and I like what you do and this and that and the other one. And we are a couple of entrepreneurs. We don't even have a startup yet. We just have this idea that we've been working on. We would appreciate if you took a look at it and give us your gave us her opinion. And first of all, it's great for the ego. Yeah, somebody listened to what I have to say.

4:29Chris Romeo Wow.

4:31Izar Tarandach But then it became an exercise in, you know, it's really cool to see how these people are thinking about these problems and different ways of solving them and see if I can give some of my experience to them and this and that and the other one. But then lately things morphed. Rather than coming with something that they've been working on, and that they would like an opinion on, it became something like, I really like your experience. I would like to have some time with you so that you can tell me your problems.

5:04Chris Romeo So, I mean, how much time do they have? How much time do they have?

5:09Izar Tarandach Apart from the fact that I already paid a series of professionals to do that.

5:14Matt Coles I was going to say, are these cyber psychologists? Have we turned the corner? We need cyber professionals to be cyber professionals.

5:24Izar Tarandach And you stop that because that's coming. That is coming. There will be people analyzing your prompts very, very soon and deciding what kind of damage you're doing to your LLM with your horrible, horrible prompts. But anyway, so these people, they're coming in and saying, we want to hear your problems. We want to find what's bothering you so that we can solve it. And I've been having this internal dialogue now that I want to take the voices out of my head and put here on the table where I ask myself a couple of things. So I was raised on the understanding that the best kind of way to get a startup rolling is to find something that's itching you new and scratch it, right? Because you're feeling it personally. It's your problem. You want to own that problem. You want to solve that problem.

6:23Izar Tarandach And by the way, let's make a couple of millions out of it, right?

6:27Izar Tarandach So you would have find this problem. It's really something that you can jump in. You can bite. You can dice and slice. Find a solution and commercialize it. That's how I've been in most of my startups, if not all of them. And I wanted to hear from you guys. Am I dinosauring myself here? Like, is everybody now—

6:56Chris Romeo They're vibe startupping. No, they're vibe startupping. Woo! They're creating startups by just asking ChatGPT for ideas of what problem should I go solve? They're just using you as the LLM in the short term.

7:14Matt Coles They replace the LLM with a Yzhar.

7:18Izar Tarandach I have not been that well trained.

7:20Chris Romeo No, I think Yzhar, I think you're 100% right. Like from, I guess, my own personal investment thesis based on my experience in starting two startups and taking them to exit, both of those companies were rooted in a problem that I was deeply invested into and that I had performed or, or I had tried to solve in other places first before I built a company. So Security Journey was, came out of my work at Cisco building a, a AppSec training belt program and then realizing, okay, I've got something here that could go to the rest of the world. So, I didn't just have an idea, I had it battle-tested in the enterprise. Approach and brought that into a startup. And then the same thing with DaVinci, having worked on threat modeling at Cisco and building a tool inside of Cisco many, many years ago, that led to the idea of DaVinci and saying, okay, what, how can I go after this same problem again? And so when I talk to a potential startup person who has an idea, has a company, and reaches out to me, similar to what they, what you, what you're going through here, I won't even consider investing in somebody who hasn't battle-tested their idea someplace else. Like if you don't have the experience and the passion for the problem, you're never going to be successful. And I can say honestly, like I went up against, with both of my companies, I went up against plenty of people who had an idea and a large check from an investor and I kicked their butts all the time. With very, and I had, thank you, Matt Stein, for confirming that for me. But I mean, I had very little, I had no investment. So I bootstrapped both companies from the ground up. I was going up against people who had millions, tens of millions, sometimes hundreds of millions of dollars in investment. I still kicked their butt at a good percentage, even though they had all of these resources. And I think that my unfair advantage was I had real experience with the problem, whereas they had an idea they thought they could make money with.

9:33Matt Coles So that, look, if you connect your earlier idea, your earlier joke of Vibe Startup, right? A lot of it seems like you're saying is they had an idea of something that somebody might wants. And then effectively, and this is probably, they were doing this probably pre-LLM, but now, well, I have an idea. There's research already available for me that I can go leverage and then make a company out of. So they're not really, they're acting as somebody who takes action to build the company and manage it. But not the idea person.

10:17Chris Romeo Yeah. And they don't, they're not, they weren't rooted in the problem, which I think was really my unfair advantage was that I was rooted in the problem and had to solve the problem in an enterprise environment first before I thought, hey, this could become a company. And so I think a lot of people, I think the other challenge that we have here in the market right now is Startups are cool. Let's just be honest. Everybody wants to say, oh, I have a startup. I'm leaving my full-time job to go build a startup. And I think it's great. I think everybody should do it in their career. Isaur, you've got some experience doing it as well. And I don't know if you'd agree, but what you learn in those experiences are things that a textbook cannot teach you, a classroom cannot teach you. Having to solve the problem, a challenging financial problem.

11:07Izar Tarandach In the end, I learned not to do it again.

11:10Chris Romeo That's an important lesson too, though, right? Like, so I think everybody wants a startup. Everybody thinks it's become very vogue to say that you have a startup, you're working on your own company, which I think is good. But the challenge is there's going to be a large percentage of people who aren't really invested in a problem. They're just trying to find a problem.

11:34Izar Tarandach And that's strange to me because the experiences that I had with VCs, I don't know if I was just lucky, but I hear from a lot of people similar experiences. The VCs that I worked with, they very quickly became engaged with what we were doing. They liked the idea, they liked the people executing, and that led them to put money on the table. But then they were part of the execution, part of the, uh, the, uh, orienteering of the startup. It could be because we were so, so green and they felt that they needed to, but I, I could feel that they had personal investment beyond the money into it.

12:15Chris Romeo Well, you don't, you never invest in ideas, right? You invest in a team of people that you believe in because they may have to morph the idea. They may have to change the idea into something else as you get rolling down down the pathway. And so the investment is something that really goes into a group of people because you believe these people could be successful. They have markers that make me think no matter what happens, no matter how we adjust the startup's scope, they're going to be successful.

12:47Matt Coles So can I— so my experience with startups is a lot more limited than you guys have, but It seems so back in the '90s, I think everyone was doing startups. You had an idea person, you had a techie idea person who wanted to drive an idea forward. They had a problem to solve. They knew how to solve it. They wanted to make it happen, but they needed leadership to make it happen. So that was the shift. The startup approach there was somebody finding a leadership, a management or leadership team to help form a company, do the pitching, do all the Yeah. You know, legwork and everything, but also build the thing. It seems now there's, there's a leadership group with no thing to build. Right. And so there's a leadership team looking for a techie and they look and look for the techie either in problems like industry reps, industry experts have, uh, or vibe. Writing it out of an LLM, it sounds like. Is that a, does that make sense? And does that fit with what you're seeing here?

13:58Chris Romeo I think a lot of people are doing that or are using LLMs as idea generation because there's not as many, there's not as many fresh ideas available at this particular point. If we think about specifically in application security or even broader into cybersecurity, other than the impact of AI and LLMs to Isar's original problem about, you know, where are the itching problems at, it's really the only thing that's changed in the past few years that I can think of that you could base a company around. And we already saw a big land grab happen over the last 12 months. As everybody tried to become the AI version of this or that. And you've got all these startups that are doing, you know, PR generation, the AI is fixing your bugs for you, the AI is, you know, fixing the vulnerabilities. But I would say the only real area that I've seen that's an itching problem is securing the AI itself. Yeah, exactly. But that's the only area where there's been a big opportunity, at least in our world, right? Like, we're very constrained in this. When we're talking about startups, we're talking about— we're not talking about consumer products, right? We're talking about application security, because I only stay in the lane that I understand. I don't, you know, if I had a nickel for every time somebody sent me a pitch deck for a pot business of some form, I'd be able to buy at least a small soda somewhere.

15:35Matt Coles Is pot making money still? I don't think it is anymore. I don't know.

15:38Chris Romeo But there was plenty of pitch decks coming around looking for investors in various marijuana distribution businesses and stuff. And it's like, you know, I just don't, I don't invest in something I don't understand. So I only invest in—

15:51Izar Tarandach It would be a high risk. Yeah.

15:55Chris Romeo For chicken. I'm gonna leave that alone.

16:00Izar Tarandach Yes.

16:01Chris Romeo Very high. I mean, Cheech and Chong levels.

16:08Izar Tarandach So, well, okay.

16:10Izar Tarandach So the issue problem now is securing AI. Apart from our AI, what are the solutions for securing AI?

16:19Matt Coles Oh my God.

16:20Chris Romeo Really?

16:20Izar Tarandach Let's give people some ideas here.

16:21Chris Romeo Perhaps somebody's going to.

16:24Izar Tarandach All right.

16:25Chris Romeo So I think it's a guardrail. I think one of the things that I was looking at what AWS was doing from a guardrail perspective on their LLMs. Which they have a service that does it. I think there's an opportunity for some guardrails that can match the innovation structure and yet happen in a separate security-focused company perspective.

16:51Matt Coles Guardrails are great. Guardrails are— obviously, we like guardrails and paper roads, right? We've talked about that on the show on multiple times. But I think on the AI side, really traceability and observability, being able to get insight into its action is going to be, and is an area of active research, and that's an area that needs a solution.

17:16Izar Tarandach Wait, wait, wait, I have to ask here. Observability and traceability inside the LLM or around the LLM?

17:26Matt Coles I think both, but it's certainly within because we don't know what come— we don't know, we, I think we still don't know how exactly the transformer architecture works, right?

17:38Chris Romeo That's gonna require the people that create the LLM for the traceability and observability of the innards of it to expose.

17:47Matt Coles Or you're doing some sort of side channel or reverse engineering or something to that effect, right?

17:53Izar Tarandach Let's get a bit more philosophical here. Do we actually need the LLM to stop acting as a black box in order to be able to secure it? Or is it enough to get some form of determinism and say, here goes my input, that's the output that comes out to a certain extent because we know that it's probably not going to be the same one every time. But is that enough to put guardrails in place or do we need to implant things in the brain itself? In other words, where does the guardrail leave?

18:30Matt Coles Yeah, I think, thinking, I mean, thinking this through, if you had some idea of what the possible outcomes were, not just what you expect, but what can actually come out, And as you said, you know, we don't necessarily, we can't know because it is based on probability. Like it is, right? And it's the fundamental algorithms, if you knew so many of them. So we know, we can only know sort of, we should only know what the possible outcomes are because of what the training set is, but that's obviously proving difficult for folks at times.

19:08Chris Romeo I think we're, I think we went about this, we jumped right into the technical deep end. Instead of doing what the people were asking you to do, which was, what are the problems you have? So I think if we tell, if we step back and jump up 20,000 feet so that we're looking out the window of the airplane and start to describe like, what's the problem that led us to that? Well, certainly prompt injection is one of the things that led us to discuss, and that's what got me to guardrails, right? Is people are putting attacks into the prompt and the LLM's not smart enough to be able to recognize that it's being attacked and so it does some illegitimate action.

19:48Izar Tarandach Look, again, again, again, you give a 10-year-old the nuclear football, tell them to sit on top of it and no matter what, nobody touches that thing, your guardrail to the nuclear football is still a 10-year-old.

20:06Matt Coles Here, here's a hot dog. Go play.

20:08Izar Tarandach Here's a hot dog. Here's an ice cream. You will always find the one thing that will turn their head away. So the problem is not in the fact that, oh my God, I can convince the 10-year-olds to let me get the nuclear football.

20:21Izar Tarandach The problem is that you gave it the nuclear football in the first place.

20:27Chris Romeo Yeah, but that ship's already sailed.

20:28Izar Tarandach No, no, it hasn't. It hasn't. It hasn't. The big problem here is not proper injection. It's the accessibility.

20:35Chris Romeo So, but if you think about in your, in your illustration here, the football, the football you've already given it, or the nuclear bomb, you've already given it. The, the companies are already giving LLMs access to their, all to their data inside the company. Like, if nuclear—

20:50Izar Tarandach And if everybody jumps, if everybody jumps from mass computer, will I jump as well?

20:54Izar Tarandach Probably.

20:55Izar Tarandach But, uh, but the, the, the fact is, We are again, and this is my pet peeve and I'm going to die on this hill. AI is bringing all the old problems up again, repackaged, and we are looking at them with a very surprised Pikachu face.

21:15Izar Tarandach Oh my God, we are seeing the same things that we've been seeing for the last 50 years.

21:20Chris Romeo Explain that in more depth. So yes, injection has been a problem for 1,000 years, maybe 100 years. But what else? What else are you saying?

21:29Izar Tarandach You look at enough pyramids and you're going to see a hieroglyph saying, forget everything that you saw before. But even excessive agency, confused deputies, all that good stuff that people are doing some very, very good work on writing top 10 lists and secure this and secure that on AI.

21:50Izar Tarandach Right.

21:53Izar Tarandach I got one for you.

21:55Chris Romeo I'm sorry.

21:56Matt Coles Time of check, time of use. Exactly.

22:00Izar Tarandach Right? People are going there and say, oh my God, this thing is not deterministic. Yeah, what can I do?

22:07Izar Tarandach We've been throwing coins for thousands of years. So, you know, don't come at me and say, which we found this new technology and it brings new problems with it. I would love for it to bring a new problem. So far, I haven't seen it yet, I think.

22:24Chris Romeo So the question was not, uh, what causes these are to go off on a rant. Uh, the question was, what are the problems that, uh, what are the problems that people can use as an inspiration to create a startup today? And so, There, and that, that was why my point was there are challenges in the AI space that people are trying or could try to solve right now.

22:54Izar Tarandach There are. No, no, no, no.

22:55Izar Tarandach What else? I'm not saying that there aren't. I'm just saying that there are, okay. There are old problems requiring new solutions. That, that's what I think I was ranting about.

23:02Matt Coles Okay.

23:03Chris Romeo And so there's an opportunity there for somebody to try to solve that problem in a unique way. As a startup today that classic tools aren't going to solve. Like, you can't just go buy a solution that's been around for 10 years that's gonna solve your agentic AI security challenges with MCT.

23:24Izar Tarandach There you go.

23:25Matt Coles Well, actually, is, so does agentic make it worse or better?

23:30Izar Tarandach Yes.

23:32Chris Romeo You mean, you mean from a, Security, you're talking about from like a security implementation perspective?

23:41Matt Coles So, so AI is the monolithic application, right? With a database. And then you have Agentic, which is, which is microservices effectively, right? Where you have different processes talking to each other.

23:54Chris Romeo I think that's, I think that's scary. I saw, so ChatGPT or OpenAI launched their kind of full-on agent thing. and it was described by them as like the agent has its own browser and it's operating on its own off to the side. And I'm like, this doesn't sound like something that ends well.

24:14Izar Tarandach Okay.

24:15Izar Tarandach Okay. So, so I'm gonna go there before everybody, anybody else goes there. The moment that the agent has its own browser, I wanna see how long it's gonna take for somebody to say, hey, solve me this differential equation. It's going to go, I can't now.

24:26Izar Tarandach I'm busy.

24:27Izar Tarandach I'm busy.

24:28Izar Tarandach Give me 10 minutes.

24:29Izar Tarandach I'm busy.

24:32Matt Coles So, so if that happens, I'm, I'm sure it will and we'll hear about it soon. But real sol— real problem to solve, I guess. If the, if the agent is using a browser and we know, talking, thinking about your, your guardrails, you know, discussion there, we know that browsers run on things like JavaScript engines. Right? And so, but these are, these are for human behavior and, and, you know, challenges with standards that are sometimes either in flux, not well adopted, not well implemented, you know, questionable or undocumented behaviors. It sounds like a problem to solve would be to have AI-aware browsers that the AI agent can use that can naturally address some of those problems, right?

25:33Chris Romeo So, so it could construct, it's like a sandbox browser almost?

25:37Matt Coles To something to that effect, yeah. So that the, so that that browser doesn't, so if the, if the agent knows how to use a browser, but the browser cannot do, you know, SSRF or, you know, CSRF or whatever. Then you're not going to get those problems. Today, we haven't really solved that in, in human use, but humans can't attack at scale as an individual.

26:07Chris Romeo Okay. So is there anything else?

26:09Matt Coles That might not be a well-baked idea, but at least it's something.

26:12Chris Romeo No, I think it's, I think that's a, that's an interesting area that somebody could dive into, but I mean, is there anything outside of AI that you guys are seeing in our industry where there's, like, I can't think of a new category that's really hit in the last year or two. Like ASPM is a few years old at this point.

26:34Matt Coles Identity is a big thing. Identity. Because now with agents, with agents, because agents are non-human identities.

26:42Chris Romeo So like machine-to-machine identity?

26:44Matt Coles Yeah. And we've never had a good solution for this.

26:47Chris Romeo It's been around in the IoT world for a long time. It's just, it was such a, it was so behind the scenes. You were, it's only if you nerded out on it, did you ever think about MTAM authentication and identification?

26:59Matt Coles But, but, but that's a thing, right? But, uh, that's, you still have to do device, you still have to issue identities to these things. Agents, agents don't come with hardware root of trust, whereas IoT can.

27:12Chris Romeo Super important question though is, Matt, what type of identity does that new computer have that you have? Anything Yzhar's gonna be jealous about? Like, does it have a certificate? I don't know.

27:23Matt Coles Yes, it actually, it does. Yeah, yes, it does. Probably has multiples actually.

27:30Chris Romeo Uh, okay. So identity, like what's, well, what's the problem in identity other than MDM? Like, is there anything else that's new in that space?

27:38Matt Coles Well, so, it's not just machine to machine. Agents are acting as humans. So you still have this delegation impersonation problem.

27:48Chris Romeo Boy, that's a scary thing. If you— we're so engrossed in AI that that is not a scary thing anymore to say. But like, I still hear that and I'm like, wait, what?

27:58Izar Tarandach What?

27:58Chris Romeo Like, agents are acting as humans? Hold on.

28:01Izar Tarandach No, please.

28:02Chris Romeo That's the reality we live in though.

28:04Matt Coles And oh, by the way, there is a potential solution for this. I'm gonna drop, take out a 40-year-old technology to say Kerberos can probably solve this problem.

28:16Chris Romeo Oh, bite your tongue.

28:17Izar Tarandach What? No, no, no, no, no, no, wait, wait, wait, wait, wait.

28:21Matt Coles Trusted third party as an arbitrator.

28:24Izar Tarandach Wait, wait, wait.

28:24Izar Tarandach As long as Microsoft AD is out there, you cannot say that Kerberos is a 40-year-old technology, because AD is still out there.

28:33Chris Romeo So, Kerberos is before AD.

28:35Matt Coles Kerberos is built into AD.

28:37Izar Tarandach Exactly.

28:37Izar Tarandach Yeah.

28:38Izar Tarandach Because if you look at how AD evolved, it was one more of those embrace and expand.

28:43Chris Romeo Now AD is now referred to as like Azure Virtual Sunshine or something like that.

28:48Izar Tarandach Oh yeah. For sure.

28:49Matt Coles For sure. I'm going back to the original Kerberos token.

28:52Izar Tarandach And it has 3 compilers.

28:53Chris Romeo Let's let Matt make his point here, even though we both concluded very quickly.

28:59Matt Coles So it's simple. It was really, I mean, the concept, okay, even if it's not the implementation of Kerberos, I, I, having a trusted third party that can issue identities on the fly and use that to authenticate to other services helps with this notion of the agents being their, so this, I mean, Kerberos would work with machine-to-machine as well, right? How do I know you're a thing? Well, I can trust you because you provide information to me and I'm a service, you know, can, you can extend trust as much as you like.

29:33Chris Romeo And you can revoke it because it's just-in-time access. You can revoke, if you revoke the agent's access at the, at the authentic authorization service, it would be dead in the water immediately. It would sound like there's a, it's not gonna last. It can't keep rolling for a day. It's immediately dead.

29:52Matt Coles You could, I mean, and again, this comes down to the implementation. We know that Kerberos, there's a time to live for the, for the tokens, but if the, if the If you take a zero trust approach, oh, this thing gave me this token because, you know, you Kerberos server granted it, should they still have access and make that ping, right? We now have more rapid network communications. We're not—

30:14Chris Romeo Any chance we could write this to the blockchain and just really complete the circle here? Oh God. A little quantum flow. I'm just, I'm just, I'm playing buzzword bingo now. I couldn't, I couldn't resist because you said AI and you said, what was the second one? Kerberos? The category. Yeah. So you just, you hit a bunch of buzzwords and it was like, oh, if we could get blockchain and quantum into this, like you probably have 10 million investors.

30:47Matt Coles You could use blockchain to guarantee the states of the systems haven't changed. that's probably overcome.

30:54Chris Romeo It seems like blockchain—

30:55Matt Coles Oh, zero trust was the other one that I—

30:57Chris Romeo Zero trust. Zero trust. Yes, you made AI to zero trust, and I'm like, we might as well just keep going blockchain, quantum, anything else we can chuck in there. All right, so Yezar, did we solve your challenge? Do you feel enlightened?

31:15Izar Tarandach I feel like we just proved that we've been Putting AI into so many things now that everything now is an AI problem.

31:23Matt Coles But it's not. It's just a different implementation of a computing system.

31:32Izar Tarandach True, but now instead of solving those problems that were there before AI, now we move those problems. Because we put AI into everything, we move those problems to be AI problems. So everybody's solving AI problems.

31:46Matt Coles Actually, no.

31:46Izar Tarandach Some people are solving AI problems that I can't even see the problem.

31:49Matt Coles I'm gonna, there's a bit of an argument maybe to make here that it's, it isn't just a technology problem. Think of a, especially think of agents as humans in the box. So instead of just having humans in the loop, you now have humans in the box. Intelligence in the box, right? They act with probabilistic tendencies, some sort of like humans, maybe less deterministic, maybe more so, right? They're just algorithms and they don't have the same thought process as humans, but they don't have, they don't function as processes that are algorithmic. right? And so you now have a human equivalent in the box making decisions and performing operations. So, you know, the human problem at the browser or at the, you know, behind the API now goes into where the agents are. The humans are now in the box.

33:01Izar Tarandach Do you like to delete your database in prod?

33:03Izar Tarandach Because that's how you delete your database in prod.

33:05Chris Romeo It actually just happened.

33:06Matt Coles Jason Lemkin has a new startup.

33:08Izar Tarandach Okay.

33:08Chris Romeo I figured out the ultimate startup idea based on what Matt just said. We need security awareness for AI agents. We make them watch a bunch of videos and then that's a company right there.

33:20Izar Tarandach Do they get their own spacebar?

33:24Chris Romeo And yes.

33:26Matt Coles And cyber parapsychologists.

33:29Izar Tarandach Oh, why wouldn't it be an AI psychologist?

33:33Chris Romeo Cyber pair. Why does it have to be pair?

33:35Matt Coles Like, I don't know. Because they're not, because they're not humans, right? So they're, they're, they're sort of semi, they're human, human computer hybrids, right? So.

33:43Chris Romeo I'm going to laugh if in 6 months I see an announcement for a security awareness company for AI. Somebody's like, somebody tunes into this and, and they're like, that's an idea. We could, we heard those, those dopes on that security table show talking about this.

33:56Matt Coles If you need some advisors, uh, technical advisors, you know, they, We do stock.

34:05Chris Romeo It's just that I want to bounce, I want to bounce another idea off you guys because it hit me and it has nothing to do with the startups. It does have to do with the AI. I'm curious about your opinion on this because it just may have just had an epiphany. Probably not. I may have.

34:20Izar Tarandach Are you sure that you want to record this?

34:22Izar Tarandach Yes.

34:22Chris Romeo It's not a startup idea. It's the question I have is with all of this focus in the industry on AI and LLMs now. Is it easier for attackers that are just doing boring things? With everybody so hyper-focused on the AI LLM side, are we somehow leaving the classic interfaces open more because we're not focused on—

34:50Matt Coles Look, I have an idea for you.

34:52Izar Tarandach Hold on.

34:53Matt Coles I have an idea for you. It's not a good idea, but it's a, it's, that's not a, it's an evil idea.

34:59Izar Tarandach Tell us, tell us, tell us.

35:02Matt Coles What if you had an attack platform that could hide in the hallucinations? In other words, if, if the AI is already hallucinating or you're getting very—

35:12Izar Tarandach been watching Inception during the week.

35:14Chris Romeo It was like an Inception moment, hiding in the— wait, am I asleep right now?

35:20Matt Coles But, but seriously, if, if, if you had an attack platform that was aware that was aware of the AI behavior and you, and you knew somebody was issuing a prompt or was doing work on as an agent and, you know, there's output. What's the likelihood that the user who's following the prompts is going to notice the side channel attack or the, or the intrusion into the network?

35:46Izar Tarandach You, you know that you are talking about a lot of people who say that, and I, I joked about it earlier, but A lot of people say that the solution for prompt injection is probably somewhere near putting an AI in front of the AI.

36:01Matt Coles Hasn't that already happened?

36:03Izar Tarandach Yes. And, and I think it was funny, but yes.

36:07Matt Coles Isn't that, isn't that the, isn't that the fox guarding the henhouse problem? Yes.

36:13Izar Tarandach Uh, I found it funny that some people did not see that as a problem, but actually as a solution. But, uh, anyway, uh, you guys know what I think about property injection.

36:24Izar Tarandach But, uh, go, go, go back to the original idea. Uh, what was it, Chris?

36:34Chris Romeo So the original question was before Matt's evil hallucination inception, Cheech and Chong moment.

36:40Izar Tarandach Right, right, right. No, no, no.

36:42Chris Romeo Yeah, that's the, the, uh, our, our, and our classical enterprise interface is more at risk now because everybody's got the— all vendors have moved their focus to AI, to focusing on doing things in the AI space. If I'm an attacker, do I have an easier chance now because people aren't paying as much attention to the old school stuff that is really the foundation of the enterprise?

37:07Izar Tarandach You know the story about the drunk guy looking for his house keys in the streets, right?

37:13Chris Romeo No.

37:13Izar Tarandach So the, uh, uh, there's a police officer walking down the street at 3 in the morning and he sees a guy under a, a, a post, a light post, looking at the ground. And he goes to him and notices that the guy's inebriated and says, what you doing? And the guy says, I'm looking for my car key, for my house keys that I dropped. And the guy says, okay, where did you see them last? And he says, over there. And he says, but why are you looking over here? And he says, that's where the light is. So it's basically the same thing, right? I mean, everybody's looking under where the light is. Nowadays, the light is under— no, the thing that's under the light is AI. Does that say that everything else that we had before stopped working?

37:57Izar Tarandach Well, not really.

37:58Izar Tarandach There's a lot of people now who are actually— a lot of people are combining both things and trying to do the— make the AI look better at the places where the light wasn't before. And that's awesome. That's great. Hey, pattern matching and all that good stuff and finding behaviors and I'm all for it. But I feel that there is a concerted effort and apparently my very probably biased LinkedIn timeline shows the same thing that everything is AI. And if it isn't, it's AI adjacent, very, very heavily wanting to be the AI, right? Not in a good way, I think. We are, I think that at the bottom of the thing, I agree with you. We are looking where the light is and we are not looking at the other things and those things still exist.

38:56Matt Coles There's a magic trick joke in here somewhere. There's a hammer and nail A joke in here somewhere.

39:05Chris Romeo It's the, uh, I don't know, I still love that, uh, it's not Batman Begins, it's The Dark Knight, where, uh, the Joker's at that, that crime lord, uh, sitting at the table and he's like, so he's like, want to see a magic trick? I switched the pencil up, takes the guy's hand, slammed them into the table. Every time someone says magic trick, I just immediately in my head, I wanna say a magic trick.

39:34Izar Tarandach Okay.

39:34Chris Romeo Did we, did we come anywhere close to what you were going for here, Izar, what your goal was for this conversation?

39:42Izar Tarandach No, let's talk about Matt's computer.

39:44Matt Coles If we generated, if we, if any, if anyone generated ideas from our conversation, it would've been a miracle, but let us know.

39:52Chris Romeo Does Eezer dream of Matt's computer?

39:55Matt Coles That's the question.

39:58Izar Tarandach Uh, probably there will be something there.

40:00Izar Tarandach Yep. Yep.

40:01Izar Tarandach Stay away from my computer. That's a whole other conversation.

40:03Chris Romeo Matt's like, I gotta move and not tell Eezer where I moved to. Eezer's peeking in the window. Is that it? Is that the computer?

40:13Izar Tarandach Can you come out to play?

40:16Matt Coles Oh no, I have to get a cable walk.

40:19Chris Romeo Haven't seen a cable lock in 25 years. That's gone.

40:25Izar Tarandach I, I, I never understood cable locks.

40:28Matt Coles Because you didn't have an Isar who copied your computer?

40:32Izar Tarandach No, seriously, like it's this huge piece of metal with this huge padlock on it that connects to this little piece of aluminum plastic.

40:41Chris Romeo It's true.

40:42Izar Tarandach And like, if you just turn too fast and the laptop goes one inch to the side. That's 2.54 centimeters for you people in Australia.

40:53Izar Tarandach If it just goes a bit to the side, the thing pops and you go, oh, I broke something.

41:00Izar Tarandach Okay.

41:00Matt Coles Yeah. Or probably use scissors to cut the cable.

41:04Izar Tarandach It's almost like, it's almost like it's a theater of Security.

41:10Chris Romeo Don't even get us started on security theater. That's a whole other episode.

41:13Matt Coles Come on.

41:14Izar Tarandach It's the casual market is the stage and all those are actors.

41:20Matt Coles It's the casual, it's the casual deterrence.

41:23Chris Romeo That's a good place to end this right here. The, uh, with a teaser for a future episode on security theater.

41:29Izar Tarandach Yeah.

41:30Chris Romeo Well, thanks for your—

41:34Matt Coles yes.

41:35Izar Tarandach I don't know anything.

41:36Chris Romeo I just don't.

41:36Matt Coles I don't know, I can't answer your questions. You can use an AI for that.

41:39Chris Romeo True, true. The Bard LLM.

41:43Izar Tarandach I'm shaken.

41:46Chris Romeo Okay, well folks, thanks for tuning in to this oddly bizarre conversation about Matt's computer and the future of startups and AI and the rise of the machines and Yzhar's weird Gollum impression, which It's kind of the end of this conversation, so thanks for tuning in. We hope that doesn't give you nightmares.

Transcript supplied by AssemblyAI. It may contain transcription errors.

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