# The Agent Access Problem: When AI Has the Keys, Who’s Really in Control?

URL: https://securitytable.ai/the-agent-access-problem-when-ai-has-the-keys-who-s-really-in-control/
Published: 2026-03-25
Duration: 2902 seconds
Season: 4
Episode: 5
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

**In this episode, we dive into the messy reality of AI agents acting inside your systems and what that means for modern security. We explore the idea of agents as actors with real access—credentials, APIs, and permissions—and why this isn’t as new as it sounds (hint: it’s just applications all over again). We unpack where things actually get risky, from over-permissioned agents to unpredictable behavior driven by prompts, and why “it won’t go rogue” might be missing the point entirely. We also question the growing hype around AI governance, whether security teams are actually gaining control or just making more lists, and what happens when agents start talking to each other… and running up your bill. Per usual, the conversation is filled with sarcasm, skepticism, and a healthy dose of “maybe just add parental controls.”**

Follow AI Security Table:

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

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

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

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

## Transcript

**Chris Romeo (0:12)**: Hey folks, welcome to another episode of the Security Table. This is Chris Romeo, joined as always by Matt Coles, Izar Tarandash, the doctors of AI. I'm trying to think of— Clear! Once again. Yeah, there you go. Charging. I've, I don't know about if you guys have watched this, but I've been watching this show called The Pit.

**Izar Tarandach (0:36)**: Yep.

**Chris Romeo (0:36)**: On HBO Max. So I'm basically a medical expert now.

**Izar Tarandach (0:40)**: I'm sorry, you didn't attend the school of ER?

**Chris Romeo (0:44)**: I was gonna say, I did, but ER was so, yeah, those are so dramatic. This is actually, feels very real. Like, and my daughter's a nurse and she said, They're consulting with doctors and nurses and everything in the production of this show to make it as realistic as possible. Like all the equipment they're using is real.

**Matt Coles (1:05)**: You know that ER basically did the same with Michael Crichton as the—

**Chris Romeo (1:09)**: Yeah, but ER was not— ER became a dramatic Grey's Anatomy. I mean, I watched ER for all the seasons.

**Matt Coles (1:16)**: Wait, ER predated Grey's Anatomy by some amount of time.

**Chris Romeo (1:19)**: So I mean, it was the same. It's what people— I mean, most people listening—

**Izar Tarandach (1:22)**: I'm there just for the machine that goes ping.

**Chris Romeo (1:25)**: Yeah, most people listening would have no idea what ER is. It's just kind of made a comeback now.

**Izar Tarandach (1:29)**: Yeah, we're old. Yeah.

**Chris Romeo (1:31)**: With the pit.

**Izar Tarandach (1:32)**: We're old.

**Chris Romeo (1:32)**: It's made a comeback. It's popping up on people's feeds now.

**Izar Tarandach (1:35)**: I remember that there was the child actor, uh, Clooney something.

**Matt Coles (1:40)**: Clooney.

**Chris Romeo (1:42)**: Yeah, yeah, it was like a child.

**Matt Coles (1:44)**: Wait, hold on. If Clooney was the child actor, then what was Noah Wiley?

**Chris Romeo (1:48)**: He was a baby on the show.

**Izar Tarandach (1:50)**: He was in the episode about the lady that was pregnant. He played the baby.

**Chris Romeo (1:57)**: Wasn't Wil Wheaton on that show too?

**Matt Coles (1:59)**: Was he?

**Chris Romeo (1:59)**: It was a joke.

**Izar Tarandach (2:01)**: Oh, who?

**Chris Romeo (2:02)**: Wil Wheaton.

**Matt Coles (2:03)**: Wil Wheaton.

**Izar Tarandach (2:04)**: Oh, Wesley Crusher.

**Chris Romeo (2:07)**: He might have been beamed in for an episode or something.

**Izar Tarandach (2:11)**: Well, we know that if you're talking about medical series, there's only actually two that are worth watching. House and Doogie Howser.

**Matt Coles (2:19)**: House.

**Izar Tarandach (2:20)**: House is one of my—

**Matt Coles (2:21)**: What was that one? Doogie Howser.

**Izar Tarandach (2:23)**: Doogie Howser.

**Chris Romeo (2:25)**: You know, House actually has a lot of applicability to cybersecurity because what is the thing that he like builds into every one of his residents? What's the lesson?

**Izar Tarandach (2:36)**: Don't trust the patient.

**Chris Romeo (2:37)**: Everybody lies.

**Izar Tarandach (2:39)**: Everybody lies. I've been referred to as the House of of computer sometime in the past.

**Chris Romeo (2:46)**: With the Vicodin or without the Vicodin?

**Izar Tarandach (2:49)**: Without the Vicodin. Monster. Monster is my drug.

**Matt Coles (2:52)**: And the limp.

**Izar Tarandach (2:54)**: But, um, probably so. Probably so, you know. Probably so.

**Chris Romeo (2:58)**: Okay, that's, that would be a good t-shirt. I am the house.comma MD of cybersecurity.

**Izar Tarandach (3:09)**: Every agent lies. Speaking of which, I'm working on the, uh, t-shirts for Vienna. So if our listener in Australia has ideas, hit me up.

**Chris Romeo (3:24)**: Now, if the listener in Australia comes to Vienna for the conference, will you give out a t-shirt?

**Izar Tarandach (3:28)**: He gets a t-shirt.

**Chris Romeo (3:30)**: He or she.

**Izar Tarandach (3:30)**: He or she.

**Chris Romeo (3:31)**: Yeah, we don't know who's—

**Matt Coles (3:32)**: it's not a security table.

**Chris Romeo (3:33)**: Who's on the other end.

**Matt Coles (3:36)**: You'll have to work on the security table somehow into the t-shirt then.

**Chris Romeo (3:40)**: Yeah, because that makes it worth a lot less.

**Matt Coles (3:47)**: Well, we know what happened with the last security table t-shirt. I threat model, therefore I am angry in Latin.

**Izar Tarandach (4:01)**: I threat model, therefore run.

**Chris Romeo (4:05)**: It's true.

**Izar Tarandach (4:06)**: All right.

**Chris Romeo (4:07)**: Well, I got to get something else off my chest here. Like this is the, we're sitting here on Monday, March 23rd. This is the Monday of RSA week. And it's the first time I haven't been there in 11 years.

**Matt Coles (4:20)**: Wow.

**Izar Tarandach (4:22)**: Have you been hit already by the tremens of RSA? Do you feel the need for us to send you swag? Yeah, FedEx.

**Chris Romeo (4:29)**: Send me some meeting requests that are meaningless.

**Izar Tarandach (4:32)**: Do you want us to send you some email asking if you want to get more details after we finish here?

**Chris Romeo (4:37)**: That would be great, actually.

**Matt Coles (4:38)**: Yeah. I have a lot of questions. How many dinners would you like to be invited to that you can't attend?

**Chris Romeo (4:42)**: That I won't attend. No, I mean, the part I'm going to miss is like, Brooke and I had a tradition of either doing lunch or dinner. Yeah. Where, uh, at Fang Restaurant and incredible food. And the funny thing is we very seldom talked about cybersecurity. In those. So it was just, it was, it's, it was a tradition that, uh, I'm gonna miss this year because I'm just not there. And I don't know if he's there either, but he, we weren't always both there.

**Izar Tarandach (5:14)**: We, we'll have a schnitzel in Vienna in honor of the tradition.

**Chris Romeo (5:18)**: You guys, it's not, it's not this, it's, it won't be the same.

**Matt Coles (5:23)**: It won't be the same.

**Izar Tarandach (5:24)**: I'm looking forward to schnitzel in Vienna.

**Chris Romeo (5:26)**: I have to be honest, don't get me wrong, schnitzel's money, but Like our, our Fang tradition was, and I've told this story before, I'm gonna tell it again cuz I want to. Um, I, when, when we first, we scheduled to go there the first time, like 7 years ago or something, I somehow had the strange impression that Brooke, he, when he told me what he used to live in San Francisco and somehow I just assumed that he used to go to this restaurant all the time. Turns out he had only been there once before. But in the, the narrative in my head is Brooke's been here, he's a regular here, the chef knows him. And so the first time we're there, we got about 6 or 7 people. Brooke's daughter's there, a couple other folks, my wife was there. And so Brooke calls over chef, asks for chef to come to the table. Now, now chef in this case is a Michelin star chef. Like he's a, he's the real deal. He looked, dresses like a normal person. He doesn't, there's no air about him. Like you, when I hear Michelin star, I think somebody's going to have like an entourage and they're going to be like, you know, the big head, super like, you can't speak to them. They're— that's chef. Super down-to-earth guy comes over and Brooke is like, looks at him and says, uh, chef, would you make us a most, uh, tasty meal? And the chef starts grabbing the menus and he, he looks and he's like, any allergies? We're like, no. And he's like, okay. And then he starts sending out dish after dish after dish, just like I refer to as a, a symphony of food. Like he's planning these dishes in order based on flavor profiles and things that I don't understand. I couldn't order them again 'cause I have no idea what they were. I mean, they told us like what kind of base they were when they, when they came out. But in the back of my mind I'm thinking, wow, Brooke is, his relationship with this guy, like having known him for years and years really opened up this door for us. And then I asked him later and he's like, I've only been here once before.

**Matt Coles (7:30)**: Well, okay, so both could be true. He could have only been there once, but he might have known him from elsewhere.

**Chris Romeo (7:36)**: No, no, he didn't.

**Izar Tarandach (7:37)**: No, no, he didn't.

**Chris Romeo (7:38)**: It was— I had a story in my mind that, that created. But yeah, so then every, every year that we were both at RSA, we would revisit Fang. One time they were filming a reality show. Cause why? I mean, why not? They were filming some type of, of food reality show. We had to sign a waiver out front to go in and sit down and eat that we could be in the back of the reality show. And all they were doing was B-roll footage. So like they had like waiters and waitresses like walking down the aisle quickly carrying empty trays that had things on top. And like we were in the background and I never did find the show. To know if we actually made it into the B-roll of this reality show. Um, so, okay. So that's, that's the story though.

**Izar Tarandach (8:32)**: It's a good story.

**Chris Romeo (8:33)**: Yeah. It's one of my favorites.

**Izar Tarandach (8:36)**: So, oh, new stories. Where are we going today?

**Chris Romeo (8:39)**: All right. Let's talk about this Medium article. Well, it's not a Medium article. It sounds like I'm a, Gen Z person if I say this is so Medium, I think I might have started something there. Um, no, they don't say that though.

**Izar Tarandach (8:52)**: I've been reading, I've been reading Medium for a long time and I still haven't contacted anybody on the other side. I still haven't heard any future predictions that are worth it.

**Chris Romeo (9:03)**: That's, that is a seriously good dad joke that you just went with there.

**Izar Tarandach (9:09)**: All right.

**Chris Romeo (9:11)**: The article says the security risk nobody talks about with AI agents. And the gist of this is that AI agents are people too, from a security perspective.

**Matt Coles (9:26)**: What's really weird is I've linked, I linked that article and I can't find it.

**Izar Tarandach (9:31)**: It's on the, uh, it's on the, uh, I can't search for it.

**Matt Coles (9:35)**: You can't search for it. It doesn't show up. Oh, I know why, because my browser switched to Yahoo.

**Chris Romeo (9:42)**: What the hell is Yahoo now? Wait, wait, hold on. Stop.

**Izar Tarandach (9:46)**: I knew that was going to trigger someone. Yahoo what?

**Chris Romeo (9:50)**: What did you tell me more? I need more context. The browser switched to Yahoo.

**Izar Tarandach (9:54)**: What? Yahoo Mail?

**Chris Romeo (9:55)**: Yahoo Search?

**Matt Coles (9:56)**: So I, I did something stupid. I let McAfee, my AV engine, filter the web for me briefly because I'm never making that mistake again.

**Izar Tarandach (10:07)**: On purpose?

**Matt Coles (10:09)**: And its default engine, no matter how much you want it to be something else, is Yahoo.com.

**Izar Tarandach (10:19)**: Which interestingly, interestingly, it brings us back to Brook, because Brook used to be the security architect at McAfee. So it could be that he's the one who said you should use Yahoo.

**Matt Coles (10:32)**: It's what all the kids are using these days.

**Izar Tarandach (10:34)**: Yeah.

**Chris Romeo (10:35)**: What would you get from Yahoo at this point?

**Matt Coles (10:38)**: I get a quick link to Google. It's like Bing, right? The first thing you do when you install Windows is you use Bing to go to Google and get Chrome.

**Chris Romeo (10:50)**: Yeah.

**Izar Tarandach (10:50)**: Okay.

**Chris Romeo (10:52)**: That's, I'm sorry. That was just, it caught me off guard when somebody said I use Yahoo. I'm like, yeah, I used to use Yahoo. It was in 1999 maybe.

**Izar Tarandach (11:03)**: For a second there, Chris was thinking it's time to buy Bitcoin.

**Chris Romeo (11:06)**: Oh dude, I'm all in on Bitcoin. Bitcoin. This, this is the Bitcoin arbitrage show. Get ready for a new episode.

**Izar Tarandach (11:16)**: No, seriously, if, if it was the time that everybody was using Yahoo, then it was the right time to buy it.

**Chris Romeo (11:22)**: Yeah. If we listen, if we would've all bought Bitcoin in 1999, we would not have our yachts. We'd be recording this from one of our yachts as we were parked next to each other in the Bahamas or something.

**Matt Coles (11:34)**: Actually.

**Izar Tarandach (11:34)**: Or, yeah.

**Matt Coles (11:35)**: Well, or you would've spent 10,000 Bitcoin on a pizza.

**Chris Romeo (11:38)**: True. That did happen.

**Matt Coles (11:40)**: Did happen.

**Chris Romeo (11:41)**: That was a very expensive pizza that I, that I bought.

**Matt Coles (11:44)**: But yes, that was, you learned, uh, 600,000 right now. No, 700,000.

**Chris Romeo (11:50)**: 70,000 a coin right now.

**Matt Coles (11:53)**: Times 10,000.

**Chris Romeo (11:54)**: $700,000 pizza. I think better been good. That's all I'm saying.

**Matt Coles (12:01)**: Pineapple and mushroom.

**Chris Romeo (12:03)**: So, so buy pizza. Yeah.

**Matt Coles (12:04)**: Fruit and Funky.

**Chris Romeo (12:05)**: This article then, do we agree that AI agents are be, are, I guess the way they describe it, become an actor inside your infrastructure. Let's start there.

**Matt Coles (12:19)**: So I do wanna highlight that this is the second time this article has been written in the past month on Medium. So I just found while I was searching on Yahoo for a non-existent article, I found—

**Izar Tarandach (12:39)**: for something that nobody talks about, that's pretty frequent.

**Chris Romeo (12:43)**: Yeah.

**Matt Coles (12:43)**: So the article we want to talk about is the security risk nobody talks about with AI agents. And there's the other article, which is AI agents are the new security risk nobody talks about.

**Chris Romeo (12:55)**: Well, I guess both of the titles are false. Yes.

**Matt Coles (12:58)**: But one of them is paid. One of them is member walls as opposed to The other one is open.

**Izar Tarandach (13:02)**: Yeah.

**Chris Romeo (13:03)**: Well, we don't, I don't believe in paywalls.

**Izar Tarandach (13:05)**: That's the one that used the more verbose agent to write. This one is the cheap one.

**Matt Coles (13:12)**: Right?

**Chris Romeo (13:13)**: But I mean, do you all agree?

**Matt Coles (13:16)**: To go with the one which is the privacy risk nobody talks about when using Agentic AI.

**Chris Romeo (13:21)**: I'm going to write an article about this, about summarizing all the things people don't apparently seem to know. But do we agree with this? Like, is an AI agent part of the infrastructure? Well, part, or the actor, are they an actor? Yeah.

**Izar Tarandach (13:38)**: It's not a matter if we agree or not. It's, it's a matter of it's happening. People, people are putting those things everywhere. Like it's the new duck tape.

**Matt Coles (13:47)**: Let's just assume for a moment that an AI agent is a system process, which it is. We give we give, uh, identities to system processes, right? Mm-hmm. Service accounts exist. Why would we treat them? Why should we treat them differently?

**Izar Tarandach (14:10)**: We shouldn't treat them different from a point of view of identity. Each one of them has to have an identity.

**Chris Romeo (14:17)**: Yeah.

**Izar Tarandach (14:17)**: We have to treat them differently in terms of lifecycle. Because their capabilities may not change per versioning, they may change per environment.

**Matt Coles (14:31)**: Well, here's an interesting question. Are AI agents, as you spin up and down agents, right? Agents get, agents, you have a main loop, agents start, and you form a pool of agents when you're doing, if you're going to do agent work. Are they independent processes? Do they fork the process?

**Izar Tarandach (14:57)**: If you're doing it right, they have to, because otherwise one is going to be, uh, uh, stopping the other. So you have to have them as independent processes. Does that lead to different identities? That's another question to be asked.

**Matt Coles (15:15)**: Well, because they may have different roles, right? They may, exactly. May need to do different things.

**Izar Tarandach (15:20)**: Yeah.

**Matt Coles (15:20)**: And, and they may have different lifetimes.

**Chris Romeo (15:24)**: Are AI agents that complicated now?

**Izar Tarandach (15:27)**: And wait, wait, wait, they, they, they may not be forks. You could just be contacting another agent of, of the network and it's a completely different ecosystem, completely different thing. No more than, no more than a client and server. So to, to Chris, uh, uh, question.

**Chris Romeo (15:45)**: We may be overcomplicating something that we already know very well, which is any fork, any client-server architecture has the same privileges and access as any other one that's in the same kind of category or same identity as any that it was given. Yeah. I don't think there's, I wouldn't expect to see very complicated role, like a, an agent being broken down into multiple roles and giving up privileges at this stage. In different segments of itself? Well, if you have like a least privileged—

**Matt Coles (16:21)**: if you have one, if you have one agent that's doing research and one agent that's actually writing an output, they have very different expectations of what they can do.

**Chris Romeo (16:32)**: Yeah, but that's like saying one application does research and one application does something else, right? My point is they don't— like, that single agent, I think, has the same access credentials at this stage for whatever that agent is doing.

**Izar Tarandach (16:47)**: There is a, there's a problem that we keep falling into when talking about AI, which is the overuse of certain words. Context, for sure. And now, role. Because as Matt said, you can have a pool of agents. Do they have a name for a gaggle of agents? A crowd of agents? A swarm? An agency of agents?

**Chris Romeo (17:11)**: Swarm sounds good.

**Izar Tarandach (17:12)**: Something. So you can have one of those and then every agent has a role inside that level of ecosystem, right? Let's call it the business level of what they're doing. So one is a writer, one is a reader, one is a critic, the other one just eats popcorn.

**Matt Coles (17:30)**: But, uh, but how many TCPs do they run under?

**Izar Tarandach (17:34)**: How many?

**Matt Coles (17:34)**: How many TCPs do they run under?

**Chris Romeo (17:38)**: Matt throwing out the old school trusted computing base for those people that don't know what TCB stands for. It's not taking care of business like Elvis said.

**Izar Tarandach (17:46)**: Then you have the role inside of the TCB and then you have the role inside whatever cloud, whatever they run, or the role that gives them, that they assume in order to get permissions to do the stuff that they need to do.

**Matt Coles (17:57)**: But they all inherit the same permissions then.

**Izar Tarandach (18:00)**: Not necessarily. If you fork them, they're running under one context or multiple contexts, right? Yeah. But if they are completely separated processes contacting just over whatever protocol you put in place, then not really. They may be part of the same swarm and have different roles inside that swarm. And at the same time have different roles inside their, whatever orchestration they work in. Right. But, uh, at the end of the day, I think that from an identity point of view, it's not very different from thinking of them as separate users or separate somethings. But my point is that it's not an abstraction that we don't know. There's no point in romanticizing this thing and say, oh, it's a completely different thing. We need new controls. We need new frameworks.

**Chris Romeo (18:52)**: We need new.

**Izar Tarandach (18:53)**: Threat modeling frameworks. No, you, you just use whatever you already have. I mean, as I like to say, 10-year-olds running with scissors, but you're going to relate to them the same way that you relate to the things that you have today. They're not magical. Hmm.

**Chris Romeo (19:15)**: So the article goes in kind of the next phase to say that folks are trying to secure AI with things like prompt restrictions, output filters, behavior monitoring, but the real challenge when it comes to agentic AI is access because AI agents have credentials, they have tokens, they have API keys, and they can interact with real systems. So it's, it's, it's, it's almost like what we're describing here is a kind of a two-phase approach to securing the world of AI.

**Matt Coles (19:48)**: It always has been.

**Chris Romeo (19:50)**: Yeah. But you have to secure, you have to secure the interaction with the AI and then you have to secure the access that the AI has to the other systems around it.

**Matt Coles (19:59)**: Right. Now replace AI with application. Hmm.

**Chris Romeo (20:04)**: I can't remember that far back to repeat what I just said.

**Matt Coles (20:08)**: Or, or replace agent with just AI. Uh, I mean, it's, this is not a new thing, right? It's amazing that this article exists, that it has to exist actually, that people have to be reminded.

**Chris Romeo (20:22)**: AI has been moving so quickly that we like, think about all of the past technological inflections that we've been a part of in our careers. There's been time to stop and think. And I mean, even like think about cloud, right? How much time did we have to figure out how to secure cloud? It took us years.

**Matt Coles (20:47)**: How much time did we have or how much time did we use?

**Chris Romeo (20:49)**: How much time did we use? Yeah, we didn't have years, but we, we, we used years to figure out there's a tail on Matt's video screen.

**Matt Coles (20:59)**: Nessie.

**Chris Romeo (20:59)**: Just, just putting out a show. Uh, but yeah, so that means like with AI though, like I don't hate the fact that somebody wrote this article. Because it is stating what we can say, oh, well, duh, this is obvious. Everybody should know this. But I don't think that's the case. Like, I don't, I don't think everybody knows this at this point.

**Matt Coles (21:26)**: But, but that, why?

**Chris Romeo (21:28)**: But it, it's not because things are moving so fast right now. People aren't even, people are, are, yes, everybody's gonna say security's important, but people are focusing on velocity and AI is changing week to week. We know this.

**Izar Tarandach (21:44)**: Look, so about a year ago, we had that discussion about prompt injection. Yes or no? Is it a security problem? And I think that where we landed was that prompt injection was not a problem. Agency was over agency was. And this article actually just states basically the same. The problem with agents is what you're letting them get away with doing or not doing.

**Matt Coles (22:05)**: Right.

**Izar Tarandach (22:06)**: Right.

**Chris Romeo (22:07)**: So So agents are basically children.

**Matt Coles (22:10)**: Yeah.

**Izar Tarandach (22:11)**: That's just like a kid with scissors apparently.

**Chris Romeo (22:13)**: With scissors. Yeah. Okay. But, but my point is—

**Matt Coles (22:16)**: Because they can delete your file system.

**Chris Romeo (22:19)**: Right.

**Izar Tarandach (22:19)**: My, my, my point is that what's getting out of control right now for me is that, okay, we, we gave them agency, right? Great. But now that we have, back when we spoke about this, we were talking about, okay, you go into your whatever client, talk to your LLM, the LLM knows how to do some stuff, so that stuff gets done. But now that they have like agents running in loops and, uh, tools that are at the disposal of the agent and the agent gets to decide how to use the tool and what to do with the tool, then my problem with securing AI becomes a bit different. We are, we have now to look at, at two levels of, um, let's call it for the lack of a better word, logic. You have whatever's happening in the code that you are using to write your agent, which, be very surprised, people, still exists. You have to write code. Actual agents are written in Python and other stuff. But then you have to look at the logic that's present in the prompts that are inside that implementation of an agent. So if, if in the past we would look at, uh, another name for what this article is implying, feature creep, and say, okay, you have code and now somebody added another function that does XYZ that we don't like, but we can track model this thing. We can see that it's going in that direction. We can say, okay, let's do it differently. But now, not in the code, somebody can change a prompt and all of a sudden a prompt opens this whole window of opportunity for that. Thing to use tools that are at its disposal to go and do other things, which are non-deterministic, unpredictable, happen, don't happen in perhaps the order that we expect or otherwise. That's what gives me pause because now I'm talking to developers at two different levels. I need to know when you are adding stuff to your agent in terms of what it can do. An idea to know when you add stuff to your agent in terms of things that it thinks that it might want to do.

**Matt Coles (24:37)**: Or have been given access to, or liberty to do.

**Izar Tarandach (24:41)**: So now the least privilege plays a different role in terms of, it's not only a principle in terms of, yeah, sure, you should do this in order not to be exploited. But it's a moving target because the things that somebody may want their, their agent to be able to do will require more privilege over time. And that privilege is going to be used and exploited in different ways that may not be completely deterministic or clear to us. And that you may not be able to scan for.

**Chris Romeo (25:17)**: Scan for an email, scan for an open— Doesn't that imply that the agent's going to go rogue? That the AI is gonna go rogue if it uses its permissions to do something it's that you weren't intending it to do.

**Izar Tarandach (25:32)**: My problem with rogue is that to me, and I could be wrong, not, not an English speaker, to me it implies intent.

**Chris Romeo (25:42)**: So you think it's gonna be an accident, AI's gonna accidentally do naughty things if we're using the children?

**Izar Tarandach (25:48)**: I think that it may. And I think that the main problem here is that we don't know when we cross that line from may not to may.

**Chris Romeo (26:02)**: Doesn't that require artificial general intelligence though?

**Matt Coles (26:06)**: I think that it only requires a bad written prompt or a, or a trick, right?

**Chris Romeo (26:12)**: I mean, so, so there's still, so in the examples you're using, there's still a human behind the nefarious. Activity.

**Matt Coles (26:19)**: Well, or another agent who prompts Agent A prompts Agent B to do something. That doesn't mean there's AGI.

**Chris Romeo (26:25)**: That just means at the request of another human in this example that we're using, or in, or other input.

**Izar Tarandach (26:32)**: The, the, the human may start with intent in one direction, but for the collection of tools that are available to the agent, for the way that the prompts have been written, for the input that it has received from yet another non-deterministic system. The stars may align in such a way that it goes and erases your database.

**Matt Coles (26:52)**: Well, so here's an interesting question. Can agents, we already know agents can write files on the file system.

**Chris Romeo (26:58)**: Sure.

**Matt Coles (26:59)**: And I'm assuming, so this is maybe a limit of my technical knowledge. There's the tool-based approach, but it's also a Python process usually. So Python has obviously read-write capabilities. And in fact, the process that you're running when you run it as a user, whether that's a human or a system user, almost certainly has been granted read/write privileges. At the very least, read, and almost certainly write just by omission of the removal of that. So can AI agents rewrite their own prompts, their system prompts or what, whatever? In other words, can they grant themselves additional capabilities?

**Izar Tarandach (27:45)**: So I've seen, for example, Claude skills that try to optimize Claude skills. Does that count as reading, uh, as writing your own prompts? I don't think so, because then you have to get out of it and get it to read the skills again and this and that and the other one. The point is, And I'm just like thinking aloud about what you just said. Don't take this as authoritative in any way. There's nothing stopping an agent from writing text that will eventually be interpreted by the same agent or a different one as a prompt, which is the whole point of automated prompt injection and whatnot. Right? So if you take this a step back, when you have two agents talking, basically what you have is them writing prompts to each other. That's the whole point of their interaction.

**Chris Romeo (28:51)**: Right?

**Izar Tarandach (28:52)**: The question here, and again, just thinking aloud, we are tending to And here comes the long word, anthropomorphizing. Thank you. The things and say they have intent, they have goals, they have this, they have that, they think about how they go. They don't. They don't. Let's make that very clear. They may have goals that are stated in a prompt, but they're not thinking how they get there.

**Chris Romeo (29:24)**: That's what I was trying to tease out.

**Izar Tarandach (29:27)**: They're throwing the coin and seeing what's the next best word to get close to that point in the latent space. Right.

**Matt Coles (29:33)**: But they could come up with a, they could come up with a reasonable response to a, to a question or a idea that a human has because they initiate the process, but through the various processing steps do something different. And that something different may be, uh, it may think in its process, I don't have enough information, I need more, and expand its capability.

**Izar Tarandach (30:00)**: So if you go into Claude, and I'm just throwing Claude because it's the one that I've been using lately, the other ones must be the same way, and you put it in plan mode and you say, let's build something that does ABC, and I think it's, uh, Shift+O or Ctrl+O or something like that, and you, and you see its internal dialogue. The impression that you have is that you are looking inside the head of somebody who's thinking very hard about the thing that you just asked. The things that it goes back and forth. Oh, I'm going to do this. No, no, no, no, actually I'm going to do that. No, no, no, no, no, the other one. Right? So it's, it's, it's very interesting from the point of view of it looks like an internal dialogue and it's valid. It looks very authoritative. It could be completely wrong. While doing that, it's going to, you are going to see it saying, oh, I know how to do this. I have a tool to do that. And it's evaluating, should I use this? Should I use the other one?

**Matt Coles (31:03)**: By the way, it's using, it's knowing that because it has been given a description of that tool.

**Izar Tarandach (31:07)**: Yes.

**Matt Coles (31:09)**: Which could itself be wrong.

**Chris Romeo (31:11)**: Training data.

**Izar Tarandach (31:12)**: Yes. Well, not training data. It's part of the context. Yeah. That the initial context that goes up actually in Claude, I just learned this this weekend. If you type /context, it will break up the parts of the context that it has right now in telling you how many tokens are taken by tools, how many tokens are taken by your question. It's really interesting to look at that. There's a lot of stuff to learn there, but the point is that at the end of the day, it's thinking context to context. They don't have memory in between that, at least the way that I understand it. So it could be that in the next round, it's going to think something different about the same tool because information changed, because it went to a different space, a different point in the latent space. And we go back to the agency thing. Okay. So you, you let it do stuff. And right now my, my main point of, uh, laughing is, yeah, sure. Nobody's going to say, go without any restrictions, do your thing and come back with results, or nobody should. But on the other hand, the option is to be pressing enter every 30 seconds and nobody's reading all that stuff that they're pressing enter to.

**Matt Coles (32:38)**: It's like the ULA.

**Izar Tarandach (32:40)**: But they're, no, they're, yeah, exactly. They are feeling good because all of a sudden they are the human in the loop.

**Chris Romeo (32:46)**: I'm in control. Yeah, it's like the ULA.

**Izar Tarandach (32:50)**: If I have to fall asleep, me too. But I don't see a more important contrast.

**Matt Coles (32:56)**: So let's get back to, let's get back to this. If you have that, if that's what happens in an AI agent and all those agents have the ability to be independently doing that, and they have all have independent needs and expectations and obligations and constraints that we want to put on them, they should all be individual identities.

**Izar Tarandach (33:21)**: Yes and no. Right? If you have a if, if you put a bunch of people inside a room and give them bits and pieces and tell them, build something, you might have one of them that takes leadership and says, this is the plan. You might have one of them that, uh, is the glue person, one of them that is the cutting person, one of them that it's the other. Inside that room, all of them have all the, may have all the same permissions. You could interchange. Between them without changing the, the, the final product.

**Matt Coles (34:03)**: They may not all have the same capabilities though. Yeah. And that's what makes them, makes them human.

**Izar Tarandach (34:09)**: Right. But, but in theory, they, they, they might all be interchangeable people. They, they just happen to be a group of 10 people, all of them with the same capabilities, all of them with the same tools.

**Matt Coles (34:21)**: Yeah.

**Izar Tarandach (34:21)**: And they just throw a coin and say, I'm the one who's designing, I'm the one who's gluing, I'm the one who's cutting.

**Matt Coles (34:25)**: We need to, we need some Pink Floyd in I'm not a big Pink Floyd fan, but we need some Pink Floyd here.

**Izar Tarandach (34:30)**: Just Another Brick in the Wall.

**Chris Romeo (34:35)**: We don't have the budget to play that in the background, so audience listeners, could you imagine that you're hearing that right now or play it from your own streaming service right now?

**Izar Tarandach (34:45)**: Right, right. But the point is that it's hard for me to put a model on top of the whole thing and say this is how it should be without actually running a threat model. Because there will be cases of yes, everybody's in the room. There will be cases of no, everybody gets a different room. And so they need different levels of access or different levels of agency. But of all the things that are in this risk that gets never talked about, which by the way, gets talked about all the time. My, my, my, the thing that I'm trying to keep in my mind right now, how to address, how to search for is, is these two levels of, uh, addressing the, the feature creep and the, what are my developers going to do next of the code and the prompt. And that's something that I haven't gotten any good answers to. Please nobody send me an email saying, can you come and see our system? We want your opinion for 15 minutes only after this, because no.

**Chris Romeo (35:56)**: Well, I mean, you're available.

**Matt Coles (35:58)**: It just costs $5,000 or 1 Bitcoin for that 15-minute meeting.

**Izar Tarandach (36:03)**: Just because I haven't gone to RSA doesn't mean that I won't be getting email.

**Chris Romeo (36:07)**: That's true. That's true. All right. Let me, let me, let me kind of wrap this article up here because there's the last point in the article I really want to discuss real quickly before we wrap up. But if we go through the conclusions, treat AI agents like real users. So AI agents should have to change their password at the same frequency that all the rest of us do, which would be hilarious.

**Matt Coles (36:31)**: Fair luck.

**Chris Romeo (36:32)**: What? 30-day password change. Like, what are you— the AI is like opening a support ticket. This is stupid that this is the policy. Um, control access, not just behavior. I think we're going to agree with that. The rise of shadow AI.

**Matt Coles (36:46)**: We don't have time to talk about shadow AI today, but I think that's a red herring in this article because it really has nothing to do with the rest of it completely.

**Chris Romeo (36:53)**: Security must consider intent. I mean, we never do that. So that's, that's just bad advice. What you intend for AI agents need lifecycle governance. Like, I'm seeing all of this, this fancy governance word. People are— everybody's throwing this around. It's, it's, um, It's compliance over, it's compliance over risk. But it used to be contained right within the GRC framework that we were doing and, and whatnot.

**Matt Coles (37:23)**: The, the GRC.

**Chris Romeo (37:24)**: Now it almost seems like it's made a comeback though, because, and it's the hot, it's a, it's one of the new hot things out there. Like, what does your AI lifecycle governance look like? Like, is it something that we actually need to be spending this much time on? Is it that different? It's something.

**Izar Tarandach (37:41)**: It is something that connects to something that we had some episodes ago when we said that everything was boring.

**Chris Romeo (37:47)**: Yeah, I remember that.

**Izar Tarandach (37:48)**: So I think what I'm seeing actually is that for a long time, developers were bored. I'm writing one more CRUD system. I'm writing one more dashboard. I'm writing one more of this. And all of a sudden, this shiny pebble fell in front of them. And it's thrilling. Things move even faster and people are afraid this thing can do my job. Oh no, it can't. And I think that the governance resurgence that you've been seeing is just security people trying to take control. Sorry, not to take control. Trying to—

**Matt Coles (38:32)**: Did you say To take governance?

**Izar Tarandach (38:34)**: No, no, no, no, no. To take, uh, uh, um, look, things are happening way, way, way too fast for teams that were already overloaded. So I think that this AI governance resurgence, not this resurgence of governance over AI is, is happening mostly because security teams are like Dude, this is moving too fast. We have absolutely no control. We don't know what's happening here. It's getting scary. Everybody slow down. Let's make lists. Everybody now, let's make lists.

**Chris Romeo (39:13)**: That's what we do in times of trouble. We make lists.

**Izar Tarandach (39:16)**: We make lists.

**Matt Coles (39:17)**: Now, interesting question is security teams and GRC teams are usually separate, right? Because audit is in GRC and security is not audit.

**Chris Romeo (39:27)**: Often for good reason, they're kept in separate spots.

**Matt Coles (39:32)**: So when you say security teams are doing lists, are they asking the GRC team to make lists for them?

**Izar Tarandach (39:42)**: I, I think that I probably have an outlier, uh, experience. I have had really good experience working with GRC teams. And in fact, I hope that they're listening to this now. I know that they're not, but I actually love the GRC team that I have right now. They're technical. They, they, they know how to look at things. They know how to listen and they know how to use the tools that we have at security to complement their work. And they are looking at using this whole AI thing to make their work, which here between us, it's very hard, which is another word for boring, uh, to, to make it, uh, easier. Right. So I'm thinking that perhaps AI is the time when those two teams finally come together and have a better dialogue. We could have reached the point in history where that division stops making so much sense.

**Chris Romeo (40:41)**: Hmm. Until AI takes over GRC.

**Matt Coles (40:46)**: Yeah.

**Chris Romeo (40:47)**: Yes. You get the, you get the GRC chatbot and it's, it just annoys you to death.

**Matt Coles (40:52)**: It's just the lists. It's just the bulleted lists with a lot of mm dashes and LLM specific words. Yes.

**Izar Tarandach (41:03)**: I, I don't know why, but you, you, you said GRC chatbot and, and the thing that popped in my mind was something that I saw like 2 weeks ago. Somebody saying, uh, how, how you don't have to pay any AI provider. And then there was a screenshot of them in some support chatbot of something, Amazon or whatever. And the person was saying, hey, you know what? I want to buy your product, but before that, I really, really, really need a Python function that calculates the Fibonacci, uh, the Fibonacci, uh, series. And the thing said, oh, fine, yeah, no, no, I'm ready to hear some recursive version of the Fibonacci series. And I don't know why the GRC chatbot took me to that, but yeah.

**Chris Romeo (41:52)**: It's a good example.

**Matt Coles (41:54)**: Now I'm curious, if there are GRC chatbots out there, can you get them to do other things? Of course, you know, that's an attack where you're consuming tokens, right? So you can deny a service or really run up the cost of a company by doing that.

**Izar Tarandach (42:11)**: Yeah. Which by the way is another thing with agents. Like sometimes they start their dialogues and the dialogue just goes crazy. I had this little one that I wrote where you could like get two different providers, give them a goal and let them discuss. And those things go crazy and you can just see like the cost ramping up on both sides. What? So it's, no, it's another way of a rogue rogue, that, that, that word again, agents, like driving up your, your bill up.

**Matt Coles (42:40)**: And what's like, it's like giving your toddler your iPad.

**Izar Tarandach (42:43)**: No. So look, think about this.

**Chris Romeo (42:44)**: You should never do that.

**Matt Coles (42:45)**: Yeah.

**Izar Tarandach (42:45)**: Right. The, the, we, we used to have like deadlocks and, and, and things like that to worry about who's writing, who's reading. But now, no, now, now we have agents running up token costs by just talking to each other.

**Chris Romeo (43:00)**: It's a feature.

**Matt Coles (43:01)**: So wait a minute.

**Izar Tarandach (43:01)**: It's a feature.

**Matt Coles (43:02)**: So we do need access controls. But what we really need are parental controls.

**Izar Tarandach (43:07)**: Yes. Yes.

**Chris Romeo (43:07)**: Parental controls for AI agents. All right. Yes. Patent pending trademark. Uh, all rights reserved. I will be writing my next book called Parental Controls for AI Agents. No, but it's even worse.

**Izar Tarandach (43:20)**: We have them today and they are called pressing enter. Right? The human in the loop. No, you have the whole loop discussion, human in the loop, human on the loop, which by the way, we should talk about at some point. Because I'm, I'm very afraid of the resurgence of the human in the loop extended range. Think about that one. You're gonna get there eventually.

**Matt Coles (43:46)**: But, but, but parental control, but parental controls, if they're the kid running with scissors, parental controls are the thing you need.

**Izar Tarandach (43:54)**: Exactly. Until you get—

**Chris Romeo (43:55)**: so we prevent the AI agent from looking at YouTube. Got it. Got it. Solution.

**Izar Tarandach (44:03)**: I think, I think that both. Oh, and using your credit card.

**Chris Romeo (44:10)**: That's a good idea.

**Izar Tarandach (44:15)**: Parental controls for AI. There might be something there.

**Matt Coles (44:22)**: We heard it right here first on the original security table.

**Chris Romeo (44:26)**: The original security table. Not to be.

**Izar Tarandach (44:29)**: The one and only not to be confused with some copycats out there.

**Chris Romeo (44:35)**: And no offense to cats, Matt. Sorry.

**Izar Tarandach (44:37)**: Totally. All right.

**Chris Romeo (44:40)**: Well, I think that was our final thought. I think that was our final thought. I think parental controls for AI is something that someone's probably going to use AI to snarf out of this episode and then turn into something before we ever have a chance to monetize it. That's why I said that whole— Patent pending trademark, just in case that would work.

**Matt Coles (45:01)**: Yeah. Prior art here.

**Chris Romeo (45:02)**: To slow the, the AI down. But yeah, that's, that's a brilliant idea. I haven't heard anybody say that yet. Parental controls for AI. So I think Matt was the originator of that.

**Izar Tarandach (45:11)**: Yeah. I'm gonna have to expand on that one.

**Chris Romeo (45:14)**: But I've said it 3 times now so I can steal it and use it as my own.

**Matt Coles (45:18)**: So it's in the transcript.

**Izar Tarandach (45:20)**: Strategy, but, but with attribution. Yeah.

**Chris Romeo (45:24)**: Only the second, only two times you have to give retribution. Retribution. Third time it's your own.

**Matt Coles (45:30)**: Attribution, not retribution.

**Chris Romeo (45:31)**: Retribution. Retribution. You got me there.

**Izar Tarandach (45:34)**: I just contribution.

**Chris Romeo (45:35)**: Contribution. Trying to think of some other butions. Um, all right. Matt's the king of the butions apparently. He can, he can really get those words. I can't think of another one to top. Bution.

**Matt Coles (45:50)**: You could use AI to figure this.

**Chris Romeo (45:52)**: I could, but that would be, that's not, that takes the fun. That's the problem with this world, Matt, is that everybody uses AI for things that they used to use their brain for. And so here's your homework, guys.

**Izar Tarandach (46:02)**: AI attribution.

**Chris Romeo (46:03)**: AI attribution. Okay, here's your homework. All three of us have watched this. If you haven't watched it, go watch Idiocracy and then come back and realize this is the future of using AI. We will eventually reach this this level of—

**Izar Tarandach (46:17)**: You see, that's, that's another form of feature creep. Idiocracy started as a comedy, ended up as a documentary.

**Chris Romeo (46:23)**: Yeah.

**Izar Tarandach (46:24)**: Without changing anything.

**Matt Coles (46:26)**: So, WALL-E, WALL-E 2, right? Not, not the second edition of, but WALL-E. WALL-E also.

**Chris Romeo (46:34)**: WALL-E also.

**Izar Tarandach (46:35)**: It's got electrolytes. Yeah.

**Chris Romeo (46:37)**: I mean, you can put Gatorade on the crops and they should grow. But yeah, Idiocracy is, is, but that's what AI, that's where AI's taking us though.

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

**Chris Romeo (46:46)**: It is.

**Izar Tarandach (46:46)**: It is.

**Chris Romeo (46:47)**: Like people don't think anymore about simple problems. They just put it in AI and get the answer to it.

**Matt Coles (46:54)**: Yep.

**Chris Romeo (46:55)**: And that's a, we could, we could talk about that for a whole episode cuz that's a dangerous skill degradation in the whole works.

**Izar Tarandach (47:00)**: Because you can start with a little problem, put it on AI and end up with a big problem.

**Matt Coles (47:05)**: Yep.

**Chris Romeo (47:06)**: Yeah. I mean, and don't get me wrong, I've used AI to design building structures. And solve engineering, like, uh, real-world engineering challenges that I had with something that I built in construction.

**Izar Tarandach (47:19)**: Can I get some addresses? Not to go to.

**Chris Romeo (47:22)**: So you said I want to visit.

**Matt Coles (47:25)**: Did it pass code?

**Chris Romeo (47:26)**: I mean, well, no comment. I'm gonna plead the Fifth, Matt, which in the United States of America is the right to not incriminate myself.

**Izar Tarandach (47:36)**: Yes.

**Matt Coles (47:36)**: Yes, that's true.

**Chris Romeo (47:39)**: All right, well, that's it for the Security Table, the original Security Table, almost. We could say the OG Security Table for those people from the '90s who understand what that means. So go ahead, folks, watch Idiocracy, watch ER, have a great—

**Matt Coles (47:54)**: listen to Pink Floyd, The Wall.

**Chris Romeo (47:56)**: Yeah, no, play Pink Floyd, The Wall at various stages of this episode because we don't have the budget to pay for it. Enjoy.

**Izar Tarandach (48:02)**: And keep your agents away from Caesar.

**Chris Romeo (48:04)**: Keep your agents away from scissors and, and add the parental controls to your AI.

**Matt Coles (48:11)**: Yeah.
