The Tool Creep Problem: When More Security Means Less Security
With Chris Romeo, Izar Tarandach, and Matt Coles
In this episode, we break down why security budgets keep growing while organizations keep falling further behind. We explore how tool creep has quietly shifted from a nuisance into an active attack surface, and why agentic AI is becoming the insider threat no one planned for.
Part of: AI agent security: agents, MCP, and prompt injection
About this episode
In this episode, we break down why security budgets keep growing while organizations keep falling further behind. We explore how tool creep has quietly shifted from a nuisance into an active attack surface, and why agentic AI is becoming the insider threat no one planned for. Izar shares a firsthand account of watching an AI agent attempt increasingly creative workarounds to escape a sandbox, revealing just how much risk lives in the gap between what agents are told to do and what they are actually capable of. At the end of the day, it comes back to fundamentals: define your agents' boundaries, limit their capabilities to only what they need, and stop confusing tool accumulation with security maturity.
🚀 Join the Conversation
If your AI agent were compromised today, would you even know it was the agent and not you?
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
0:10Izar Tarandach I'm gonna take care of it.
0:11Chris Romeo We do need a soundtrack for the security table. Speaking of the security table, welcome folks to what appears to be a conversation already in progress. Chris Romeo here, joined by Matt Coles. He's our terandosh. We were discussing the soundtrack to The Life of the Security Professional. We had Paradise City, Guns N' Roses. Welcome to the Jungle, Guns N' Roses. Seems like we have a bit of a pattern here.
0:42Matt Coles And of course, since today is today, it's the Imperial March.
0:47Chris Romeo True. Oh yes, it is May the 4th here in the Eastern Hemisphere. Is that where we live? Or Western? I don't know. Somewhere.
0:55Matt Coles I don't know.
0:56Chris Romeo Somewhere. So, okay, we can put the Imperial March in there on this. But if we're also— if we're getting witty about it, I would say the another song would be specifically to the security professional would be U2, I Still Haven't Found What I'm Looking For.
1:14Izar Tarandach It's right by the car keys.
1:16Matt Coles As long as it's not— as long as it's not Sunday Bloody Sunday.
1:20Chris Romeo That could also be an indication for the security professional. —of some of our worst days.
1:28Izar Tarandach I myself am particular to painted black by the Stones. Stones. Okay. I see a red door and I want it painted black.
1:40Chris Romeo That's very, very much a security requirement. Right? No colors anymore.
1:47Izar Tarandach I want them to turn black. Yeah.
1:51Chris Romeo That was what it was like in the old days. All right, what other— Matt, give me one more song that applies to the— a double entendre song. Everdream. Everdream by who? Nightwish. Isn't Nightwish a cologne?
2:08Matt Coles I don't know.
2:10Chris Romeo Uh-uh.
2:11Izar Tarandach We, we, we are missing the, the one, the only. Where is it? Well, since my baby left me. I found a place to dwell. It's down at the end of Loneliness Strip and Heartbreak. I bet you're Heartbreak Hotel.
2:27Chris Romeo That's the life of a security professional.
2:30Matt Coles Heartbreak Hotel. That's hilarious.
2:32Izar Tarandach I feel so lonely I could die.
2:35Matt Coles Maybe also we could do a whole episode. We could do a whole episode about applying songs to cybersecurity.
2:41Chris Romeo Yeah. Folsom Prison Blues. Oh yeah. If you got in real trouble. All right. Well, we should talk about something related to cybersecurity, not just how much we enjoy the music of cybersecurity, which is realizing kind of a weird thing to say.
2:56Izar Tarandach Oh wait, there's enjoyment there too.
2:58Chris Romeo So we love reports and we're big fans. We don't like writing reports or doing book reports, but we like analyzing other people's reports. And so we had one come across the desk, making us sound like professional journalists. 2026 State of Cybersecurity Report, Bridging the Divide. The preparedness gap between threats and defenses is widening. What does the research say about how to reverse the trend? This is from Avante, which I don't know, remember what Avante does, but they do something.
3:31Matt Coles They do something. They do something. They sound familiar. They do.
3:33Chris Romeo Other than writing reports. So they have lots of cool web graphics and stuff and junk.
3:41Izar Tarandach It's a very pretty report, I have to say.
3:45Matt Coles Yeah.
3:45Chris Romeo So I had one of my AI friends give me a couple of spicy takes from this report just for us to noodle upon. And so the first one I want to get your thoughts on, they've got this chart in here, Threat Preparedness Analysis, where they're showing the gap between the difference. The gap represents the difference between threat level and preparedness, and they're showing 2026 versus previous years. So 83% of organizations say their security budgets are increasing. Board engagement is up, leadership support is up, yet the preparedness gap widened by 10 points anyway. So the spicy take is, where is all the money going? Are we becoming experts at performing security investment while operationalizing none of it?
4:40Matt Coles Is this the security theater all over again?
4:43Izar Tarandach Oh yeah, I'm going to have to bring it back to, uh, my Caesar's approach. The security brutalism— we got all the bling that we wanted across the years. We've got all the tools that we want to cross the ears. We're not using half of them. I'm not saying we, we were at work. We, we, we industry-wide, we, the, the, they're not very royal way. So the point is that we have added a lot of stuff, a lot of uh, uh, tools out there and, uh, a lot of people are just not using them or not using them to the full extent that they could be used. And I would even go a step further and say that a lot of tools out there, they don't make sense anymore. That there is a lot of stuff out there that, uh, has been overtaken, overridden by, by events or by reality. And you may not even need that tool. That's just churning back there and producing findings that somehow you have to deal with, even if it's to hit ignore. So I don't think that just throwing money and getting more tools is the response.
5:59Matt Coles Touch upon that for just a second. The tools are not producing quality outcomes, or the tools are producing still so much outcomes you're You get drowned, you're drowning in it.
6:13Izar Tarandach I think it's not a, okay, I'm just talking about the tools that you probably shouldn't be, sorry, you needn't be running anymore. It's not that they are producing a lot of stuff that you have to deal with. It's just that with all the changes in technologies and in approaches and all that, There are certain tools there, the names are irrelevant, that still keep churning in the back of your security toolkit and producing results that you should be analyzing because at the end of the day, you're paying for that and you got functionality and it's plugged in your other stuff. Can you— But that other tools are getting better.
7:01Chris Romeo Can you call out a particular class of tool? I'm not sure what you're talking about here.
7:06Izar Tarandach A bunch of scanners out there. Like SAST, DAST? Some stuff that in the past we have said is dead and should be dead and might be dead. And it's something scary.
7:19Chris Romeo He's trying to protect himself here without—
7:20Izar Tarandach I'm trying to be nice. It's Monday morning. Come on. And don't be nice.
7:24Chris Romeo We're never known for being nice on the security channel.
7:26Izar Tarandach I know, but people have been telling me that I should be nicer.
7:29Chris Romeo That it would make me— You know what you should tell them to do?
7:30Izar Tarandach You should tell them to shut up. I did, I did, but then mother said, you know, you don't talk like that to me. But anyway, so there is these tools and I don't know, I have this deep feeling that as Uri said, some of this is just craft, it's just like fluff, and some of it has even turned into an attack surface, which is something that the report actually takes into consideration later, saying that, hey, you know what, security tools usually work at higher privileges, and you may have a bunch of blunt tools in your toolkit that are out there and are still running with higher privileges, and now they just turn into more attack surface that is not being managed.
8:27Chris Romeo One of the things that I'm staring at this table in the report about the threat preparedness analysis, and I'm thinking, I'm going down this list and I'm going, which of these things existed when I started in security? Uh, all but 3 of them, I think. I mean, API-related vulnerabilities came later because we didn't have You could argue we didn't have API, not at the scale, not at the, not the way that this is.
8:56Matt Coles Yeah, not the way that tradition now—
8:58Chris Romeo credentials, software vaults. We didn't really have supply chain threats because software was still being written off the shelf at that point. When I started, there wasn't, uh, and there were some open source libraries, but it wasn't at the level obviously that we are now. Um, but insider threat, phishing, poor encryption, ransomware— ransomware came a few years into my career, but like most of these things have existed for almost 30 years. Yeah. And we haven't really stopped them.
9:29Matt Coles Well, some of these aren't threats also, but that's okay. I mean, poor encryption is not, poor encryption is not a threat. It's a flaw. True. Right. The threat, the threat is data theft.
9:41Izar Tarandach And some of them are not under your control. Supply chain threats. Okay. You can, you can scan and fix to, until, till the cows come home. But the, uh, at the end of the day, you're going to have to use, uh, a third-party, uh, surface. You're not going to write everything from zero.
10:01Matt Coles Well, okay. Maybe, maybe today you can, you can go back to writing, to writing everything bespoke.
10:07Izar Tarandach Right. And, and, and yeah, it's a no. Well, okay.
10:12Matt Coles But, uh, can versus should, right? We've, we've had this. True. Very true. Just 'cause you can doesn't mean you should.
10:18Izar Tarandach Very true. And when you start dealing with supply chain threats, you're going to, to start dealing with exceptions. But beyond that, beyond that, the threat stays the same. The attack surface changes, right? So once upon a time it was libraries, then it became packages, then it became, uh, dependencies. And now if you look at, uh, nowadays, hey, your cloud skills are a new supply chain. Tracked. Remember when it was just—
10:47Matt Coles remember when it was just containers you had to worry about?
10:52Izar Tarandach But that's the thing, containers, it was more of a, uh, an agglomeration of stuff, right? Because you were getting all the, the third-party stuff, putting it into a neat box, and you would be scanning that, that, that neat box. But just to connect to the previous, uh, uh, point, then somebody went and owned Trivi. So the thing that you're using to scan your containers all of a sudden is also a vector of attack. But, uh, all right.
11:19Matt Coles That's a, that's very recent. I mean, that's, that's within the past year, right?
11:23Izar Tarandach So last, last month, I think.
11:25Matt Coles Yeah. A couple of weeks.
11:26Izar Tarandach But, uh, the, the, the point is that some of these threats are going to be—
11:30Matt Coles and by the way, it wasn't just Trivy now, it's Checkmarx, I believe, right? Also recently.
11:35Izar Tarandach Believe so. And, well, talking, speaking of not naming the security partners, but, oh, you already did already. So, so, no, no, no, no, no, no, true, true. Yeah. But, um, so I think that what I'm saying is that some of these are going to roll into new forms and new technologies, but that doesn't mean that the old ones are going to go away. It's just going to say that, hey, now there's a market for new scanners and new tools. That are going to scan this new flavor of the month of yesterday's threat. So it's even more, it's an even bigger opportunity for tool creep type.
12:20Chris Romeo Ooh, new term. That's another good song for the creep cybersecurity professional mixtape by Radiohead. Creep by Radiohead.
12:29Matt Coles Because I'm a creep. As opposed to Tool, the band, just as a band.
12:34Chris Romeo Tool Creep. Tool Creep. Tool and Radiohead making a combination band. Tool doing a cover of Creep.
12:43Matt Coles Okay. Can you imagine?
12:46Izar Tarandach Yeah, but the show has to be by Phish. That's true. It is a PH. It's Phish with a PH. Or at least the selling of the tickets has to be by Phish. That's like a complete double meaning.
13:03Chris Romeo So the second hot take that I pulled out of here was what we talked about in the previous episode, just confirms what we concluded there with our brilliant take on human in the loop for Agentic. Uh, the, the report found that 77% of cybersecurity professionals expressed some level of comfort with deploying agentic systems and allowing them to take actions without human review.
13:34Matt Coles I have a, I have a bad take on this. Is that because plausible deniability? The tool did it.
13:41Izar Tarandach Yeah, but somebody pressed enter. Only once. Only once after another, another, another.
13:48Matt Coles Then you put it in a loop, let the tool do its thing.
13:55Izar Tarandach Mm-hmm. I would like to— Oh boy, what is happening? I would like to introduce you guys to my new keyboard. I just finished it this weekend. It has exactly 2 buttons.
14:16Chris Romeo Wait, do you want to move it to camera?
14:19Izar Tarandach It's 2 buttons. It's a— A red one and a black one. Okay, so the black one is enter. The red one is escape. This keyboard is an agentric keyboard. It's all you need to actually interface with your agent. Right? Okay. And the fun thing is it works. So yeah, so the The fatigue is, is, is true, but also the, um, responsibility transfer, as Matt said. The moment that I, I type enter, it means that I did a cursory review on what my agent plans to do, and I am totally fine with it going and doing it wherever the, the chips may fall, as they say. Now, one thing that I haven't yet, uh, come to complete terms is the, the, uh, um, what's it called? The autonomous agents and this feeling by some people that they are bound to become insider threats. I mean, I get it that, that some people are writing their own, their own agents and giving them like, I don't know. Their own OAuth, uh, uh, and by, when I say the own, the author's OAuth, uh, keys to go and interact with all kinds of systems. And then they walk away from the keyboard and the agent does its thing. And it's like, oh, it's like I'm sitting on the keyboard and I'm doing my thing, but actually I'm on the deck, uh, watching the clouds. And a lot of people keep telling me that this is the new inside the threat frontier. And, uh, I don't know, I'm, I'm, I'm of two, two minds there that, yeah, I, I, I see it coming. I see the possibility, but then I think from a security professional point of view, what are the things that I would like to happen before that actually happens? And for sure, identity management is one thing, but who's to say that that's going to be followed and it's going to be enforced and whatnot. And again, I see a lot of tools promising, oh, we're going to find all your agents and we are going to hold them accountable and whatnot. But I don't see it happening.
16:46Matt Coles This whole notion of shadow AI.
16:52Izar Tarandach If, if, if the agent is supposed to mimic the user and it does it well, Then is it really shadow AI? Like, how are you going to distinguish between a user doing something and the agent doing something on behalf of the user? You're going to look at efficiency? Oh, that session is much more efficient. It must be an agent.
17:17Matt Coles It must be an agent. Bob doesn't usually work through lunch. Yeah. Well, so you could look at it as it's automation, right? I mean, we, we developers have written scripts for ages and have it do things on their behalf. Yeah, exactly. So if the agent does it, if the agent does it well and doesn't do it, do anything unexpected, I guess, well, maybe it's all unexpected, but within bounds. Uh, is that a problem? Probably not. But now we're talking something completely different, right? We're not talking— well, I guess, are you talking insider threat because there's, uh, somebody is— somebody's job is being taken by the automation? Meaning they're, they're deferring— they're let— their, their brainpower is now the tool.
18:19Izar Tarandach No, I think that it's more of an insider threat because it could theoretically be corrupted way easier than a user. Like, let's put it like this. For the past, uh, I don't know how long Chris said, but, uh, for the past X tens of years, we have been fighting, uh, phishing, right? So we've been training people not to click on suspect uh, uh, links. Yep. And those things sometimes they don't come from pattern recognition. They come from here, look at this thing, read it, understand it, use your common sense, intuition, and don't click on it. Right.
18:58Matt Coles The computer can't do that.
19:01Izar Tarandach Yes. Although some people say, no, no, no, the computer can do that even better because it's doing an unbiased, uh, check on the blah, blah, blah, blah, blah. But no, you still have prompt injection. And although I don't think it's a threat, I think it's a very powerful vector to threats, right? And you can influence this agent that's acting on your behalf, I think.
19:28Matt Coles Well, in fact, in fact, that's a feature in some cases, right? Like, exactly. Like ClaudeBot, right? Send it an email and have it do actions for you, right?
19:36Chris Romeo That's what, uh, Steve Wilson in the— we did an interview on AppSec Podcast where his Open Claw bot got hacked because the security team sent— started sending emails to it that were from a slightly different domain than the domain that he works for, and it fell for it just like a human would.
19:56Izar Tarandach And, and that's a— that, that's actually a very good, uh, article. We, we should reference that at some time, but, uh, My point is that the agent becomes the insider threat once we have this illusion that because it's automated, because it has all kinds of guardrails, it's capable to stand attacks that a human wouldn't. When you only have to send a PDF with some crazy text inside it to have it turn into something completely different that it shouldn't be able to do. So I go back to prompt injection is not the threat. Is the vector. The threat is the identity capabilities that get way too more than they actually should be able to, right? So again, it's the same threats, it's just the attack surface changed. Hmm. Well, the ease of exploitation has changed.
20:52Matt Coles The capability, the human, right? The human may have been an admin. But now the agent is almost certainly more privileged than the human is.
21:03Izar Tarandach Exactly. Because it's trying to mimic something that the human would be able to do. So it has its own administrative, uh, uh, capabilities.
21:15Matt Coles So interestingly, I, I've been thinking about this, that really agents should have their own identity, but maybe what you're saying is if agents are acting on behalf of the human, The agent should impersonate the identity of the human.
21:30Izar Tarandach Mm-hmm. Mm-hmm.
21:31Matt Coles Mm-hmm. Because, because if the human, if the human has limited access capabilities, the agent shouldn't have any more access capabilities.
21:40Izar Tarandach Okay. What I'm saying is actually what you just said, the other— maybe they should be prepared instead of impersonating, but the way things are today, I've been talking to some people in, in similar positions as mine. The way that identities for agents seem to be evolving is that Joe Developer wants an agent. Joe Developer wants the agent to do the same things that he does. He gives Joe Developer's OAuth tokens to the agent. All of a sudden, the agent is Joe Developer. Now, for all the monitoring, for all the monitoring stuff that we have, we are going to see the agent as Joe Developer.
22:21Matt Coles Right. Because it's impersonating him.
22:23Izar Tarandach What should happen— Exactly. What should happen is Joe Developer should have a way to register that agent with the organization to say, hey, I have an agent. It does A, B, C, D, E. Give me an identity. And now you can monitor that identity for anything that falls out of the bounds of doing A, B, C, D, E. But it shouldn't have any more— So now you have a way to check what it's doing. But you're right, on a, on a, on a, on a programmatic way, if the agent is trying to mimic what Joe Developer is able to do, it shouldn't have any more permissions than what Joe Developer is doing. If Joe Developer is writing an agent that's going to work as Susie Admin, then yes, it needs to, the the credentials, the level of operation that Susie Admin would have. But that's a different thing. If you give it its own separate identity and you make it known to the security team, hey, these are the boundaries this thing should be operating. It's going to be operating as an admin, but it's only going to do move files from here to here, open ports here, close ports there. Then security can monitor the thing based on expectations. And anything that falls out of that, all of a sudden you don't have to draw this extended behavior profile. Oh, the agent is doing something that we didn't see it do in the past 2 weeks. No, it's only expected to do ABC. It's doing something that doesn't look like ABC. Shoot that thing.
23:56Matt Coles You know, it's interesting. You, uh, this goes to an interesting point I've been thinking, well, not just thinking about, but talking about proactive versus reactive. If you're monitoring what the agent is doing and looking for bad behavior, sure, you can find it and take action, but you're being reactive. You're reacting to the behavior of the, of the agent, not being proactive and saying, put it in a sandbox. What can it do? And don't let it do anything else. Right? So what you're, because I think what you're, what you're ultimately highlighting is a challenge that security teams and, and the security systems themselves. Forget the tools, right? That's all, that's all reactive stuff. Think about the permission system and how granular or not granular those permissions are, right? What we, when we develop agentic workloads or workflows, and so in order to work to develop agentic workflows, you have to take and take the action that a human was doing and break it down into its representative parts, right? Well, you really should be doing that with authentication and authorization systems because it's those individual— let's go back to capability-based security, right? What capabilities should the agent have, right? So it can't do more than it should be allowed to do. And you won't get it 100% right. You do still need the reactive side. But you can be proactive and not give it full capabilities.
25:35Izar Tarandach So it is the threat modeling aspect of not giving it too much agency, right? But the point is that when you started talking about this bit, I was thinking, okay, he's going into the human in the loop, human on the loop thing. But no, no, no. Actually, I think that what you're highlighting is that nowadays we do have these mechanisms of runtime observability, where you can see what your thing is doing, but we've been looking at it from the point of view of performance. Oh my God, my daemon is going crazy and resource— not even before we get to the AI bit. We have had this observability capabilities for a long time now and people are used to using them, but from a monitoring point of view of bandwidth usage, storage usage, this thing is running away, it's not that loop, I don't know what is happening, but you have those ways of looking inside something that's running and saying it's acting well, it's acting bad. Now we can turn those things into agentic capabilities and start saying, I'm not only observing, but I have these hard guardrails that these are the agency guardrails that I'm giving to my agent. I don't expect it to do anything but these things. Or inside these areas, and then alert for anything that happens outside of that.
27:01Matt Coles You know, this is a system like what Android implemented with intents, right? You have an application that has certain intents. Now, those intents are, they're sort of, what are they? I'm trying to remember the quote from Pirates of the Caribbean. They're more like guidelines, really.
27:21Izar Tarandach Yeah, I think that they were more of an advice of how you're going to use it. And there was a cascading system in there saying that, If you don't have any intent that captures this specific thing, then fall into this generic case of, I don't know, opening a browser or whatever, right? So it's exactly that. It's the cascading expectations that you have of some kind of system that's running and how it's going to deal with exceptions. And what I'm saying is. We shouldn't have agents deal with exceptions. Exceptions should be automatic. I give up. I don't know what to do with this. And then you treat it on the monitoring side by saying, hey, my agent got something that it doesn't know what to do with. Let me go and look at it. And this is not human on the loop. So this is actually a responsible approach to the agency of the thing that you just wrote.
28:17Matt Coles So is that what In effect, is that what NVIDIA built with OpenShell? In other words, the agent framework or agent harness is able to do the enforcement and keep the AI within balance.
28:31Izar Tarandach I don't know, I don't look at that. So tell us a bit about it.
28:36Matt Coles I only know a little bit. I've only just started to research it, but OpenShell I think is a controlled sandbox environment for agents. Well, specifically for things like OpenClaw so that it can only do— I have a sense, although I don't know for sure, so don't quote me on this, that it functions in part like a jail, like a true jail or something, or just a jail in PSD, or like a— I'm trying to think now. It gives a container for the agent to run in with only what it's been granted access to. But I don't think it's been— it's not like the agent itself has been privileged. It's that the environment that the agent operates in has been privileged.
29:31Izar Tarandach So here's an anecdote. I think that I even mentioned this in the past episode, but for some reason I had to play with Claude's cowork. Don't ask why. And I wanted to be very, very, very careful.
29:47Matt Coles Had to? Yes.
29:49Izar Tarandach So I used my gaming box to play with. It's the only thing in the house that I have that runs Windows. And well, yeah. And I decided to go one step further and give it a USB key to work with. I wanted it not to write or read anything on my disk. I said, hey, your playground is this USB key. And assumption, nothing but an assumption. I think that Microsoft must, uh, that, uh, Anthropic must have done something in there. And the thing thought that it was being successful in writing files to the USB key. But in reality, it wasn't. So at some point, I had to screenshot the freaking file manager and say, no, the files you are telling me are in this key are not there. And yes, I flushed the memory and I checked that there wasn't anything pending and all that. No hidden files. But that's not the point. The point is that at some, like after 3 bounds of the files are there, no, the files are not there. I started to—
31:00Matt Coles Did it say, did it say, oh, you're, you're totally right. There are files.
31:06Izar Tarandach At some point, I think it did, but the point is that it started trying ways of writing those files that to me looked very much like what I would do if I was trying to escape a jail. So it started writing, um, Python scripts that were encoding the contents of the file as Base64 and trying to write it to a terminal that would then copy it, like copy from the terminal into the file. And A, it was fascinating. No. B, it very much gave me the feeling, the impression of something that very much wanted to escape the boundaries that it was put into. It didn't make a consideration of, hey, perhaps I shouldn't be trying to write on that directory. Perhaps you should give me a different directory to write to. That never came up.
32:09Matt Coles Because, because of the training bias, right?
32:13Izar Tarandach It went single-handedly towards the mission that I gave it, which was write these files in this directory. It wants to appease you.
32:20Matt Coles It, that's its training bias. It wants to be successful to your, to what you're asking it to do. Right now, interestingly, because you're on Windows, I wonder if you check alternative data, alternate data streams, you know, ADS or, um, what's the other technology there that Windows supports for having, uh, multiple paths to the same file. It may have written, it just wasn't visible. Use your forensics knowledge to see if it's still there.
32:49Izar Tarandach I was very—
32:50Matt Coles Maybe it wrote to the block device and didn't write to the file system.
32:53Izar Tarandach At some point, I was very afraid at some point that it was going to start thinking about downloading a compiler and just trying to write its own binary and try to go through memory. You gave it access to the internet?
33:09Matt Coles And USB key. It's cloud co-work.
33:13Izar Tarandach You have to give it the tools that it needs to do your work, right?
33:17Matt Coles I did read over the weekend that somebody connected a local LLM to cloud and it worked, so. To Cloud Code. To Cloud Code, yeah.
33:27Izar Tarandach Yeah, yeah, no, you can run.
33:28Matt Coles Or Cloud Co-Work, whatever.
33:32Izar Tarandach It's not only Cloud Work, it's, I think most of the popular harnesses, you can, for example, run ollama Claude blah, blah, blah, name of the model, and it will start Claude talking to that model on Ollama. So it's going to download the model, it's going to bring it up, and it's going to run Claude talking to it. It's actually very fun. I have it on the RPI here with a Gemma 4. And if you have patience, it's actually—
34:00Matt Coles because you run it on a Raspberry Pi.
34:02Izar Tarandach Yeah. Yeah. I have an RPI 5 in here. But, um, it's like talking to a sloth.
34:12Chris Romeo It's, I want to bring us back to this report for one more statistic because I think it's, this is an interesting one for us to, to confirm or deny. And so what Avante found, 43%, and I looked down further, they surveyed like 1,215 people. So we're talking about, you know, 500 people or so from those surveyed. Amongst— 43% amongst security pros surveyed report high levels of stress due to their jobs only.
34:49Matt Coles I don't know what they're talking about. Only 43%? Whatever happened to the 'it's all bad all the time,' 98% of cybersecurity professionals being stressed. So, you're arguing that this is an improvement? It sounds like half of the security people have suddenly found bliss. No, no, no, no.
35:10Izar Tarandach Half of the security people just had heart attacks and don't count anymore. But I have been thinking about retiring as the IT manager of a Zen monastery, but I think that even that would be a bit too Too stressful. No, look, I think that we spoke about burnout in the past. Yeah. The good thing about now is that, uh, now you have this whole surface that you can use to bitch about being burned out. And so far it's even if you use a locomotive, it's not going to, uh, talk, talk about it to other people.
35:48Chris Romeo Oh, you're talking about using a model or an LLM as a therapist. Is what you're describing.
35:53Izar Tarandach Yes. Uh, not as a therapist. No, no, don't go there. Very dangerous territory. Do not go there. Kids do not try this at home or anywhere else. True. But, um, it, it, it, it's a, it's a great place to, um, commiserate.
36:10Chris Romeo So wait, the model commiserates with you?
36:13Matt Coles Yep. It always wants to appease you. That's a training bias. Mm-hmm.
36:18Chris Romeo True.
36:18Izar Tarandach Which is annoying. Mm-hmm. When you want to commiserate, it's great.
36:24Matt Coles So if you want a friend, always say yes.
36:27Chris Romeo So another song for the security mixtape. You've got a friend in me.
36:35Matt Coles As opposed to I got you, babe.
36:37Chris Romeo That's weird now. All right. You crossed the line. Crossed a line. Do not refer to your LLM as babe.
36:45Izar Tarandach It's more like a bridge over very, very, very troubled water.
36:54Chris Romeo This is a heck of a soundtrack.
36:55Matt Coles Sweet Child O' Mine.
36:58Izar Tarandach Sounds of Silence. Getting back to Guns N' Roses.
37:03Chris Romeo Everything returns to Guns N' Roses is what I'm learning. They're part of the fabric of the universe, holding this whole thing together. Yeah.
37:15Matt Coles So stress, Or for the stress case, jump. Jump.
37:21Chris Romeo Okay. That was not appropriate at all. Not appropriate at all. Not appropriate at all.
37:30Izar Tarandach No, no, no. We do not condone that.
37:32Matt Coles Don't try this at home. Don't try this at home.
37:35Chris Romeo Don't try this at home. So yeah. So Izar, you agree with this? You think it, you're with Matt that you think it's, it's actually better now than it used to be if we're at 43% or?
37:45Izar Tarandach No, no, no, no, no, I don't, I don't, I don't, I don't. I think that we worked so long about hardening systems, we forgot to, to a certain extent, harden people. And now we are dealing with the burnout and mental health debt that we built over the last 30 or so years.
38:10Matt Coles Yeah. I mean, the, the report says 43% report high levels of stress. So yeah, probably everyone is stressed.
38:18Chris Romeo Some people feel it more.
38:23Izar Tarandach I mean, you just have to go to an OWASP event and see the line for coffee.
38:29Matt Coles No, that's a misleading, that's a little bit, that's, that's causality versus causation or correlation versus causation.
38:36Izar Tarandach You have, I don't know.
38:38Matt Coles I, I, I look at the line, you're at a conference and people are sleep deprived, that doesn't mean they're stressed from their jobs.
38:45Izar Tarandach I look at the line for coffee on an OWASP event and it reminds me of string theory because everything is vibrating and that creates the fabric of the universe. Everybody is vibrating. Did you hear about the last one? No, no, no. Did you, did you? No, no, I was the last one.
39:02Matt Coles If you had little devices that could capture kinetic energy to electricity and you could beam it, like you could power by conference alone.
39:12Izar Tarandach Oh, you could power the conference. Easy, easy. Totally easy.
39:16Chris Romeo This has really taken a number of left turns.
39:21Matt Coles Well, then eventually we'll get it right.
39:23Chris Romeo Somehow we've landed on powering conferences. I don't know if you've ever seen Monsters, Inc. with the scream. Oh yeah, yeah, with the scream.
39:32Izar Tarandach But wait, I'm going to go one further. What could power a data center better? The vibrations of an OWASP conference or RSA? Now remember, there is an alternative here. Burning shit is also a form of energy production. So the vibrations of OWASP conference or burning the bullshit of an RSA conference? But wait, wait, wait, there's more. Oh God. The OWASP conference, the RSA conference, or the power spent sending spam after an RSA conference.
40:15Chris Romeo Now that's a huge amount of resources that are applied.
40:21Izar Tarandach If you could unleash that amount of power in one single blast, uh, I think it would be more powerful than, uh, the, the, the, the, the Death Star.
40:37Matt Coles Well, you know what it is? It's, it's, it's somebody at RSA who realizes that they haven't used their token budget for the month and they just need to burn it all at once.
40:48Chris Romeo Ooh. It always comes back to the token budget. It does. Authentic sampler. That's what I hear. All right. Well, I think this is enough. May the force be with you. So I don't really know how to describe what just happened.
41:02Izar Tarandach Energy. A lot of energy. A lot of energy was expended.
41:06Matt Coles Uh, and breathe and see data lights.
41:09Chris Romeo Blinky, blinky lights were flashing in data centers around the globe. Tokens were consumed and the security table, the original security table, who would, that also has a really cool keyboard. So I bet you there's not, I bet you there's other ones.
41:28Izar Tarandach Available in our, in our, uh, uh, merch store. In our merch store.
41:33Chris Romeo You two can have. Yeah, but we only have one, so it costs $4 million, that keyboard, so.
41:39Izar Tarandach Yeah, I'm gonna sign on the bottom.
41:41Chris Romeo Well, then it's worth less if you sign it, so please don't sign it. Uh, but I guess I will say adieu to our Audience, much ado about nothing. This has been an episode of the original Security Table.
Transcript supplied by AssemblyAI. It may contain transcription errors.
Originally published as The Security Table. Part of the AI Security Table archive.
