40 minSeason 4, episode 11

The Agentic Access Problem: When AI Becomes Its Own Administrator

With Chris Romeo, Izar Tarandach, and Matt Coles

AI security & autonomous agents

In this episode, we explore what happens when AI agents meet the security principle of least privilege. As agents gain the ability to request permissions, make decisions, and interact with systems on our behalf, the line between human and machine responsibility starts to blur.

Part of: AI agent security: agents, MCP, and prompt injection

Episode title image: The Agentic Access Problem: When AI Becomes Its Own Administrator

Listen to the conversation

About this episode

In this episode, we explore what happens when AI agents meet the security principle of least privilege. As agents gain the ability to request permissions, make decisions, and interact with systems on our behalf, the line between human and machine responsibility starts to blur. The discussion covers prompt fatigue, over-permissioned agents, and why "because the agent told me to" may become the next security anti-pattern—before taking a hilarious detour into EULAs, cookie notices, and Matt's unexpected habit of reading both.

🚀 Join the Conversation

If your AI agent requested administrator access right now, would you know whether it actually needed it?


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 · 40 min conversation

0:08Chris Romeo Hey folks, welcome to an in-process episode of the original Security Table, not to be confused with any imposters under the table out there on the interwebs. So Isar and Matt were just beginning a debate about, I think one was taking the side of mythos is gonna save the world and the other was saying it wasn't. But perhaps you guys could wait, wait, wait, wait.

0:37Izar Tarandach Were, were you listening to the same conversation that we were?

0:40Chris Romeo Is that, no, I only half listen to any conversation that I'm in.

0:43Izar Tarandach Good, good, good.

0:44Chris Romeo So I probably didn't hear exactly what you were saying because I was, Chris has his AI, uh, record the conversation so he can play it back later. Yeah, this is just the AI generation of my voice actually with the AI's answering.

0:57Matt Coles And of course, AI is in security news all over the place these days. So we have to—

1:02Chris Romeo What does AI stand for?

1:04Matt Coles AI?

1:05Chris Romeo What?

1:06Izar Tarandach Absolutely insufferable.

1:08Chris Romeo True. Uh, then once again, to our future robot, uh, overlords, uh, that was Izhar said that.

1:20Izar Tarandach Yeah, but it's, it's not that we don't like you. We just don't want to inconvenience you. Okay.

1:27Chris Romeo So what were you guys, what were you saying about—

1:28Matt Coles we like, we like suffering by doing it manually.

1:31Chris Romeo Yeah. What were you saying about Mythos here? So like one of you was making the argument that Mythos should be able to fix the problems. Can somebody set that up for me?

1:39Matt Coles Well, no, we were playing with it. As everyone, as everyone should know by now, Mythos is the, is the, uh, what's the word I want to use? The, the pinnacle of, the pinnacle of security. Today, uh, as a— for those who have access to it, and if you have access to it, call us. Um, for those who have access to it, uh, Mythos is of course the latest model, the latest model from Anthropic, or a recent model from Anthropic. I actually don't know if they're the latest anymore, uh, for bug hunting. And of course it's, uh, it's pretty popular because it's been under restricted access, uh, through that Project Glasswing. But there's been reports of, uh, of it finding vulnerabilities in, in very, uh, interesting places that have never been discovered before. And so the question was, if, uh, today, if you have human coders who are developing, or architects, architects and coders who are developing applications, all the, all the plethora of applications and systems that exist out there, Mythos is gonna almost certainly gonna find vulnerabilities, right? It found, uh, it found a curl issue. It found multiple Linux issues. It found, I believe, at least if I remember correctly, I have to go look at the sources here. Um, it found a kernel bug, I think, in Apple macOS. Uh, there's a host of other things, uh, that's been attributed to Mythos. Um, and so human coders are going to, are not infallible and will, you know, generate bad stuff. Almost certainly. We know this. I mean, this is why OWASP and OpenSSL and others exist, right? Is to try and minimize this. But then I was pondering, well, what happens when you have spec-driven development or vibe coding and the AI is developing the application and you have AI and analyzing the application for security. At some point, will the analysis feed the development that they can have a feedback loop to prevent coding issues? Because the alternative is the AI injects coding issues for Mythos to find, which is a self-fulfilling prophecy. Book, go buy Anthropic now.

4:07Izar Tarandach So here's the thing. I think that on a theoretical basis, perhaps you got something, a beginning of something in there. Although I think that the gaps between discovery and training and use play a part in here. But, I have difficulty thinking how that would work when in the real everyday world when you write, okay, let's call it a spec for the lack of a better word, to start coding something and you preload your set of amazing instructions. As soon as that window starts hitting 65, 75%, those amazing instructions are being forgotten and you don't even know how many times have people told the thing to use a certain tool to, I don't know, quickly read code instead of grep or whatnot. And over time, you start developing your thing and you notice that your tool stops being used and grep gets used again. Why? Because the agent forgot that bit of the context, which is still there, but the magic that is AI, forgets to use it. And when you call— no, no, seriously, that's how it feels. And when you call it—

5:32Matt Coles that's how it feels or how it is? I mean, just feeling personally.

5:35Izar Tarandach That's how it is, but it feels like, it feels like you're basically programming with your grandma, right? Because— hold on, explain this.

5:45Matt Coles Explain this.

5:45Chris Romeo How is this like programming with your grandma?

5:47Izar Tarandach It's like, okay, use Graphify every time that you have to read source code, okay?

5:52Chris Romeo That's what you tell your grandma?

5:54Izar Tarandach Not anymore, but— she hit you with the wrong advice.

5:56Matt Coles Am I swearing at her?

5:58Izar Tarandach But wait, wait, wait, wait, wait, wait. But then you go, you write, you write, you write, you write, and you say, okay, let's find where I wrote the blah, blah, blah, blah, blah. And the thing goes grep. And you look at it and you go, wait, but I told you a moment ago to use Graphify to read this thing. Oh yeah, you're right. I forgot. I should make a note to do that next time. And then the thing goes and makes a note and makes this big show of adding it to its agentic memory and shit and whatnot. And then on the next paragraph, you go and you say, okay, let's find where we define blah, blah, blah. grep. Dude, the fuck? Kidding. I just told you to use the tool.

6:37Matt Coles This is the rated R episode.

6:39Izar Tarandach Oh yes, you're right. Let me make a note of it. I mean, if I go and read the memory of my thingy, I wonder how many times the same note appears in there. Every time with more strident words of, oh, I should really remember this thing.

6:55Matt Coles So I thought, I thought RAG was supposed to solve this.

6:58Izar Tarandach No, no, no. RAG is a different thing, right? So you have the, uh, this, okay, let me caveat this. This is a hobbyist talking. I have absolutely no training, no understanding, no anything of AI. This is just what I figured out from day-to-day use. So I could be talking out of my cordon here.

7:24Chris Romeo Way to sell your expertise here.

7:26Izar Tarandach Right? No, I'm just caveating this just in case somebody who actually knows how this stuff works is on there.

7:32Matt Coles It's the humanist approach to—

7:34Chris Romeo I can let you in on a little secret. Nobody understands how this stuff works.

7:39Izar Tarandach There are some people who very forcibly put out their voices that only they understand. So to those prophets out there, just in case one of them is listening to us, I says, my precious.

7:53Matt Coles And there's all the experts on Maltbook who probably will weigh in.

8:00Izar Tarandach That too. So where I was going was, you have your context, you have your reg. Using the magic that is, wow, so much magic. Using the magic that is embedding.

8:12Matt Coles Pixie dust.

8:13Chris Romeo Okay.

8:14Izar Tarandach Using the magic that is embeddings, something then reaches into your RAG, brings the context that was in there related to the thing that you are trying to do, puts that into the context and kicks that back to the LLM. So the LLM doesn't know about the whole RAG container and go searching for stuff in there. It's something that's prepackaged. The context goes there, gets magicked away, and comes back to you. So it's two different stages of the thing. And the agent is not going to hit the RAG all the time for, hey, how should I be writing code? It will be looking for like examples of past code of the same context or things like that. It's not going there to figure out. For the lack of a better word, personality. We are lacking so many better words to deal with this AI thing. So I think that it's different stages of the pipeline. And this thing that the agent forgets instructions that it got a while ago, I find that very annoying because clearly you put stuff in there to save tokens and to put your code in a certain way, apply certain standards, requirements, and whatnot, and the thing goes and forgets about it. So if I needed something that forgets about everything every 5 minutes, I would do it myself.

9:50Matt Coles Why not just re-inject into the context?

9:52Izar Tarandach Because then it, it, I think that that's what happens when you do the compact thing, right? It like does a Swedish Chef thing and woo, woo, woo, woo, woo from the context.

10:08Matt Coles Really? Does it make sound effects? Because that would be awesome.

10:13Izar Tarandach Actually, you can do that with a hook.

10:15Matt Coles Does it have like the Knight Rider, but I imagine it's in blue instead of in red going back and forth. Mwah, mwah, mwah, mwah, mwah, mwah, mwah, mwah, mwah.

10:25Izar Tarandach I'm sure that somebody somewhere has already done that.

10:30Chris Romeo Added an LLM to an '82 Trans Am.

10:35Izar Tarandach And somebody clearly has done something with a prompt for the thing to always be saying, "By your command." So it probably happened somewhere. But no, that's a different episode. I did some stuff with TTS models. So that Claude, whenever like I run a bunch of them in Parlor and I always get lost who's doing what. So I have this hook where it passes to the synthesizer to tell me which window is calling for my attention.

11:08Chris Romeo The synthesizer?

11:10Izar Tarandach Yeah, yeah. And it's pretty cool unless I am in the middle of a meeting, then it's very vexing. But because I'm on the meeting on the headphones, but the thing routes through the speakers. So sometimes the people in the meeting go, Claude is looking for your permission on window blah, blah, blah. And yeah, Claude, go ahead, do your thing. But anyway, so why would we—

11:36Matt Coles Sir, yes sir, may I have another, sir?

11:38Chris Romeo Well, they say in one of the documents that we were potentially looking at today about the one about the, uh, AI, the Agentic AI services, um, kind of documented about, you know, kind of security features and whatnot for Agentic. One of the things it said was privilege risks over-permissioned agents could be a potential challenge.

12:08Matt Coles Mm-hmm.

12:09Chris Romeo So it sounds like that's where you were, uh, where you are in your own investigations.

12:14Izar Tarandach Well, we, we've been talking about that for a while, right? And right.

12:17Matt Coles Our agent agents aren't humans, but what are they? In terms of permissioning?

12:23Izar Tarandach Yes.

12:24Matt Coles Do you make it like early days of OpenClaw where, here, take my stuff, please.

12:31Izar Tarandach And do something with it because I'm not. But I think that for a while we've been telling, like from the early days of prompt injection, that the problem wasn't the prompt injection, it was the over-permission, right? So I think that we were prophetic in there as well.

12:51Matt Coles You know, least privilege has always been a thing, right?

12:55Izar Tarandach Yeah, that's the thing.

12:56Matt Coles We seem to have lost sight of basic security principles.

12:59Izar Tarandach It's not that we lost sight.

13:00Chris Romeo Did we ever really have them in sight though?

13:04Matt Coles I think we had them in sight. They were sort of like the vision on the hill, you know, the holy, like from Monty Python, it was the Holy Grail on top of the castle, you know, from that scene.

13:15Izar Tarandach I think that they were inspirational, but that we were making strides towards going there, right?

13:23Matt Coles We invented a thing called stride. Not we. No, no, we invented stride for that purpose.

13:30Izar Tarandach It was invented. It was born. It's self-generated. But, um, I think that in between all the just-in-time training that we we have. And I think that even people who are able to package and sell, and I'm opening a huge air quotes here, uh, security prompts for agents, right? It feels like we were able to, at least in some way, shape, or form, to codify the knowledge of secure coding. So you could say that that's a sign that we were getting there.

14:10Matt Coles Well, secure coding is easy, easy, but in your quotes here, right? That, well, 'cause they have, there's defined constructs, right? It's a more defined problem space, right? You can look at a piece of code and say, I have a buffer overflow, and you may intentionally have implemented it, or you may have had some, math error, right? Buffer overflows happen because of math errors mostly, right? You, you, you, you have a 0 to 32 buffer, you put 33+ in it, all bad.

14:44Izar Tarandach You have a math problem. Stop making errors.

14:45Matt Coles It's a math problem.

14:46Izar Tarandach It's not a math problem.

14:47Matt Coles It is a math problem.

14:49Izar Tarandach No, it's a math problem.

14:51Chris Romeo It's a math problem.

14:53Matt Coles It's, it's a PepCap problem.

14:55Izar Tarandach Uh, Matt, stop making math problems.

15:02Matt Coles Uh, but security principles are hard.

15:07Izar Tarandach Yes.

15:07Matt Coles A problem. A question. What is the least privilege?

15:15Chris Romeo I feel like this is a trick question.

15:16Matt Coles It is not a trick question, but it is not a trivial one either.

15:20Chris Romeo Yes.

15:21Izar Tarandach Okay.

15:21Chris Romeo So I think that's the least amount of privilege that a process can, can still perform the work that it needs to do.

15:29Izar Tarandach Right.

15:30Matt Coles And how much is that? How do you quantify that?

15:36Izar Tarandach Once you close it enough and things stop working, you just found the boundary of the least. My problem is, and my fans agree, my problem is that, uh, Once upon a time, it was easy to think in terms of those things because your privileges were, can I read this file? Can I read that one? Can I write that one? Can I open a socket?

16:05Chris Romeo Yeah, I mean, this is least privilege and this idea of privilege came from the Unix context where you were shedding privileges from a particular process.

16:16Matt Coles Well, you had, you still had a current, you still had a kernel, the kernel owns resources. And then as application wants to use those resources, how much should the kernel let it do that? Right. That was that notion of least privilege. Don't have, don't give it more capability than it needs.

16:33Chris Romeo Then it needed. Yeah.

16:34Matt Coles Right.

16:34Chris Romeo But that, I mean, that came before now least privilege gets applied to everything.

16:38Matt Coles It does.

16:39Chris Romeo It was a construct that existed in a Unix.

16:42Matt Coles But the privilege model now is different because originally it was a single application on a single system. The system owned all the resources. Now You have, now you have something broader than that stack of privileges, right? And a stack of systems that are, or things that are part of that privilege layer.

16:56Izar Tarandach It's what is your local privilege? It's what's your privilege inside of the deployment that you have? It's what's your privilege in terms of assessing other systems?

17:06Matt Coles And, but it's not, it's not even just resource access now. It's, it's broader than that, right? Because, because the AI is, uh, The AI is a human in a box, right? Effectively.

17:17Chris Romeo Oof.

17:18Izar Tarandach Oof. Okay. We, we'll cruise over that one, but, uh, okay.

17:26Matt Coles It does language, it does language processing and makes decisions.

17:34Izar Tarandach Yeah.

17:35Matt Coles And can, and can use tools.

17:39Chris Romeo Yes. So, yes, it is. Wait, so that's what defines a human?

17:41Matt Coles Yes.

17:42Chris Romeo We're getting philosophical now.

17:44Izar Tarandach No, no, no, no, no, no, no. What defines a human is doing all that but being way more annoying on the way.

17:51Matt Coles So AI is human in a box.

17:55Izar Tarandach Yeah, you know, I'm dying to say what else is a human in a box, but that's so morbid that I will stop myself from doing that.

18:02Matt Coles Okay, so you know how our podcast has like a soundtrack, so maybe this is Alice in Chains' Man in a Box.

18:11Chris Romeo I mean, we could just sing it here for more context.

18:13Izar Tarandach I have to go back to that one.

18:17Chris Romeo But if we have enough budget to cover an Alice in Chains song over the top of the podcast.

18:23Izar Tarandach But Matt, where are you going with the human in a box?

18:26Matt Coles I was just talking about least privilege.

18:29Izar Tarandach Oh yeah, being in a box would definitely put me in a—

18:34Matt Coles How do you tell the human in a box Or how do you allow the human in the box to decide least privilege when the humans outside the box can't decide least privilege?

18:50Izar Tarandach But that's the thing, you can decide what it is. I feel—

18:55Matt Coles You could decide a threshold. I think you could decide a threshold. I don't think you can decide—

18:59Izar Tarandach Personal experience here. And only personal experience. I think that the problem is not that people don't know how to do this privilege. I think that people don't see a point in, in doing this privilege. Plus, there is a very creeping feeling that comes from working with certain developers where they say, I'm just going to run this thing as root. And you know what? I never know what privilege it's going to need to be needed tomorrow. So I might as well just let it run as, as give it all, give it all.

19:33Chris Romeo Isn't that more of a classic problem though?

19:35Matt Coles That is a classic problem.

19:36Chris Romeo In this modern day and age, you can't really run something. Can you even run containers as root anymore? Uh, yes. You have to go out of your way. It's not like it's, it's not coming default like that. You gotta come, you gotta make a design decision to say, I'm going to run as root.

19:53Matt Coles Right. Or yeah, with excessive, with privileges almost certainly beyond what you need, yeah.

19:58Izar Tarandach You have to put an asterisk here and there, but I think that it's still doable and easily doable. And even, oh, nobody told me that I shouldn't do it this way.

20:10Matt Coles If anyone, well, actually, if you ask an AI, all right, bad idea here, but if you ask an AI to tell you what the least privilege is, you may not. And you, and if you're a new developer, or if you're, say, a new developer who's vibecoding because you don't have a development background, will you know what least privilege is? Developers who go to school for this sort of thing, or developers who have experience with this thing, know what, what that, at least that concept is.

20:38Chris Romeo Where'd they learn that from?

20:40Matt Coles What, from all of you guys? From us? No, no, wait, wait, our security teams, right?

20:46Izar Tarandach Wait, so first of all, developers—

20:48Chris Romeo I wish I could, I wish I could say this with that, with that amount of confidence that all the developers know what least privilege is.

20:54Matt Coles Do you?

20:54Izar Tarandach No, I, I, wait, I think that all developers know there's a line.

20:58Matt Coles Haven't you delivered this?

20:59Chris Romeo Of course, of course. But I mean, I'm talking about the, the, I'm thinking about the universe of developers that exist out there, the, which probably now GitHub's studied probably 32 or 35 million people. On this planet, how many of them know what least privilege is? And only a very small percentage have been trained. It's true.

21:21Izar Tarandach Look, I heard of people, how can I know what the least privilege is before I know all the privileges that this thing is going to need? So I'm going to do the whole thing and once it works, I'm going to start taking privileges away.

21:33Chris Romeo That's how we've always done it.

21:36Izar Tarandach Right?

21:37Matt Coles But, uh, but the behaviors, the latent behavior, the behaviors that come out of AI make that difficult.

21:45Chris Romeo Right. Yeah.

21:46Izar Tarandach I'm thinking that it's not, not even the behaviors. I think that it's the speed. And I think that again, the, the, the argument that I don't know what I'm going to need tomorrow, so I might as well just give it everything. It's very tempting, but when you put on top of that the layering of privileges and people all of a sudden have to think. About different layers that before were like, oh, somebody's going to take care of that. I'm just going to deploy whatever Terraform or CTS is, or what's the name of the other thing? Anyway, any of the infrastructure as code stuff, and all the privileges are going to be embedded in there. And to an extent, that's still true to that layer. Layer up, layer down, talk to a new system. That there might be access stuff even in prompts, right? One prompt could require certain access and not another slightly different might be interpreted as something that doesn't need that one. So how do you deal with that now? And how do you deal with the AI coming in saying, hey, I need that other one that you didn't know and just go into the console and enter those stanza here and I'm going to have everything that I need.

23:09Matt Coles There's also intent, right? The permission that you give it versus what you intended to use it for versus what it could use it for.

23:19Izar Tarandach But what I think that I'm pointing out here is that the AI is going to always ask what it needs to get to the goal, right? It's a greedy algorithm. It doesn't really care about your feelings and your politics. It's always going to ask, I need more access. And people now are self-training to give it whatever it asks.

23:41Matt Coles You know, that's interesting that you say that too, because your pre-AI application could never ask.

23:49Izar Tarandach Yeah, exactly.

23:50Matt Coles More importantly, you could tell it, you could code it to ask, but you'd have to architect that yourself. Now the system can ask.

24:00Chris Romeo That's a really kind of scary proposition.

24:05Izar Tarandach Uh, I'll give you the, the next step that's even scarier. The AI asks for the, uh, the privilege. You, you, the developer, go into the ServiceNow interface that you have for asking the platform team for privileges. And they very rightly ask you, why do you need this privilege? And the answer is, because my agent told me.

24:32Chris Romeo And their answer is, oh yeah, okay.

24:34Matt Coles Yeah, okay. Well, because it's— because, you know, you're probably talking to— potentially talking to an agent.

24:40Izar Tarandach You are potentially talking to another. So at some point they are going to talk to each other, and we are just going to look— well, in the good, good case, we are going to look from the side. In a bad case, we're just not caring, let them do whatever they want.

24:51Matt Coles And let them do whatever they want.

24:53Chris Romeo Yeah, come on, human in the loop, we still need it somewhere.

24:58Matt Coles Well, we still need it. Will we have it?

25:01Izar Tarandach Many times the human in the loop passes out when it reaches the top.

25:08Chris Romeo Wait, what causes the human in the loop to pass out? Just exhaustion or?

25:12Matt Coles No, no, have you ever seen the, the, that slingshot, that slingshot ride where they go up and down, up and down and they, they reach the peak and then they pass out from fear or or, or, or, or gratitude, gratitude, putting your, your blog somewhere, somewhere else.

25:26Izar Tarandach But, um, how many role—

25:27Matt Coles Human in the Loop. Yeah.

25:29Izar Tarandach Human in the Loop is not a very reassuring, I mean, I know you—

25:33Matt Coles Especially with, uh, was it prompt fatigue, right? So, a what? Prompt, prompt fatigue.

25:38Izar Tarandach Definitely.

25:39Chris Romeo But apart from that, that's a thing now, prompt fatigue.

25:42Izar Tarandach It's been, it's been, but, uh, where, what is it?

25:44Chris Romeo Prompt fatigue on me as the operator or like the AI gets tired of me?

25:49Izar Tarandach No, it's when you have been prompted so much to read something and press enter if you agree that you stop reading it, you just press enter because nothing bad happened before. Why would it happen now?

26:00Matt Coles And at some point, at some point you just run, run the agent with allow all privileges.

26:05Chris Romeo Gotcha.

26:08Matt Coles Cause I don't want to be bothered to review every, every request.

26:12Chris Romeo This reminds me that how glad I was when the Brave browser came out with the cookie notice blocker. Right. So that I no longer have to click accept blindly for all of my cookie EU connected cookies.

26:25Matt Coles Isn't that what, isn't that like NoScript under Firefox? Oh, I don't have to click. I don't have to click accept on cookies. Cause I never see it.

26:35Chris Romeo I haven't seen a cookie notice in years at this point.

26:38Matt Coles So I, uh, I enjoy reading those actually. I'm, I'm disturbed. Stop.

26:44Chris Romeo Wait, stop. Hold on. No, no, no, no. We're not going any further until we unpack this problem. Wait, you just said you enjoy reading the cookie notices. You're gonna have to explain in some more depth about what's enjoyable about that.

26:59Matt Coles You know, it's interesting.

27:00Izar Tarandach I don't, I'm not sure I really want an explanation. I prefer to think of math the way that I always think of math.

27:06Matt Coles I used to, I used to read the EULAs too, but I gave up after a while.

27:11Chris Romeo You can also just skim it. You can also just skim a 75-page EULA.

27:13Matt Coles I skim it, uh, just in case they inject something.

27:17Chris Romeo This is before, I mean, you've been doing this obviously longer than AI's been around. Oh yeah.

27:20Matt Coles Yeah, just, but I mean, just in case they inject something new, because all the, all the YOLOs are cookie cutter, right? And look, but if you will look for it, you know, sometimes they have Easter eggs. Um, but the, the prompt, so the, the cookie prompts, right? Because sometimes you'll get the cookie prompts like accept or reject. Sometimes you'll get the simply accept, you don't have a rejection choice, right?

27:41Chris Romeo Okay.

27:42Matt Coles And then sometimes you'll have these slider bars and some of them we on by default and you have to uncheck them. Sometimes they'll be off by default and you get to check them if you want. And then, and then other times they're confusing because there's performance cookies and then there's tracking cookies and there's targeting cookies and there's all these other types. And Kim, if you're listening right now, let's have a debate about privacy engineering.

28:08Izar Tarandach Wait, wait, wait. I lost you on— the only thing I can think about now is cookies. I haven't had breakfast. I haven't asked yet.

28:18Matt Coles Talk about least privilege. What's the privilege of the application to take my data?

28:23Chris Romeo You said they read the—

28:25Izar Tarandach Yeah, I'm stuck on that too. I'm stuck on that.

28:28Chris Romeo What does that mean?

28:29Matt Coles Well, they pop up a little—

28:30Chris Romeo What are you reading?

28:31Matt Coles It pops up a little box that says what cookies would you like to accept.

28:34Chris Romeo I see. Okay. You're not clicking through to some like 100-page document and reading that on the website you visit. Oh God, no. That was my concern.

28:42Matt Coles They do have those. Occasionally the sites do have that with more information about how they use the cookies, yeah, it's all cookie cutter.

28:48Chris Romeo I don't care. Because I was afraid like we were gonna have to do an intervention here or something. Like, Matt needs a hobby.

28:55Matt Coles When you learn about privacy, when you learn about privacy, when you, when you learn about privacy and you get, you get interested in these sort of things.

29:08Izar Tarandach How do you find that Matt in the supermarkets is the one in the Uh, snacks at, uh, aisle going through all the packaging.

29:15Chris Romeo Yeah. Reading all the labels. I'm reading cookies. I'm just looking at all the labels trying to figure out.

29:22Matt Coles I do read the labels.

29:25Chris Romeo See, I think we may be onto something here. I think we may have uncovered a pattern. Maybe Izar and I, we should be profilers for the FBI. 'Cause we seem like we can really find patterns.

29:34Izar Tarandach That's why he likes S-bombs so much.

29:37Matt Coles I hate S-bombs, but that's okay.

29:38Izar Tarandach He's the one who reads them.

29:39Matt Coles I do not read S-bombs.

29:41Chris Romeo He is the one who reads them cover to cover.

29:43Matt Coles I do not read S-bombs.

29:47Chris Romeo Come on, have you? Okay, let's— if you've, if you've ever read an S-bomb in your life, raise your hand. Come on, Matt.

29:56Izar Tarandach No, come on.

30:00Chris Romeo Very reluctantly. If you're on the audio-only stream, he starts very reluctantly raising his hand slowly with one finger.

30:07Matt Coles Did you accuse me? But you're actually—

30:09Chris Romeo it's you.

30:09Izar Tarandach I believe that you cannot criticize something without reading Okay, that's a safe bet.

30:16Matt Coles Wait, did you read an SBOM or did you read an SBOM spec? Yes, but if Steve Spriggett is listening right now, we love you.

30:28Izar Tarandach But, uh, the truth of the thing is that, um, as every security, uh, practitioner, I, I do have, uh, sleep problems and, uh, reading SBOMs is an awesome way to overcome them.

30:42Chris Romeo Some people count sheep. Some people count sheep.

30:44Izar Tarandach 3, 4 packages in and you are down. You're down.

30:49Chris Romeo Could there be something more boring?

30:52Matt Coles Well, there could be. I've read—

30:54Izar Tarandach than YTS Bones?

30:55Matt Coles I've read hardware bill of materials too.

30:58Chris Romeo Oh my God. YTS Bones. You're just like, this is, it just keeps getting worse. You're like, oh, I just read hardware bill of materials, which is like part number 47242. Are you looking at like Matt's like, oh my God. Wait, wait, wait, he used 242? Don't, don't mess with my part numbers scheme that I just invented in my head. All right, Matt, what other bizarre things do you read? This is the title of this episode is Matt's Bizarre Reading Habits. So we got cookie notices. Cookie notices. We have, uh, hardware bill of materials. We have EULAs, which nobody, even lawyers don't read EULAs. Matt reads Ulysses.

31:39Matt Coles I skim them, but you don't read SBOBs.

31:42Izar Tarandach I don't read SBOBs.

31:42Chris Romeo You don't read SBOBs.

31:43Matt Coles I don't read SBOBs.

31:44Izar Tarandach That's a name.

31:44Chris Romeo What other names are on that episode? Is there anything else bizarre that you read?

31:48Izar Tarandach That's a name for an episode. Matt reads Ulysses. And I'm sure that we could, uh, sell them to the apps that do the, the pre, the, the pre-sleep thing. That's like, imagine, imagine Matt going, hmm. And now by clicking here, you accept all the terms.

32:09Chris Romeo All the terms of my billing.

32:11Izar Tarandach And you slide into sleep. Take a deep breath and click.

32:16Chris Romeo Click.

32:16Matt Coles I just want to know how many things I'm going to violate by clicking the EULA.

32:22Izar Tarandach You already violated the terms of not reading the EULA.

32:25Chris Romeo Let's, let's, um, we got to unpack this some more. So. You're reading the EULA to know how you're breaking the EULA?

32:37Matt Coles Oh, come on. You always break the EULA. You always break the EULA. There's no way you, you—

32:41Chris Romeo Well, I don't know. I've never read it. I know I'm breaking it.

32:45Matt Coles You're not reading—

32:46Chris Romeo Ignorance is bliss.

32:48Izar Tarandach You're not breaking the law if you don't know the law. Everybody knows that. That's like—

32:52Chris Romeo That's my— I would argue that in court all day long.

32:54Matt Coles Is it willful ignorance or willful—

32:56Izar Tarandach No, it's the way the law is worded.

32:57Matt Coles Is it worse to read it and know what you're violating? On YouTube, go.

33:00Chris Romeo I think it's worse to read it and know what you're violating, 'cause then you know better. Okay, so wait, what, what am I, how am I violating the EULA?

33:12Matt Coles I don't know. What do you do?

33:13Izar Tarandach You use software? Nowadays people don't, right?

33:15Chris Romeo No, I don't really have software anymore.

33:17Matt Coles We can't, we can't, we can't violate our EULA.

33:19Chris Romeo No, I don't have software. I mean, I guess I have an operating system. That's probably it. Everything else is in the web.

33:25Matt Coles Uh, And that is software.

33:30Chris Romeo Hmm. It's something. Do you have a browser software? I do, but that's called software. Has a EULA because it isn't, didn't cost anything.

33:40Matt Coles It doesn't have a EULA because it didn't cost anything.

33:44Chris Romeo Let me check. Where would I find a EULA, Matt? Since you're the EULA expert, do I report a broken site or Brave help? Where is the EULA? Where is your EULA?

33:55Matt Coles I mean, uh, uh, Chrome has a, has a EULA.

33:58Chris Romeo No results found.

34:00Izar Tarandach For what?

34:01Chris Romeo I'm gonna, there's no license agreement for Claude for Brave? Apparently not. I'm gonna ask Claude to explain it to me.

34:08Matt Coles Cause terms of service also.

34:10Chris Romeo I wanna be a EULA expert like Matt. They have like 6 on their belt. Oh, you have to read the logs for that. Belt colors for reading EULAs? Like, are you a EULA reading black belt or?

34:23Matt Coles I can, no, I just skim it. So, uh, it's just, uh, I'm sure that there's a Guinness record for that.

34:29Izar Tarandach Like who got the longest deep into an EULA.

34:34Matt Coles You know, there are, there are sites for you. There are sites that analyze EULAs and other agreements for reading comprehension level.

34:43Izar Tarandach Somehow I'm sure there are, and somehow that does not surprise me.

34:48Matt Coles There's a, well, Mozilla had Batteries Not Included, which looked at privacy policies for, uh, reading comprehension level, which is where they get me because most of it I don't understand.

35:02Izar Tarandach Oh wow.

35:02Chris Romeo These EULAs are actually pretty intriguing. Apparently Brave can terminate my access to use the browser at any time if I fail to comply with the terms. Never knew that.

35:13Matt Coles I don't know how they would do that.

35:15Izar Tarandach Well, you know that they are so intriguing that they had like classical books based on eulogies, like the, the Eulogy for the Red October.

35:24Chris Romeo The Eulogy for the Red October. This is riveting.

35:28Matt Coles The Eulogy for the Red October.

35:30Chris Romeo This is riveting television. It was last updated May 11th, 2023.

35:35Izar Tarandach Oh, oh, you're looking for the Eulogy in Paris.

35:38Matt Coles Yeah.

35:38Chris Romeo That's it. It's like, yeah, it's like, it was a dark and stormy night. That's how it starts.

35:43Izar Tarandach Wow. Cool. Oh, A Tale of Two Eulas. That's a classic. Yeah.

35:49Chris Romeo It was the best of times. It was the worst of times.

35:52Izar Tarandach No, it was the best of Eulas. It was the worst of Eulas.

35:57Chris Romeo I'm glad I give you guys— Apply Eula to classic literature.

36:00Matt Coles I'm glad I give you guys so much to think about.

36:05Chris Romeo I was hoping for a Joe Pesci impression.

36:09Izar Tarandach Do I amuse you? Do I amuse you? Am I funny? Jack Nicholson, you want the eula? You can't take the eula. You can't handle the eula.

36:21Chris Romeo That's what Matt says at meetings.

36:25Izar Tarandach You can't handle the eula.

36:28Chris Romeo We have really diverged here.

36:32Izar Tarandach Yep.

36:32Matt Coles We were started talking about AI. We went to least privilege and here we are on eulas.

36:38Izar Tarandach Well, in the big context of things, it's all one big thing. We haven't exploded the 1 million tokens yet.

36:46Chris Romeo This is— I told, I asked AI to give me 5 movie quotes and substitute EULA into the quote. Says The Godfather, I'm going to make him a EULA he can't refuse.

37:04Izar Tarandach Jerry Maguire. No, no, that's actually a cookie, a cookie dialogue. I'll make him a cookie dialogue he can't refuse.

37:10Chris Romeo Jerry Maguire, you had me at, I have read and agreed to the EULA. And then they actually, the AI pulled Isar's real one, A Few Good Men, you can't handle the EULA. It's number 3 on the list.

37:22Izar Tarandach AI's had to restrain my humor.

37:25Chris Romeo Number 4, Gladiator, are you not bound by the EULA? And then number 5, Apollo 13, Houston, we have a EULA.

37:35Izar Tarandach Yeah, but it skipped use the EULA, Luke. Luke, I am your EULA.

37:46Chris Romeo So wait, that means that Darth Vader was kind of a robot looking dude. Like he might've actually been software and hardware.

37:52Izar Tarandach Wait, wait, wait, wait. I haven't even started. Prepare for ludicrous Yule. Wait, what?

37:58Chris Romeo Ludicrous like the rapper?

38:00Matt Coles No, that's like Spaceballs.

38:02Chris Romeo Yeah.

38:03Matt Coles Ludicrous speed.

38:04Izar Tarandach Have fun reading the Yule, boys.

38:11Matt Coles Princess Bride.

38:11Izar Tarandach Princess Bride, come on.

38:12Matt Coles Princess Bride.

38:13Chris Romeo I don't remember Yule as being in Princess Bride.

38:16Matt Coles Have fun storming the castle.

38:18Izar Tarandach Have fun storming the castle. Have fun reading the Yule.

38:21Chris Romeo Finding the Ullr, that's what we tell Matt now.

38:23Izar Tarandach We're the knights who read, we, we're the knights who read the Ullr. Yeah.

38:28Chris Romeo And the knights who say knee.

38:30Izar Tarandach Knee.

38:30Chris Romeo All right. I think we've pretty much taken up enough time from the people watching.

38:38Izar Tarandach Yeah. We promised to have a subject next week.

38:40Chris Romeo Yeah. Well, there was a subject. We had a subject.

38:42Matt Coles We talked about least privilege and privacy.

38:44Chris Romeo Least privilege and Ullr.

38:45Matt Coles Yeah. We did that.

38:47Izar Tarandach Check. Check on security.

38:48Chris Romeo We had, uh, yeah, we, the, the EULA episode can now go into the archives.

38:53Izar Tarandach We indeed no longer will have to tackle this topic if people click on the right place.

38:57Chris Romeo Yeah. Only if you click on accept.

38:59Izar Tarandach Yeah.

39:00Chris Romeo All right. Well, don't violate the terms. Don't violate the terms. Um, I'm trying to think of some other wisdom. I just, I'm at a loss.

39:09Matt Coles Least privilege. So always apply least privilege.

39:11Chris Romeo I've always applied least privilege to everything that you do, even when buying a car. You can figure out how to do that.

39:17Izar Tarandach No, we, we need, we need a, a last closing Eula quote.

39:22Chris Romeo Uh, okay, well we're just, you know, we'll put up a little, a little window on the YouTube video that says, you know, 5 minutes later, dot, dot, dot.

39:30Matt Coles 5 minutes? You know how long it takes to read a Eula?

39:35Chris Romeo No, we're like, he needs another quote. He's thinking.

39:37Izar Tarandach I'm going somewhere of like Back to the Future. Great eula? No, too short. Oh, where are we going? We don't need eulas.

39:48Chris Romeo Yes, that's it. That's the end of this episode of the Security Table. Remember folks, read your eulas.

Transcript supplied by AssemblyAI. It may contain transcription errors.

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