40 minSeason 4, episode 1

Crystal Penguins and AI Chaos: What Could Go Wrong in 2026?

With Chris Romeo, Izar Tarandach, and Matt Coles

AI security & autonomous agents

We’re predicting what 2026 has in store for AI and cybersecurity. We explore the wild possibilities of AI integration gone wrong, from people accidentally connecting their AI to sensitive file systems to blaming their AI agents for losing critical data.

Episode title image: Crystal Penguins and AI Chaos: What Could Go Wrong in 2026?

Listen to the conversation

About this episode

We’re predicting what 2026 has in store for AI and cybersecurity. We explore the wild possibilities of AI integration gone wrong, from people accidentally connecting their AI to sensitive file systems to blaming their AI agents for losing critical data. The conversation takes a thoughtful turn as they debate which jobs might fall to AI automation and if the human touch is still irreplaceable? Examining real examples like the "Y'allbot" weather monitoring system and photorealistic AI actress Tilly Norwood to illustrate how rapidly AI is transforming industries.Tune in and learn how to navigate the AI-powered future responsibly.

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

0:09Chris Romeo Hey folks, welcome to the first episode of 2026. Well, no, I guess not the first episode of 2026. First episode of the Security Table in 2026. Probably were other episodes on January 1st of something. I'm Chris Romeo, joined as always by Matt Coles and Izar Terrindasch. And welcome to the 2026 edition where we will make vast predictions, which if you were to take these predictions and create wagers on whatever that site is, you can wager on anything, you will most likely lose all of your money.

0:48Izar Tarandach Speaking of losing, I don't know where I put my crystal ball.

0:51Chris Romeo Oh, it's in your, out your nostril, your isodromus outfit.

0:56Izar Tarandach Yeah. For Deco. No, no, that, that's for last year. This year I'm going to use my crystal penguin.

1:02Chris Romeo Crystal penguin. Okay. All right. Folks, this, it looks like a real penguin if you're listening on audio, like a stuffed penguin, but apparently it's has crystal somehow embedded within it.

1:14Izar Tarandach It's skeeper.

1:20Chris Romeo Skipper. Well, you gotta go first if you got Skipper, then Magic Red One. Oh, right, right, right.

1:25Izar Tarandach 2026, 2026. Let's try to channel 2026.

1:30Matt Coles Channeling 2026.

1:31Izar Tarandach I predict—

1:33Chris Romeo Oh, wait, that's insane.

1:35Izar Tarandach Yeah.

1:35Chris Romeo Insane.

1:37Izar Tarandach I predict that somebody big is going to connect things they shouldn't connect to their AI. And then they are going to be very, very surprised when bad things happen, but like very, very surprised.

1:54Matt Coles Uh, why would that surprise anyone?

1:57Izar Tarandach I don't know. I just wanted to start with one that's like 100% of actually happening.

2:02Matt Coles Well, that happens a lot apparently. Uh, people connect their AI to their local file system and bad things happen already.

2:12Izar Tarandach What could possibly go wrong?

2:14Chris Romeo You gotta define what bad things means, 'cause, uh, for some, one person's functionality is another person's bad thing.

2:21Izar Tarandach Sure.

2:21Matt Coles That's true. As soon as the AI issues rm -rf, boop, that'd be hilarious.

2:27Izar Tarandach Oh, in that case, in that case, in that case, I predict that somebody's going to have an accident with their file system. And when somebody else asks them, why did you tell the agent that it could access the file system? The person is going to say, well, it asks so nicely.

2:40Matt Coles I'm pretty sure you're supposed to ask it, not it asking you.

2:47Izar Tarandach No, no, no, no, no, no, no, no. It asks you in the beginning, hey, do you trust me to access the files in this directory?

2:53Matt Coles Oh, and you say, of course. If Emily asks if you're a god, you say yes.

3:00Izar Tarandach Yeah, of course. Like it's an AI. It knows what it's doing. It's been trained.

3:05Matt Coles It's been trained. So there's an almost guaranteed prediction on somebody will lose their files, blame it on the dog again, the future of AI peril in 2026.

3:17Chris Romeo So that will likely be a real thing. All right, man.

3:21Matt Coles It'd be funnier. It'd be funnier if they, if they lost it to the AI dog.

3:26Chris Romeo True.

3:27Izar Tarandach We should train a model, call it Fido. And teach it to do people's homework so that kids can say Fido ate my homework.

3:34Matt Coles That's right.

3:35Izar Tarandach Yeah.

3:36Matt Coles There you go.

3:37Chris Romeo Well, you don't even need it to do the homework. You just need it to be a shredder of the homework assignment. Right.

3:42Izar Tarandach Yep.

3:42Matt Coles Well, it needs to, it needs to be plausible deniability, right? So it has to have done something and then deleted it and then, eh.

3:48Izar Tarandach Mm-hmm.

3:51Chris Romeo Why wouldn't you just have the dog turn in the homework? And that involves—

3:55Matt Coles Well, 'cause then the dog—

3:56Izar Tarandach But we do know some, some dogs that might as well. Okay. What else? What else? So we did that.

4:01Matt Coles Oh, fetching the paper. If they fetch the paper, that doesn't exist anymore.

4:06Izar Tarandach I have an agent for that.

4:08Matt Coles Digital paper.

4:09Chris Romeo You'd have to be 150 to be receiving a— Hey, wait, hold on. Do you get, is there a paper delivered to your driveway each morning?

4:20Izar Tarandach No, but I do have a neighbor that every morning I pass by and I see a paper thrown in there.

4:25Chris Romeo So, okay.

4:26Izar Tarandach The industry still exists.

4:28Chris Romeo It's, it's hanging on by a thread. Yeah. I understand.

4:32Matt Coles And then, and then someday, some kid is still, it'll be drone delivery.

4:38Izar Tarandach Some kid is still planning to go to college by throwing papers and stuff.

4:42Chris Romeo I don't think they have any more paper boys. I don't think that's a job anymore. Paper people. It's all like, yeah, it's like adults do it. Like since the 1990s, adults do it driving cars.

4:53Matt Coles They drive around in their car and they, yeah.

4:55Chris Romeo Get really good at throwing papers out the window and landing them perfectly.

4:59Matt Coles So across, across the car, out that window.

5:02Chris Romeo Yeah.

5:03Izar Tarandach I could see myself sitting on the porch and flying a drone that delivers papers down to people.

5:09Matt Coles Yeah.

5:09Chris Romeo But that would be for fun. That would just be for your own enjoyment.

5:13Matt Coles So now really fun would be get one of those t-shirt cannons and use that to throw paper.

5:19Izar Tarandach Oh, put it on the back of a Jeep. Put it on the back of a Jeep and just drive through.

5:23Matt Coles You get a paper.

5:23Chris Romeo Everybody gets a paper. 5 papers. Okay, man, what do you got?

5:30Izar Tarandach Back from the 60s.

5:32Matt Coles Oh, wait, I just made the— I made the first one. What do you got?

5:35Chris Romeo That wasn't a— something about a newspaper, it's all I remember.

5:38Matt Coles Oh no, the, the, the fire— never mind. Uh, oh, let's see, it's so hard because they've all been done already. Uh, So we know AI is going everywhere and everywhere is becoming AI.

5:59Izar Tarandach Oh God.

5:59Matt Coles Somebody will relinquish their job to an AI. The AI will do it better, surprisingly, and it'll get a raise and become CEO.

6:17Izar Tarandach Can we be honest for a second and think to ourselves, January 2026, What is one job that you can imagine an AI doing better than the person who's doing it? With the AI in the level that it is today. That's completely fire and forget. You write a couple of agents, they talk between themselves, and you go away and the job gets done. Hotel concierge. What's one job that can do that?

6:46Matt Coles Hotel concierge.

6:48Izar Tarandach Hotel concierge.

6:50Chris Romeo Yeah.

6:52Izar Tarandach Until somebody comes in and asks for the Saturday night special and whatnot.

6:58Matt Coles Or, or, or, or even the person at the front desk, right? So if you're checking in, here, can I take your, your payment information? Here's a room for you. Here's your key. Yes, you have to have a physical. You have to have physical. So maybe there's a robot involved.

7:11Izar Tarandach That, that's if everything goes right. What about the exceptions that hotel concierges deal with day in, day out?

7:18Matt Coles Isn't that what the agents are for? You spin off a new agent that can handle the exception? You teach it a skill?

7:25Izar Tarandach No, but then you have to stop the thing and teach the skill, and you have to fail on the first time that the skill is asked because it doesn't know the skill.

7:32Matt Coles You just go and watch, uh, you watch 20 years, you watch 20 or 30 or 40 years of, of customer service engagements to learn all the exceptions.

7:42Izar Tarandach Put aside for a second the what-ifs. What is the one job today that an AI can do perfectly with AI in the state that it is?

7:51Matt Coles News reporter?

7:53Chris Romeo Weather, weather person? Because they're wrong 90% of the time, so.

7:58Matt Coles Well, actually, so, uh, has anyone seen the Yawlbots?

8:02Chris Romeo Mm-hmm. The who?

8:04Matt Coles An AI, an AI by, uh, Ryan, Ryan Hall, who runs a YouTube channel. called, um, I think it's what, Reinhold Jahl, uh, and he has, um, he has a Jahl bot, which is basically an AI-trained model looking at weather and, uh, actually doing this 24/7. So you can find that online. AI is already doing the weather. So weather person or news anchor probably would be able to do it too. I mean, that's, uh, that's for me to read a story and speak a story.

8:41Chris Romeo On the weather side though, those folks are meteorologists, so they do have some scientific knowledge and, and experience and study of weather patterns, weather systems.

8:54Matt Coles So yeah, but why couldn't you have the meteorologist still feeding the, the weather person? So in other words, weather person is still on air.

8:59Chris Romeo Okay, so you're just talking about the on-air, so pretty, pretty much any on-air personality that was being fed data from some other source could be an AI at this stage, right?

9:10Izar Tarandach So what you're telling me is that we are basically at max headroom level, MTV circa 1980-something.

9:17Matt Coles Well, so my understanding is there is also an AI actress currently, right? There's, uh, Ms. Newton, I think her last name is. I don't know her first name is on Presence, but she's an AI.

9:31Chris Romeo You, you're, you're, you just used pronouns describing this actress as an actual person.

9:37Izar Tarandach At least he didn't say that the name is AI Livia Newton-Jones.

9:42Matt Coles So, Tilly Norwood. Sorry, Tilly Norwood.

9:45Izar Tarandach What?

9:46Matt Coles Tilly Norwood.

9:47Izar Tarandach So, okay.

9:49Chris Romeo It is Tilly Norwood.

9:51Matt Coles Photorealistic AI-generated character.

9:55Izar Tarandach Wasn't Norwood the name of some CIA campaign?

9:59Chris Romeo So Tilly Norwood, the AI, has appeared in a television show.

10:05Izar Tarandach Yes.

10:05Chris Romeo Which, do we know which one? Does it say? I guess we can check IMDb.

10:10Izar Tarandach It's, uh, but wait, as an actress, can it give an informed political opinion?

10:18Matt Coles I, I don't know.

10:21Izar Tarandach Because if not, it's not doing the job perfectly. But, uh, Yeah.

10:29Matt Coles So, and, and wait, and Tilly is part of an AI talent studio. So they have more of these coming.

10:38Izar Tarandach So basically what you guys are telling me is that the level of AI performing jobs is following a perfect script, either for news or for acting or for weather.

10:54Chris Romeo Reporting. It doesn't seem like we're there yet where you have to worry that AI is going to take my job. We're not quite there.

11:00Izar Tarandach So can we make a prediction that people are going to still be saying that and no, it's not going to happen in 2026?

11:07Matt Coles Or are things going to be interesting? We heard the, we heard, I think it's the story of Salesforce where they eliminated a lot of staff for AI and then realized, wait. We need people.

11:20Chris Romeo Yeah, that may have just been, that may have been a ploy to, to clear the deck. You know how big corporate entities work though. They're sometimes they, well, we, we, we're, we're doing a new AI policy. Okay, let's slash 10% of the headcount and then nobody will think it's actually a layoff. They'll think it's a, it's an AI move.

11:38Izar Tarandach There was a smaller company, a payment company, I don't remember the name, that they actually did do that. Like they, they sent everybody home and then half a year later they said, well, actually we need you guys. It wasn't Salesforce. It was, uh, something with A. I don't remember.

12:00Chris Romeo Hmm.

12:01Izar Tarandach Okay.

12:01Matt Coles So Salesforce laid off, Salesforce laid off 4,000 employees and automated them with AI agents and then started rehiring.

12:07Chris Romeo Yeah.

12:11Matt Coles Customer support. So they use them for customer support. Reading the article right now.

12:14Izar Tarandach Customer support. That's another one. Customer support. So I, I see a lot of, uh, AI going into customer support.

12:22Chris Romeo There's a lot of, I mean, pretty much every customer support experience is AI driven with humans at like level 4 or whatever.

12:30Matt Coles Yeah. I mean, that's, that's like FAQ reading, script reading, and then any exception gets bubbled to human.

12:36Izar Tarandach What was that, uh, That movie that, uh, I don't remember. I think that it was Ghostbusters, that there is this demon that feeds on people's rage. Do you guys remember that movie?

12:47Matt Coles That's Ghostbusters 2.

12:48Izar Tarandach That's Ghostbusters, right?

12:49Matt Coles Ghostbusters 2 specifically.

12:52Izar Tarandach So I think that the whole AI and customer support thing is a big plot to harness all that energy. Because I don't know one single person that can go through a loop of customer support with an AI that's not enraged. Like in the middle of the, the experience.

13:09Chris Romeo I mean, I'm gonna be honest, I haven't used customer support of any companies in about a year. I use AI as the customer, my customer support to solve whatever problem I have. So true.

13:24Izar Tarandach I, this weekend actually I had a, an experience with that where I asked, uh, uh, 2 AIs actually for advice in helping me solve something extremely frustrating. Turns out that it's something extremely frustrating for a lot of people and there isn't a solution, but the AI was extremely well convinced. Both AIs were extremely well convinced that yes, there is a solution and here's step 1, 2, 3, 4. And by step 4, the file that they're telling you to look at doesn't exist. And by step 5, you're contacting aliens to do something.

14:04Matt Coles But, uh, so as a prediction, as a prediction, all customer support problems will be solved in some way by some AI.

14:15Izar Tarandach Well, given that there's a bunch of like companies that are dominating the chatbot market for, uh, uh, customer support. I'm thinking that if they don't get solved, at least there should be a bot that you can ask. So yeah, no, I don't think that it's going to get solved, especially because, and this I've been seeing myself, a lot of the interactions that are happening with these chatbots, they involve credit card information and health information and this and that and the other one. And then the problems that we've been seeing, they just keep appearing again and again and again and again. So I'm going to get on the soapbox and remind people, just because you can't doesn't mean that your AI should. Do us a favor. Don't let your AI do much shit.

15:05Chris Romeo Let me throw my layup AI, since you guys did two AI predictions, I'll throw a layup AI prediction guaranteeing that mine will be correct just like he's ours. AI will get 100% better in 2026. Than it was in 2025. Finger roll, layup.

15:26Izar Tarandach Yes, but yes, but apparently we won't have the memory to run it.

15:33Matt Coles Yes, that's right.

15:36Chris Romeo Memory shortage.

15:37Matt Coles That's the other, that's the other. DRAM will become, DRAM will become the next Bitcoin.

15:42Izar Tarandach Yeah, exactly. It is.

15:44Chris Romeo I know.

15:45Izar Tarandach So well put.

15:46Chris Romeo I know, I know a guy who is in the kind of resale business. He's in the big time kind of resale business. So he'll go and like buy pallets of stuff, like technology stuff. And he was saying, he was telling me that the RAM, the DRAM market, the old stuff from like 5 years ago is like quadrupled in price in the last year. And so people are buying—

16:12Matt Coles Yeah.

16:13Chris Romeo Like it's a semi-truck business filled with Technology, you know, he'll buy, he'll buy trucks filled with technology stuff.

16:20Izar Tarandach Mm-hmm.

16:21Chris Romeo And then sell it on eBay.

16:22Izar Tarandach Can I lift the truck and see what falls?

16:27Matt Coles Well, so those are all great warmup predictions. Now, because this is called the Security Table and not the AI Table, oh, should we make any security predictions?

16:35Izar Tarandach Wait, nowadays isn't it all the same thing?

16:39Chris Romeo Yeah, I was, now I, I was gonna do the same thing, Matt, ask for, now can we have a round of security Or privacy, because we love both.

16:47Matt Coles Yep.

16:47Chris Romeo Focus. Or safety.

16:49Matt Coles We could do safety too, but, but certainly security, privacy.

16:51Izar Tarandach Yep.

16:52Chris Romeo Okay. I mean, people will wear orange vests. There's a safety prediction. Not really. That was—

17:00Izar Tarandach there will be a huge privacy impact that will need people to talk about loss and nothing is going to happen.

17:10Matt Coles Of course there will be. That's not really a prediction.

17:13Chris Romeo That's probably already happened this year and it's only like 12 days long.

17:17Izar Tarandach It happens every year.

17:21Matt Coles That's a, that's a given. That's a law of computing. That's not—

17:25Chris Romeo all right. I'll give you one. I'll give you one to chew on here that I've been thinking about quite a bit that I've, people have asked me this question as well. So it's a little bit cheating, but the, I'm going to mix AI back into AppSec now. So. The, I, I mean, the divide between the old school AppSec companies and the new school AI market-based AppSec companies is gonna, I, I don't, I actually think it's gonna close. The gap's gonna start to close in 2026. I think the bigger companies are starting to catch up now, and I think we're gonna see the checkmarks and those, those scale of companies really embrace their AI and start to lead with it. I'm not going to say anything about how the fact that it's going to provide better results or anything like that.

18:16Izar Tarandach I'm not going there, but I'm going to go the other way. I think that customers are going to start saying, well, presenting AI fatigue and saying, all this AI that you put into my stuff, it's just making me more confused. Please give me back my menu with my list in the middle.

18:38Chris Romeo I wish that was the case. I mean, I was in, I was in talking to a company that was, uh, preparing to do a sale and they got stuck on the, the large company that was buying a, the, a product was very much caught up on their AI strategy and was almost, almost using that as the definitive purchase choice or, or what was driving their decision on what they were gonna buy was based on the AI strategy, not on the core technology that this company provided.

19:16Matt Coles Hmm. So I think the AI, the AI strategy can be a breaker too, right? I mean, you have, if you have larger organizations that have privacy or compute policies, right? The AI strategy can, can make or break that deal. In terms of whether it's local processing, you know, on-prem versus pushing all your data to the cloud for the AI to operate on it.

19:40Izar Tarandach Okay. So here's a prediction. Some company is going to put their whole strategic decision and market pipeline and all that good stuff in the hands of very, very good, very well-defined, very well-performing AI processes. And then one day the AI provider, the third-party provider, is going to go away, and this company is going to be looking around and going, what do I do now?

20:06Matt Coles Because you're talking about the whole cloud, uh, you're talking— we're talking this, this was the prediction from the cloud services. The cloud, the cloud is going to take all your data and then poof, it's going to go away and your data is now missing, right?

20:17Izar Tarandach But that's the thing. How do we solve that? How did we solve that thing with the, the whole, uh, cloud agility thing? Which was the same thing as crypto agility, right? You, you, you design your systems so that you can pivot and, and put the thing that you need somewhere else or make the thing that you need happen in a different way, but without changing the higher levels. But on AI, that, that, that gets different because assuming that, uh, you are relying on specific models that are run by a specific vendor. When you pivot away from that, the performance of the stuff that you so well engineered to work with that model may completely change to work with a different one. The constraints are different, the training is different, the approach is different, the results are going to be different.

21:05Chris Romeo Or you just tell the new model to act like the model you transferred from.

21:10Matt Coles But also data portability is a problem, right? I mean, it's, it's, it's whether or not the data is being acted upon by the model or the data is being integrated into the model. It's not really portable if it's been integrated.

21:21Izar Tarandach But that, that, that's the thing. When, when it's, when it's, uh, when it's integrated, that, then you have the whole discussion of who, who owns the model, and that's a contractual thing.

21:30Chris Romeo So you have your own, uh, copy and enterprise customers are, have very specific pro, uh, training data policies now that are being added to, to contracts and stuff that say, thou shalt not use my data to train your AI systems. But in the cases where it is, $100 billion.

21:51Izar Tarandach In the cases where it is necessary, then you get an instance of the model, not the model that gets served to everybody else. Yeah.

21:56Chris Romeo So that, so they can be terminated. So that, yeah.

21:59Izar Tarandach So the, and then you have joint ownership and in theory you could bring the model in and run it inside if you had the appropriate hardware. Right. So in terms of mobility, that's even a good scenario to have. But most people are not like that. They want the generic AI thing. They want the thing to answer. They want the thing to give pretty things and Lord knows what happens. And in that case, you, you, you, you, you focus on that model. Some, what I've seen in some places is that people just get like, it's almost like, uh, uh, people get infatuated with models. Oh, look at this one. It does exactly this thing that I needed to do and it does it so well when it's the only one in the world. Right. I mean, you just need to go to Reddit and see that the same wars that we had via Emacs, which are still open, are now being had between different models, right? People are fanboying the models to death. And sometimes—

22:58Matt Coles My Gemini beats your ChatGPT any day, right?

23:01Izar Tarandach Exactly. But now it's more like what I see at least, it's the local running people. My twin 2B30B FP4 is much better than your FP8. My FP8 is bigger than your FP4, whatever.

23:18Matt Coles So they're compensating for something and it's parameter size.

23:22Izar Tarandach Exactly. As we always do, right?

23:24Chris Romeo So it's, it's no longer sports cars that people are using to compensate. It's, it's their, it's their model.

23:32Matt Coles It's their model size. Yeah.

23:34Chris Romeo It's a weird world.

23:34Matt Coles Model size doesn't matter. Does model size matter?

23:36Chris Romeo It's a, it's a weird world we live in. Weird world.

23:40Izar Tarandach It's not the model size, it's what you use it for.

23:45Chris Romeo Well, great. There goes our PG-13 rating.

23:47Matt Coles Gone.

23:50Chris Romeo Sorry, parents.

23:53Matt Coles All of you.

23:54Izar Tarandach I forgot where I was going with that one.

23:56Chris Romeo I thought you were, are you done or did you have more?

23:58Matt Coles Did you have more? I, I, I was gonna say a, a, a particular security one, which was the attack vector moves into the, the, the cert, the access side. So behind the MCP server and all those people who are vibe coding those access code, those access, uh, access codes, scripts and codes, right?

24:22Chris Romeo War games. I was thinking war games too.

24:25Izar Tarandach Immediately, immediately went there.

24:27Matt Coles They become, they become the new attack vector and somebody will lose big time. I'm certain. Right. So the AI will be perfectly fine. The MCP server will be perfectly fine. It's all the thing behind the MCP server that does all of the local access becomes compromised because somebody vibecoded it and left a debug port open or something without authentication.

24:49Izar Tarandach Hmm. Let's talk AI identity now. How many companies are going to have people running around asking, who's Jeff AI? Who's Jeff AI?

25:05Matt Coles Well, here's an interesting question. Does AI, why are we on, why are we stuck on AI? First off, this is a security show.

25:12Chris Romeo Because it's the hot thing.

25:13Izar Tarandach Because it's 2026.

25:15Chris Romeo Everybody's stuck on AI.

25:17Izar Tarandach Yeah, let's talk Bast.

25:19Chris Romeo No, that was my, that was gonna be my, my crazy prediction.

25:23Matt Coles My crazy— DAS will come back and say, "I will make it work." Don't, that was gonna be my prediction.

25:28Chris Romeo That's what I was gonna say next. Don't steal my jokes. Come on. Is you gotta, you, is your, is my writer writing for you too? What's going on here?

25:36Matt Coles Predictions. It's not a joke. It's predictions.

25:37Izar Tarandach Okay. Okay.

25:38Chris Romeo So no, it wasn't predict— that was, I was seriously gonna, gonna make that my prediction that, that an AI-powered DAS will bring DAST back from the ashes like a phoenix where it will rise and revolutionize enterprise security.

25:53Izar Tarandach Here's one that if you know, you know, I dread that thought.

25:59Chris Romeo That's it. That's the prediction.

26:03Izar Tarandach That's the prediction. I dread that thought.

26:05Chris Romeo I get it. We, Matt gets it too. I mean, we, we're nerdy enough to, yeah, nerdy and localized enough. Yeah, that's true. It's all context. Context is king, they say.

26:18Izar Tarandach Oh, oh, oh, that's a prediction. That's a prediction. There will be at least 4 new meanings to the word context.

26:24Matt Coles Yes, sir.

26:25Izar Tarandach Uh, okay.

26:29Chris Romeo So what'll be the ST, like short abbreviation that'll be added this year that will somehow relate to AI? Cause like we got I asked and Dast and Sast.

26:45Matt Coles It would be just AI.

26:46Chris Romeo What would we get? It would be A asked.

26:48Matt Coles A asked.

26:49Chris Romeo I asked.

26:49Izar Tarandach Madi. Madi.

26:52Chris Romeo Madi.

26:54Izar Tarandach My AI did it. Okay. Like it. Oh, if I put an H in there, then we have the Madi.

27:05Matt Coles Oh wait, we already, so. Prediction: somebody will redefine maid. Ooh. Ooh. Which already has AI in it.

27:17Izar Tarandach Is that going to give a new, a new meaning to evil maid attack? Yes.

27:23Matt Coles The evil maid attack.

27:25Chris Romeo Yes.

27:27Izar Tarandach Maid. Metal zone AI device. Uh, Midsummer Night's Dream.

27:39Chris Romeo Hmm. Shakespeare.

27:40Izar Tarandach Oh, that's good for hallucinations. That's good for hallucinations.

27:43Chris Romeo Yeah. All right. What about, uh, threat modeling predictions? I mean, if we're going to go full circle, we went from 30,000 feet to 5,000 feet. Let's go to the ground level of where the magic's happening. Everything, everything will stay the same. There's my prediction.

28:01Matt Coles AI will continue to be used as a thought generator, as a sounding board for money, or a thought provoker.

28:14Izar Tarandach Wait, wait, wait, the two extremes of the thing here, thought generator and a sounding board, two different sides of the coin, very different. In one of them, the AI is the active party. On the other one, the person is the active party. So I go to the person active party as the right one.

28:32Chris Romeo Mm-hmm.

28:33Matt Coles Well, so here's an interesting prediction. Will threat modeling resist AI-ification?

28:42Izar Tarandach No, no, no, no, no, no.

28:44Chris Romeo To its own detriment.

28:46Izar Tarandach But, but to its own detriment, that's true. But I think that approaches like, uh, the Vici's AI infused are going to become more popular because they are more effective than full, here's my diagram, give me my threat model approaches. Simply because people are going to notice that, yes, they made, that makes more sense.

29:11Matt Coles Right?

29:11Chris Romeo Well, it's also— Part of the beauty of threat modeling is the knowledge that you gain in that process and that, and what you just described there. And if I just have a magic box create the threat model for me, I don't learn anything. It doesn't improve my future designs.

29:25Izar Tarandach Exactly.

29:26Matt Coles Well, so that's an interesting, it's interesting you say that because, so there was a presentation from OWASP recently. Somebody did a presentation on, on the AI webinar and they said it something interesting that threat modeling is becoming audit ready or part of audit and compliance.

29:45Izar Tarandach AI, uh, sorry, threat modeling has always been audit ready. What I mean is, since the days when somebody would come up with a spreadsheet and say, this is my threat model. And somebody on the audit side would make a V, it has always been audit, right?

29:59Matt Coles But the, the implication is, I mean, if you run a, if you run threat modeling as a, as just a tool, here, here's some, here's a diagram, here's a diagram or a spec, give it to AI, AI spits out a bunch of, bunch of threats. It's a scanner.

30:13Chris Romeo Yeah.

30:13Matt Coles Not a conversation starter.

30:14Izar Tarandach Mm-hmm.

30:15Chris Romeo Mm-hmm.

30:16Matt Coles Does AI, does threat modeling in the age of AI become less of a conversation and more of a scanner. Do we need a new, do we need a new SAST ver— do we need a new term for SAST for threat modeling?

30:29Izar Tarandach That's not a problem with threat modeling. That's not a problem with AI. That's a problem with people.

30:35Matt Coles And how they perceive the activity because they see activity threat modeling as why isn't it like static code analysis?

30:42Izar Tarandach Those are the smart people. The rest of the people are just looking for the silver button that they press. They get the, the InstaThreat model, and everybody's happy. The auditor is happy, they are happy, the AI that did the threat model is happy, everybody's happy. But that's exactly my point. I think that just to not bury the lead, I think it's said, my whole point with AI security, especially AppSec security for AI, is that There's nothing new under the sun, right? People are just building this whole AI security, sorry, security for AI industry. And I keep looking at them and saying, why? You're just repeating the same things with new shiny words.

31:30Matt Coles Because it, because everything has to be reinvented, Maestro.

31:33Izar Tarandach Exactly, right? Because otherwise we don't move ourselves up in the LinkedIn scoreboard. But the, the. My point is that, and okay, yeah, I come from a very specific place to say this, but I say this with 100% certainty. The one thing that makes the most sense in security for AI, where AI is today, is threat modeling. Because that's the one process that can look at all the brittle stuff that AI brings to the table. And put some certainty in there.

32:11Matt Coles But you can't have an AI do that. You can't have an AI do that because you need access to, you need to, you need the internals and you need the controls that are outside the AI that the AI won't know about.

32:24Izar Tarandach Not only that, but we haven't yet found a way to codify business processes in a way that expresses exactly, and we see that when we talk privacy all the time, right? You have a business process. It has stuff that boils out of it. Edge cases, requirements, requests, exceptions.

32:49Matt Coles Black holes. And we haven't—

32:50Izar Tarandach Exactly. And we haven't found yet a way to codify those apart from writing the system that we are trying to threat model, which is the only single representation of a system that 100% means this is the system, right?

33:05Matt Coles Actually, I'm gonna push back just a little bit. We know how to write business processes. There's two, two ways, right? Flowcharts and business process modeling notation.

33:14Izar Tarandach But they still have a lot of ambiguity. My point is we don't have one single way that leaves all the ambiguity out so that you can give it to a model and trust 100% what comes back from it.

33:25Chris Romeo Uh, not a request.

33:26Izar Tarandach It's a request. If there is somebody smart enough out there, to come out with a way of writing out business requirements and processes in a way that's not ambiguous over time, I would love to see that.

33:39Matt Coles I think that can be done. I think you can write a business process that is unambiguous.

33:47Izar Tarandach I think that you have to take care to do it. It hasn't been done so far.

33:50Matt Coles You have to take care to do it though, right?

33:51Izar Tarandach Yes.

33:52Matt Coles Because you have to account for all the boundary conditions. Or have a failsafe, like a go-to, right? I, this, this is an undefined condition, and now I can address that undefined condition.

34:04Izar Tarandach And you, what you described right now is you just wrote code for a business process. So assuming that that code is 100% representative of the business process that you're trying to, to explain, you just wrote the one representation of that business process that is 100% parsable and understandable. And provable.

34:26Matt Coles So here's a question. Can we do business process threat modeling without an AI today?

34:35Izar Tarandach Yes, but we will always be left asking ourselves, what did we forget?

34:40Chris Romeo Sure.

34:40Izar Tarandach Because the ambiguity is always there, but that's going to happen with an AI as well.

34:44Matt Coles And, and then we document those, we as humans document those as assumptions. We assume that this functions as we expect, or this, we assume that there's a condition here that we don't know about that will probably happen, right? Can AI make assumptions? Sure. It makes assumptions, but can it generate them?

35:06Izar Tarandach So I'm gonna go out there on a limb and say what I always said with a bit of a modification. AI knows what we all know, and it can't imagine what none of us can imagine.

35:19Matt Coles Right.

35:20Izar Tarandach Can't imagine.

35:21Matt Coles Which is why AI is an assistance to threat modeling being done by humans.

35:26Izar Tarandach Exactly, because there is no— whatever idea the AI gives you means that somebody else already had it because it's taking it from somewhere.

35:36Chris Romeo Yeah, it's never had an original idea, right?

35:38Izar Tarandach Adam, add on random stuff as much as you want, raise the temperature as much as you want. It's coming from somewhere that it was trained to. There is a probabilistic path between the first word and the last word that actually draws back to some, something that was already written. So to have that assumption made, it means that it happened somewhere. So it's not unique to your system. It's recognizing a pattern and following it and saying, well, okay, in, in case of system like this and this, chances are that this assumption exists. Perhaps not with that thought.

36:14Chris Romeo We gotta land the plane here on our predictions for 2026.

36:20Izar Tarandach Oh, that's a prediction. In 2026, an AI is going to land the plane. Hmm.

36:27Chris Romeo It's a good prediction, actually.

36:28Izar Tarandach What's it gonna look at the end of the runway? Lord only knows, but it will. Yeah.

36:32Chris Romeo Do you want to be on that plane?

36:34Izar Tarandach Nope. Nope. It's gonna be a military plane.

36:37Matt Coles So that was an interesting prediction from earlier. Could an AI agent replace an airline pilot?

36:45Izar Tarandach Oh my God. I, I, I was going to answer something, but I would get into so much trouble. Nope. Nope.

36:51Chris Romeo No, I'm gonna say, I'm gonna answer no. I don't think it, it, for 70% of the duties, sure. But when things go off the rails, that 30% where things go off and, and go horribly wrong, pilot. Goes to experience.

37:10Izar Tarandach How do you know that an agent is capable to fly, is capable of flying a plane?

37:16Chris Romeo Because of flight simulator.

37:18Izar Tarandach It will tell you. No, it will tell you.

37:20Chris Romeo Yeah.

37:21Matt Coles Can you fly a plane? Of course I can fly a plane.

37:23Izar Tarandach No, it's more like, uh, hi, my name is Joe. I'm a pilot.

37:27Chris Romeo Maybe it's like in The Matrix where they just download, the AI just downloads things like Neo did. Like I know kung fu.

37:35Matt Coles Well, that would be skill development, right? So Claude can do that right now.

37:39Izar Tarandach Yeah, no, isn't that a well-trained model? Like that's seen enough stuff. You ask it to do kung fu, it does. And you ask it to cook, it cooks.

37:48Matt Coles Actually, you train it on the Matrix scripts. It can do kung fu.

37:55Izar Tarandach Oh, whoa, whoa. That's a prediction. That's a prediction. Okay. Okay.

37:58Matt Coles Okay.

37:58Chris Romeo Okay.

37:58Izar Tarandach There's going to be a market for fine-tuning models. Like people will come and say, I like this model. I'm a fanboy of this one, but it doesn't do enough what I needed to do. Now I need somebody to go out and grab enough data and fine-tune this thing so that it gets better at that one specific thing.

38:16Chris Romeo All right.

38:17Matt Coles I think, and that will, and that will generate a data breach just to tie it back to security and privacy.

38:22Izar Tarandach Always.

38:23Chris Romeo Oh, it's always about the branding. Matt's always about tying up loose ends. So. All right. I think that's enough wacky, wild, crazy predictions for 2026. As always, we will not be accountable for any of these predictions that we made.

38:40Matt Coles We may be AI right now.

38:42Chris Romeo We could be.

38:43Izar Tarandach Talk to my agent.

38:44Chris Romeo Uh, Max Headroom.

38:45Izar Tarandach Oh, you, you see, I always wanted to, to be able to say that, talk to my agent. Now I can.

38:52Matt Coles Wait, are you gonna open up your Are you gonna open up your agent?

38:56Chris Romeo Man.

38:56Izar Tarandach Talk to my agent. Yeah.

38:57Chris Romeo You gotta start. I thought you had an agent.

39:00Matt Coles Shodan will find agents online.

39:03Izar Tarandach Last security prediction.

39:05Chris Romeo Okay. This is it. Final one.

39:07Izar Tarandach There's going to be two, at least two great security conferences this year. And wait, wait, wait. I'm hearing the voices coming from the other side. One of them is going to be in Vienna and the other one in San Francisco.

39:25Chris Romeo Wow.

39:25Izar Tarandach Wait, is it San Francisco this year? Is it San Francisco? Is it San Francisco this year? Is it San Francisco? I don't know.

39:34Chris Romeo All right. Well, folks, thanks for tuning in to another predictive episode of the Security Table. We look forward to catching you on the cyber beach somewhere in the future.

39:46Izar Tarandach And don't do this at home. We've been trained for it.

Transcript supplied by AssemblyAI. It may contain transcription errors.

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