Ever wonder if your AI is its own worst critic? In this episode, we dive into new research showing that AI models are much worse at reviewing their own code than a competitor's! Plus, we uncover a terrifying new trend: hackers aren't just stealing your data anymore... they're stealing your entire AI model.
Sources & transcript: Episode #45 page
Read transcript: https://share.transistor.fm/s/c627ea17/transcript
Apple Podcasts web strips note links: use its existing Episode Webpage action for clickable sources and the transcript.
Automated transcript; may contain errors.
Sources/context added Nov 9, 2026; selected-topic references, not verification of every claim or identification of the original on-air references.
Sources
Join industry experts and thought leaders as we dive deep into how artificial intelligence is transforming cybersecurity, shaping defense strategies, and creating new opportunities in the digital landscape.
Automated transcript; may contain errors. Formatting/encoding cleanup only; not a full human audio or factual review. Corrections: info@aicyber.land.
Hey, welcome back to the AI Cyber.land podcast where artificial intelligence and cyber security combined to make a leaning tower a pizza, you know. So, >> it sounds tenuous. It sounds like it might fall. >> Yeah. Why does why is the tower leaning in Italy? Do you know what I'm talking about? Right. >> I know what you're talking about. Was it a found? It seems like a foundation issue.
Seems like a ground was too soft or something or >> a bad foundation model and then they built on top of it, right? So, >> so we want to have a good cyber security baseline to understand things so we don't tip over too. >> Yeah, it is kind of amazing that it's like still there just like slanted, you know, that's probably like most corporate IT stacks right there, you know, and you don't know why it's still alive, but it is.
It is. >> We're still hanging on even though it's a little wonky. Oh, I already got sucked into a tangent. That's not good. All right, we're your eyes and ears for all things cyber security and AI related. So, the breaches, the breakthroughs, and everything else in between. So, stay tuned. Uh, if you like random side rants, you should click the subscribe button because I can guarantee you there will be more of them.
Uh, but in seriousness, have you se what what have you seen in the last week that you want to talk about, Shelby? >> Okay. Um I want to share some some research by a guy named Rodrigo Cardiat and um he looked at who is finding like more like who's finding the bugs in software. So in like like before you commit it I guess. So question for you.
>> Yeah. >> Um we're going to just kind of talk about the two two big players here. Cloud code and codeex. Okay. >> Okay. >> Um out of those two which do you think is probably writing more bugs overall? Hm. It's a good question. Do they say which model that is used in the study or just in general? Just in general. >> It might be general. >> In general, I'm going to go with open AI codeex is doing more bugs.
>> Oh, interesting. So, that wasn't actually the question that this particular research set out to answer. But according to So, I don't know actually know if this data is normalized, but so far from the data, it looks like more are actually coming from Claude. And I don't know if that's because they're on the Cloud Code total or not. But here's the interesting thing that they did figure out, which is that models are not as good at reviewing their own code as the competitor's code.
>> Oh, I can see that. Yeah. >> Um, and I guess I guess people might be like that, too, because if you're having faulty logic or writing an issue into your code, you're probably the least likely person to notice it cuz you put it that mistake in there in the first place and if you knew it was a mistake, you wouldn't have written it there. I don't know.
>> Yeah, that makes sense. I So, yeah. >> Yeah. So you probably want to use one to write the code and the other one to like double check it. Right. >> Exactly. So that's the main takeaway and I'll kind of explain a little bit about how um Rodrigo went about doing this research. He analyzed 500 PRs from cloud code and 500 from codeex um and he analyzed the comments which were kind of used as the ground truth in this um to do things like sentiment analysis, upvote rate, down vote rate and the git archaeology.
>> Okay. And that was able to kind of understand like are you catching what you need to and when they when he switched it um so and then I guess he used like the slash review feature um and he did it three times. They got three times to try to catch things per PR and um it was interesting that he did start to see different patterns in that.
So, like for example, the code that was made by Claude tended to have um more bugs that dealt with like wrong data or else just straight up missing behavior in in the logic. Um and then code made by by codecs tended to have more semantic issues and error handling failures. So, I thought that was interesting. It's almost kind of holding up a mirror to see where they're weak.
Um, but yeah, and when he looked at like the the what is it called? Like the way that the AI is thinking, right, as it's going through its reasoning. >> Yeah. >> Um, and kind of leaving behind hints of how it's thinking, he found that GPT 5.5 >> tended to do like this approach of like going really in depth first. So he called it like depth first verification.
Um but then it would get kind of mixed up when there was like conflicting um like prompts happening or or commands I guess. And then um Claude Opus 4.7 tended to take like this breadth first um approach before it drilled down. Um so I thought that was interesting. Anyway, so the takeaway I think for us is do not use the same model that you use to write the tool or write the code to do the analyzing because it's not going to find as much as if you used the opposite and and just just cross those.
So um so Rodrigo helped GPile build something called model inversion which is just using AI to detect just kind of looking at artifacts and little things in your PR to detect who wrote this code which model wrote this code and then use the opposite to uh tool >> just kind of automating that little process of that. >> Yeah. Try to figure out which one actually read it.
That's cool. Yeah. >> So there you have it. That is called Model Inversion by Grapile. >> Okay, cool. I'm going to check that out. That's a that's a good tip. I feel like there's a lot of uses for that. I am frequently like changing around my like software factory workflows. Uh so yeah, it'd be nice if it could figure out which model of the week I'm using and then know automatically, hey, I should use this other model to do the review, right?
So >> there you go. Make it simple. >> Yeah, that's cool. I I've been using GPT6 Astra quite a bit. And uh man, it loves to just think and think and think. It loves to create more work for itself. That's what I've concluded, which is like a genius solution. It's like we'll have them burn tokens and then when they're not looking, we'll have to create more work so it'll burn more tokens.
So, got to get that revenue up. >> It's like a real overachiever. It's like let me think about this some more. >> Yeah. Let me think about this real hard. And while I'm thinking about it, I'm also making money for my daddy Sam Alman. Chick. Did you see that there's a uh Sam Alman movie coming out this Christmas? This holiday season? >> Oh, what? >> Yeah. Yeah.
I mean it's like uh there's a guy he's playing Sam Alman and it's like the scenario the which is you know based on history where Sam Alman was let go as the CEO uh there's that week where their like board basically fired him and then uh Miriam Morate got put in I believe and then then they reinstated Sam like after a week. So that's what the movie's all about >> and they got someone pretty oh they got Andrew Garfield to to be he's going to be Samman in the movie.
Do you know who he's he was? >> He was a Spider-Man, right? >> Yeah. Yeah, he was a Spider-Man. Yeah. Yeah. >> I feel like he's too likable. You got to find someone that we all hate to be Sam Alman, you know? >> Yeah. Yeah. I thought it was He's the trailer. He seems to like mimic his mannerisms really well. >> Wow. >> Alman's got kind of like a weird walk and some other tails.
>> So, he seems to like mimic that stuff really well. So, I think he did a good job. Uh but yeah, he doesn't like really strike you as like having the same core personality as Sam Alman. So, it'll be interesting to see how the movie pans out. Trailer looks good. Uh, yeah, I like movies. I'm That's definitely one of the things I enjoy, which >> that it comes out in December.
>> Yeah, it's it's like sound like around the holidays this year. And uh I don't know what the title is, but Andrew Garfield also made another movie called Under the Silver Lake, >> and it's kind of like a cult hit movie, so you either like love it or you hate it. And uh yeah, I'm definitely in the camp of I love it. So you should uh everyone should watch it.
>> The Silver Lake. Okay. >> And then uh you can hate me for recommending it to you. So So >> Well, thanks. That made me really excited for it. >> It's like a Have you ever seen Dotty Darko? >> Yes. Yes. >> It's kind of like a Donnie Darkoesque. It's It's not It's not like the same, but it's kind of similar to that. So >> if you like Donnie Darko, you'll probably like Under the Silver Lake with Andrew Garfield.
So >> Okay. That's We're not I'm This is not an advertisement. I'm not getting paid to say this, by the way. So, uh >> but if if anyone does want to pay Bryce to say these things, he's going to say them anywhere. So, >> yeah, hit me up. Send me an email. I will I will say about anything. Uh >> what you see in the news though lately?
What's going on? >> Oh, yeah. So many companies. What do you think the most expensive thing they built this last year was? >> AI. >> Yeah, they built AI. And several of them built their own AI models >> and their models are particularly good at specific tasks to help their companies make more money. Like we've talked about previously, Crowdstrike just released some models to help blue team and red teamers.
>> Oh yeah. >> Well, would you believe it? Attackers are now targeting your AI models to steal them like they steal intellectual property. Uh Mandy just released a report and it details two different organizations. One is a healthcare company that had built their own proprietary AI model to like help doctors. And the other one was kind of like a media generation company where they would, you know, create ads and things like that for different uh, you know, customers.
And, you know, they they went in basically stole everything like they stole the weights, they stole the source materials, they stole all that stuff. And um, it's kind of interesting, right? Because it feels like that would be kind of harder to monetize, right? rather than just doing like a standard ransomware operation. >> But um they are also so like in a standard ransomware like operation, it's like you got your house and let's say Bryce has a house and then when Bryce isn't home, Shelby comes over and just changes the locks to my house.
And then when I come home, I can't get back in my house. That's like ransomware, right? They like lock your computer so you can't get into the computer. And then people were like, "Well, I'll just buy a new computer. They're not that expensive. Like it's cheaper to buy a new computer than it is to pay you. So, um >> that's so funny. Or buying a new house, I guess.
>> Yeah. The analogy doesn't really work with a house because houses cost a little bit more, right? But um so then what the attacker started to do is they're like, "All right, we're not just going to you like change the lock on your house. We're also going to steal all the stuff you care about. So we're going to steal your journal. We're going to steal your pictures.
And if you don't want us to put your journal on the internet, then you got to pay us money, right?" So that's kind of like ransomware 2.0 0 where they're at now where they like but this is like more like ransomware 3.0 where they're like where you spend years designing your dream house and then they just roll in and steal the plans and use them for themselves.
They're like we got the blueprints let's go. So uh yeah it's not great >> sneaky. >> Uh they are also quite sure in an ironic twist that the actors are using AI to breach into the organizations to steal their AI. Uh so yeah, >> of course. >> Yeah. Yeah, that's kind of a given. We've talked about that a lot on the on the show. Uh this is the same they've attributed this back to Team PCP.
Uh I don't know if we've talked about Team PCP on this before, but they are the ones responsible for most of the supply notable supply chain breaches. So uh they, you know, infected the NPM packages uh that a lot of software engineers were using and they've, you know, are behind a lot of those uh supply chain attacks, right? So they they just figured out, you know, like instead of like trying to hack everybody, if they just hack the place where they store their library that they can get into everybody from there.
So anyways, there wasn't any like correlation between the supply chain attacks and the AI theft, but it would make logical sense that those two things correlate at some point. So what does this mean for you out there? Like if your company is building any type of intellectual property, especially something where you're spending a lot of resources to make a model, you know, you definitely want to take extra precautions to ensure that like that information is not going to walk off.
And you know, you need to kind of treat those resources as crown jewels. I mean, you know, like, yeah, you know, I think a lot of companies are so pressured right now to get AI out the door, to build the best AI capabilities that, you know, they're not really thinking about, uh, you know, what happens if someone comes in and steals all our hard work. So, anyways, I just thought it was an interesting twist.
Uh, something a little bit different than what we normally see. And, uh, >> I've got a solution in your directives for your AI. Just write. If me not actually equals me, self-destruct. >> Self-destruct. We Oh, I got a great idea, Shelby. It's a poison honey model. So, it looks like it's like a normal model that you would want to steal, but when you like run it locally on your laptop, it just destroys your whole laptop.
You know, like there goes, "What tools do I have access to?" Oh, the terminal. Great. >> RIM- RF. Delete. >> Doesn't look like home. >> Yeah. Yeah. >> Better self-destruct. >> Yeah. The rest of this model just just garbage to distract you. The real purpose of it is delete all your stuff. I wonder could someone build like a malicious model and like upload it to Hugging Face and then like try to get trick their friends into downloading it and like like hey bro run this new model, you know?
Like >> I'm seeing it as more of like a time bomb thing, like a trap, right? It's a it's a good working model, but if all of a sudden it's not running from your company's internet or whatever your company's network, it doesn't work. >> Or maybe like that's a good idea. Maybe also like a sleeper agent where like if you say the right word that it triggers the right path inside the model and then that deletes your whole that deletes your whole computer or take some action, you know.
So like >> there you go. It's like spy stuff. >> Yeah, spy stuff. >> It's counter attacks. >> Zoolander. They both kind of relate, you know. So wasn't there some word in Zoolander which he would say and that would cause the male models to to try to murder people? I don't know. I haven't watched that movie in a while. So, >> I haven't seen it actually. >> No, never seen Zoolander.
>> Never got around to it. >> You're missing out, Shelby. Missing out. That is >> It's like one of those things. It's so popular in culture that I can still quote a lot of the lines. So, >> true. >> Does that count as having seen it? >> No, it does not count. Does not count. If you think that counts, leave a comment down below. We're going to see who's right and who's wrong.
If you think Shelby should watch Zoolander, do a like button. The more likes, the more likely she is to do it. I just want to throw out there, you should watch Zoolander. Uh, did you see anything fun this week, Shelby? >> Oh my goodness. I just got back from a hike with a handful of kids and it never occurred to me that we should bring swimsuits because we're entering fall.
It's getting a little chilly and where we hiked, the water is very cold. It's also coming off of a waterfall. It's shallow and it's very cold. Oh, no. Just fun fun uh little piece of advice. When you go with kids to the water, just bring a swimsuit. It doesn't matter that it's ice cold water. They'll jump in anyway. >> It was a little chilly this morning. I'm surprised.
And then you add the altitude and then you add the chilly water. That's uh >> that's some dedication right there with those kids, you know, to >> They're crazy, but they were so happy and wet. >> Just living carefree, their best life. >> What do you have for us, Bryce? What you been up to? >> Uh in the third installment that everyone's been waiting for so patiently, I built the rock Pokemon.
>> Yay. Geodude. Yeah, >> Geodude. Yeah, you knew the name, man. You're so good at this. You crushed it. >> Well, thank you. >> So, I got Geodude. I got Charmander. I got Jigglypuff. We'll see what comes next. Uh, I might I might go do something else. Uh, I also >> I need a Squirtle. >> Oh, I do too need a Squirrel. That would be so good.
All right. All right. I'm going to get in trouble, but more Legos are coming. Um, I also saw the new Spider-Man movie cuz we were talking about Spider-Man earlier. I really enjoyed that. I thought that was really good. So, uh, check that out. Even if you haven't seen any other Spider-Man movies, it's it does it does pretty well as a standalone. So, uh, did you have you seen that yet, Shelby?
>> No. >> Okay. All right. Well, it's worth a it's it's worth a watch. That's what I That's all I'm going to say. Uh, anything? Okay. So, I think that's a wrap on today's briefings. AI and cyber security move fast and we're here to keep you ahead of it. Uh, so until next time, stay safe, stay informed, and we'll see you. Bye. Bye.