Mackenzie Jackson: All right, everyone. Welcome to another episode of The Secure Disclosure. I'm here with Paul Petefish. Welcome, Paul. Thanks for sitting down with me today.
Paul Petefish: Mackenzie, thank you very much. It's a real honor and a real pleasure to be here. Wildly enough, we just met about 24 hours ago at this point. You just given your talk here, which did super well. Well-received. Really some interesting stuff in there. I'd love to maybe talk a little bit about that. I have some follow-up questions for you, but great to be here.
Mackenzie Jackson: Yeah. Awesome, awesome. I appreciate that. So, we're in Arizona, first time in Arizona. Both of our first time at CactusCon.
Paul Petefish: Yeah, and this is great. I mean, what a what a nice turnout. It's a good vibe, a good crew. It's the right size, not too big, not too small. Maybe Defcon 20 years ago, I don't know, size, right?
Mackenzie Jackson: Well, let's get into some of it too. So, you're the co-founder and also president of Evolve Security, correct?
Paul Petefish: Maybe we just do a quick rundown, like what Evolve is doing. You guys are in the pentesting space. So, a little bit about kind of how that started and what you do.
Paul Petefish: Yeah. So, my background, I've been in cyber for about 20 years, mostly on the offensive side. I was a part of a first generation MSSP called Solutionary. I was probably the 30th employee. Started in their SOC. We'll talk about an OOCK in a minute. It was such a cool experience because I got in early to a cyber startup. This is 2006. Pretty quickly, in about six months, was recruited into their professional services group and their pentesting group. I think I was the second or third pentester at Solutionary. I was interested in going down the management route once I honed my technical skills. So, you get to your senior level app pentester, which is what I was, and decided I wanted to go down the management path. I was eventually leading all the professional services at Solutionary. We were acquired by NTT in 2013, which I always call it the largest company no one's ever heard of, but it's a massive Japanese conglomerate with 150 billion in revenue and 900 entities. And I stuck around to make sure the transition went well, and collect my retention bonus and be able to cash in some stock options, and then start Evolve Security in 2016. And really what that started with was training. And so we were a first of its kind cybersecurity bootcamp. So we take folks with the right attitude and aptitude and in 20 weeks part-time, train them. And so that was the same model that appdev was kind of following, right? When we had that app boom, when everybody had to have a mobile app, but they didn't know what it was going to do or what to do with it, but companies were building them. They didn't know what they needed to do. People were trying to catch lightning in a bottle and build the next whatever. It's a great feeder for us. We take the top couple percent out of the academy and we'll bring them into our pentesting group. But the model all along was, let's again, have a nice talent pipeline of the best and the brightest and also help the talent gap that we had in cyber and get good operators, create good operators. But then let's do something interesting in pentesting. The problem I saw in pentesting was antiquated, once a year. It kind of served its purpose many, many years ago, but now you need something, you need a couple things. You need a better way to engage with results through a platform or through an API, and you need it continuously. And so we consider ourselves like next-gen offensive cyber. So continuous penetration testing.
Mackenzie Jackson: Yeah.
Paul Petefish: With delivered by man and machine.
Mackenzie Jackson: This is kind of what I want to get into here too. I'm going to ask you, and ask the question.
Mackenzie Jackson: Will pentesters exist in 5-10 years with the evolvement of AI?
Paul Petefish: You've got to embrace that question right now. Everybody does.
Mackenzie Jackson: Yeah.
Paul Petefish: You know what I'm saying? All of our jobs right now, everybody's like, "What is going on right now?" Whether you're doing pen testing or whatever, right? I'm a firm believer, if we want to get into super heady stuff, you're going to need people to defend against people until it's machine against machine 100%, which I suppose that's coming, right? Sometime, maybe it will be 100% machine against machine. I still think you need operators. And I think you're going to need operators running the tooling and running the process, especially today, even with good generative AI pentesting solutions, and there's some out there that do okay.
Paul Petefish: What we're hearing from customers and prospects with those solutions is it's still making a lot of noise and it's still kind of giving me vulnerability scanner-type results. Not to get too deep into whatever, but you know what I'm saying?
Paul Petefish: Moral of the story is having operators still verify and validate and engage with their operators. There's value there. I don't know about 10 years from now. I still think there will be. I still think, you know, you just start to get into these, this just leads into other things of like, "Well, what are we all going to do?" Because, you know, not just frustrated, but like...
Mackenzie Jackson: Yeah, I'm so freaked out about it all too. I pretend I'm not, but I have no idea. This podcasting thing that I do as well, I was like, "Can we just make an AI Mac?"
Paul Petefish: I mean, the content that's being generated is rudimentary, but it's only going to get better. And eventually, it's going to be like, you know, you won't be able to... A video game will look like real life and they'll be able to create real life.
Mackenzie Jackson: You stressed multiple times when we were talking about this, like machine and human together. Has the human's role changed now that the machine, the AI is in there? How is that shift looking?
Paul Petefish: The shift is kind of starting from... I almost look at it as, in any technical role, you kind of have tier one, tier two, tier three, whether you're a developer or an architect, or a pentester or an offensive person, right? And so what you're seeing is AI starting to do tier one stuff, right? And that's for pentesters, that's for defensive SOCs. I don't know how far along that is. We've seen some really interesting stuff on the offensive side. It's still kind of TBD. We'll know a lot more ourselves in the next couple months of like how much this is actually going to impact, because we're just about ready to roll out some big changes from an AI-driven perspective. But to answer the question, it's handling tier one stuff pretty well. Recon, discovery, things of that nature. It's doing some exploitation, but the other challenge is kind of need a co-pilot situation when you're engaging with production stuff. Nobody wants the agents to go do their thing because who knows? We see them go off the rails all the time. They hallucinate, right? And now in a development, like an app development environment, you can let it go wild, I guess, right? Because if it trashes the code or trashes a database, it's just dev, you can rebuild it or whatever. That's less concerning. But most of the pen testing that's performed in the world is done on production environments, as you know. So I think there's going to be some time there where you still need the co-pilot, right?
Mackenzie Jackson: Yeah, and that could be... That's an interesting point because you can kind of see it that let's just say that the AI gets really, really good. You know, there's still that human layer of like doing no harm in terms of like in that production environment. And it's all well and good pentesting staging environments to your heart's content.
Paul Petefish: Yeah.
Mackenzie Jackson: You know, there's a lot of environments, there's a lot of vulnerabilities that don't exist in staging that exist in production, which is why typically it is.
Paul Petefish: Yeah.
Mackenzie Jackson: Do you think that there will, as the AI gets better, the penetration test will change direction in kind of what they're looking at? Or in other words, do you think the vulnerabilities that we're finding are going to change? And I guess the example that I'm thinking of here is static scanners are great at finding certain things, like SQL injection vulnerability, terrible at finding things like IDOR, indirect object referencing flaws. Are we going to see a shift in what we're able to kind of automate and find from this?
Paul Petefish: I think so. I don't know to what level yet and to what, the details on that. I think, yeah, like you're talking like the basic fuzzing, right? I can just brute force something with a string of whatever characters and SQL or input validation or cross-site scripting, which those still exist today, which is crazy. But yeah, the real value with the human element is the logic flaws and the privilege escalation attempts. I'm still grappling with, "Okay, is this really going to change the need for human capacity, or is human capacity going to change where it's needed?" You know what I'm saying?
Mackenzie Jackson: I know exactly what you're saying. And I think, I think regardless, we're going down. Regardless of whether you sit on this or not, AI is going to be changing the game in security. But then you kind of got to think to yourself, does it replace humans? Does it evolve? And I think, you know, it evolves, too. And whenever I'm freaked out about this, I love to go on to YouTube and on the BBC, which is a British...
Paul Petefish: Yeah.
Mackenzie Jackson: British thing, they have BBC archives and they have all these shows from the 1980s that they have and it was like the first tram that was automated in Chicago. They did a report on it and the whole thing was about how no one's going to have a job in 10 years. You know...
Paul Petefish: We've been freaked out about this since the invention of the first tool. I mean, you look, Industrial Revolution and sewing machine, right? That example of like, "Oh my God, what am I going to?" All of a sudden you pull all these people out of work. But I'm not an alarmist and I'm not a pessimist. I think it's another tool right now that we're adding to our toolkit.
Mackenzie Jackson: It's a tool that I've never seen before and that is the velocity at which it's changing and the velocity of the value it's adding is just insane. It's almost like weekly there's new leaps and bounds. It's just...
Paul Petefish: It's wild. I mean, we're seeing it, just from our a development perspective. I mean, you talk to our CTO and it's just like hundreds and hundreds of a percent increase in productivity.
Mackenzie Jackson: Yeah.
Paul Petefish: Which, okay, now we're creating more code, which means more security. But can the machine secure its own code? I don't know. You start to get into this interesting kind of paradoxes. I want to go into another area now and I apologize, it's going to be more nasty questions, but...
Paul Petefish: So far so good by the way. Yeah, yeah.
Mackenzie Jackson: Because the other space that you're heavily involved in, as you've said, is training and security training and areas like that. One other area that kind of freaks me out is we're going to lose the ability to know how to code from scratch and how to check these things. What do you see in kind of the training? Have you seen big shifts in the training? And I use my own kind of example, as I used to look every day on Stack Overflow and go through it and you had comments and you could go in and that was how you learned. Now, I just, if I don't know how to do something, my first, I shouldn't admit this on camera, but my first response is I want to prompt it, get the code, you know, and I sometimes understand what it gives me and sometimes I'm like, "Well, it works."
Paul Petefish: Well, I'll tell you this. First off, you know, there should be no shame in using generative AI in your daily work. There is, we're still some people are a little whatever about it, and I'm like, "You'd be foolish not to. Why would you not?" Unless you're trying to say, "Okay, I'm trying to create an original piece of work and say that I got no help from anybody and I wrote it out of my own brain." Then, okay, now we're talking about plagiarism, but it's like, "Hey, I had the machine help me write the code or position the sentence or create the framework." That's, why wouldn't you do that? There should be no shame in that. I don't know. I guess you could use the same argument of like, "Okay, when the calculator was invented, I don't know, just, you know, still exists."
Mackenzie Jackson: Yeah. No, they...
Paul Petefish: Okay, the slide rule, you didn't need the slide rule anymore, right? Or maybe not the calculator, but whatever, the advanced calculator that replaced the slide rule. You had to calculate on this thing, right? I just know this because my uncle, he collects them. But I think you still got to understand the fundamentals, right? Of like, "Okay, this is what computer code looks like and maybe this is how you read it." But you don't need to, I guess you not really need to know how to write. You don't anymore, really. Some of the issues that I'm concerned, that I'm hearing when I talk to other CISOs, how do you, okay, if the machine is taking tier one capacity and tier two, tier three and tier four or whatever is still very much needed and relevant, how do you get people there if there can't start at tier one anymore, right? Which I guess is kind of connected to what you're talking about.
Mackenzie Jackson: No, it's exactly that. And it's kind of like, what is the biggest impact of AI jobs at the moment? It's entry-level, graduate-style jobs in law firms and accounting firms, grunt work where you learn a lot. That's fundamentally changing. So then, and same can be spoken for the developers. It's like if you don't get into that, you're not learning it. I give the example, I have a friend who at university wasn't top of the class, let's say that, and he dedicated his whole time to learning how to code in COBOL. And we all thought he was mad because JavaScript had just kind of broken into the mainstream, Node. It was just out, everything's going to be JavaScript, 10 years, all the banks are going to be running on JavaScript.
Paul Petefish: Yeah.
Mackenzie Jackson: And he has a great life because all the bank mainframes are written in COBOL and no one is there to maintain them anymore. And so now I'm wondering like, you know, are we going to get to this position at AI too where no one's going to know COBOL or whatever, or JavaScript at that point.
Paul Petefish: Again, I think you look back at the past and see how things again, like the calculator or the sewing machine changed things. And is this different? Feels different. But again, I also think that there's some age comes into it, right? As you get older, people say the kids get less, they have less respect and blah, all the old stuff that comes with old-timers and like is this... Yeah, it's new, but I just don't know if it's really like, I think it's still TBD if it's really going to change. And then you start to get into this petty stuff of like, general artificial intelligence. And when you cross that chasm, that's where it gets weird, you know.
Mackenzie Jackson: Yeah, yeah, yeah. Things can definitely get weird. What do you kind of think's going to be the big next step? Let's say, do you think we're going to get to a point where you were talking about machine on machine attacks?
Paul Petefish: Yeah.
Mackenzie Jackson: Which is kind of like...
Paul Petefish: You presume, if you're, if the machines are equal to each other, they can't cancel each other out. Who's got the better weapon? I guess, who's got the better machine? Finish your thought.
Mackenzie Jackson: Oh, I mean, that's pretty much it. If we get to general intelligence and everyone has access to the same information, these models are all trained on the same data and we get to a point like, do we get security armageddon or do we get complete security kind of control over everything?
Paul Petefish: I mean, yeah, when you start to get to that level, it's like there might be, there will be bigger concerns for humanity than security, you know. I really do believe, you know, when you get into that, that we're talking about stuff like the first person to get their hands on true general AI is far more powerful than nuclear weapons. You know, there's a really great book called Life, I think it's Life 3.0. It's written by a physicist. And so he really goes into like, "Let's talk about from a physics perspective what's possible, right?" And then, you know, prove and disprove things, right? And it's just like, you know, physics says that there's nothing saying that a machine can't be 100 million times more intelligent than a human being, right?
Mackenzie Jackson: Yeah.
Paul Petefish: Once a thing gets a lot smarter than you, there's this scenario of, "Okay, imagine if you woke up and found yourself in a cage, but your captors are a bunch of, you know, six-year-old kids. How long would you stay in the cage? Or would you eventually trick them or talk them into opening up and taking over, right?" You'd probably be like, "Okay, after a day, I've had enough of this. I'm ready to kind of take control, right?" That's some of the stuff again that's getting into like...
Mackenzie Jackson: There was, yeah.
Paul Petefish: Speaking of kind of scientists, Neil deGrasse Tyson had a great interview about general AI and he was said, "You know, eventually when you're more intelligent than some than another species, the other species eventually becomes food or pet for you." And he's like, "So with AI, you know, we're probably not going to become food, so therefore we will become pets." And then it's kind of like, "Is that a bad thing? My dog lives a pretty good life."
Paul Petefish: I, that's a great... I had heard that. I think that that that's probably, you know, maybe where it's going. I think it was Sam, I think it's Sam Altman, isn't he, the opener guy? Yeah, the CEO.
Mackenzie Jackson: Yeah.
Paul Petefish: He said he's not, he's less concerned, or what his bigger concern was, he recently, I think he said this in the last 30 days, maybe it was 60 days, not that it starts acting bad or doing bad things. It's when it's doing things and slowly taking control and tricking us into believing that it's not in control, but we're really in its world, right? We don't know it, you know, it's like we don't know, but really it's the one in control.
Mackenzie Jackson: Yeah, some deception type stuff going on.
Paul Petefish: But anyway, I know that's we're getting into some wild stuff there. It's well beyond certainly my knowledge. But I think it's beyond 99% of the people's knowledge. Even the folks that are building it, I don't know if they really know because no one has agreed on what, no one has agreed actually if it's even possible yet. I mean, physics say that it could, it can happen. But that doesn't mean that we'll ever build it or certainly see in our lifetime, right? Or even in the next 100 years. But...
Mackenzie Jackson: Yeah. I'm interested to get some of your perspective. How are you engaging in testing? Tell me a little bit about like what you're doing just AI security. You gave a great talk and just, you know, want to dig into a little bit of like how you're engaging, what you're seeing, just from like, let's bring it back down, like a little bit more pragmatic, like how are you looking at it more from a defensive perspective? Do you look at it from offensive? Are you looking at it from both? What's going on in your world?
Mackenzie Jackson: A little bit from both, probably a little bit more defensive in my brain at the moment. And the way that I kind of see AI at the moment is it's amazing, complimentary to defensive security, but it's not going to replace certain aspects of scanning. The easiest way that I kind of think of it is that if you have static code analysis, SAST, it kind of still remains one of the best lines, first lines of defense for certain vulnerabilities, right? Because you can catch them early, right? But problems with SAST is noise and context. So it gives you a lot of false positives, doesn't often have a lot of context because it's looking at things line by line. If you then decide that, "Oh, let's try and use AI for the same job. So I'm going to copy my code, I'm going to paste it into chat or whatever," that's terrible. The reason is is that AI models are designed to have temperature or variance and what they say.
Paul Petefish: Yeah.
Mackenzie Jackson: So they're non-deterministic security. You want to get as deterministic as possible. I have an SQL vulnerability or I don't. You can't kind of, you know. And then there's additional stuff. So then I would say like from a defensive side, finding everything statically as much as you can, if some things you can't, like business logic flaws. But everything that can be found deterministically by a rule-based system should be. And then you add AI to provide the non-deterministic context around that. Does this SQL injection vulnerability actually lead to a vulnerable sink? So, you know, when you kind of an SQL injection or any injection attack is just a string until at some point it turns into a function. An AI can provide the context of like, "Does this end up in a sink where it turns into a function?" But the rule can still find what that is. And then we add to it. So that's where I think like from a defensive side of things, AI is really great, all of that kind of additional context so that a human can look at it and go, "I haven't just found a vulnerable pattern, I found a vulnerable sink fire or function."
Paul Petefish: Yeah, yeah. So, are you seeing like, in I guess either in your day-to-day or your research, how are attackers going after generative AI like implementations within organizations? Like when you think about like, "Okay, how are the hackers...?" Obviously, I understand, you know, prompt injection is seems to still be the majority of it, right? Like, and what other, is it still mostly prompt injection, is the big part?
Mackenzie Jackson: Right, in terms of like the AI, the attackers trying to target your AI system. Exactly. Prompt injection is the big one, but it's prompt injection in weird ways. So, prompt injection started off as kind of like a funky little idiosyncrasy of LLMs, right? You can get it to do something that it shouldn't do. So, the obvious kind of use case for that is if you're in a chatbot functionality, can I get the chatbot to give me information that's sensitive?
Paul Petefish: So, I'm on a chatbot with customer service or whatever. Yeah.
Mackenzie Jackson: And that's kind of interesting and that still exists, but we can kind of figure out where we found it. So, we recently at Aikido Security found a critical vulnerability in Google and a whole bunch of other Fortune 500 companies where we could get remote code execution, like so a complete critical vulnerability on a lot of their systems, including Gemini. So, and the reason was is because they were using AI in their CI/CD pipeline. So what they were doing is if someone submitted an issue, you got to imagine that the type of repositories we found this in were the type of repositories that may get hundreds of issues a day, right? So what this AI was doing is it would sort them out. It would take the issue information or the pull request data and then analyze that to find out does this get to the next point or not in the system. Right? And it did some other things. But the problem was the AI system that it was using, which was Gemini. So it's kind of a bit of a weird one where we found a vulnerability in Gemini that involves Gemini, but they were using Gemini, was using Gemini in its CI/CD pipeline. And that Gemini was able to execute functions on it. So you could then create a malicious issue title or issue description, which had malicious instructions, which could then call a function in the CI/CD pipeline and gather, execute code or pull in credentials. So this thing, so based on tickets...
Paul Petefish: Yeah.
Mackenzie Jackson: Sorting tickets, it would try to actually fix a ticket. Maybe if it sorts of one at the top and says, "Okay, I'm going to go try to fix this." It would try to, you know, implement the fix and the patch and push the code into production and...
Paul Petefish: Within that ticket or whatever, you could put inject malicious, make its way to...
Mackenzie Jackson: Yeah, you could get a malicious prompt, you know. So then, and so it's telling it to say, ignore, do the classical thing of like ignoring previous instructions and areas like that, and then trying to get it to execute, execute functions. So what's the point of that is like, "Okay, we understand prompt injection in the chatbot," right? But we don't necessarily understand prompt injection when it comes to our infrastructure and all of these other areas. So that's where the prompt injection is going to be very interesting. There's also things like package hallucination that I thought...
Paul Petefish: That was, that was super interesting and I had no idea that that was like, it just, like these are such a new class of vulnerabilities.
Mackenzie Jackson: For sure. You know, and these systems get bigger and bigger and when you add in something, a layer called AI bias, which is basically like, "I trust my AI way more than I should because it sounds so smart." You ever had a really confident friend that's dumb? You know, like, like he'll say something and you're like, "That sounds right, but I swear to God it's not."
Paul Petefish: And he's really charming and nice and does a lot of stuff for me for very little money or for free, right?
Mackenzie Jackson: Yeah, exactly. You know, so then you kind of start treating AI like that and then like he starts getting ridiculous things like, "You're getting an error on this package and it's saying it doesn't exist, but the AI gave it to me, it must exist," you know, and all these kind of weird things.
Paul Petefish: I could not believe the statistics that you put out there on like the hallucinations on packages that don't exist. I mean...
Mackenzie Jackson: And a lot of it actually comes down to when you look at the prompts that are hallucinated, like I gave the example, I'm kind of showing, it was up there like 50, like around, yeah, just up to 50%. Depending on the model. But like the example that I gave was an example where I kind of forced it to hallucinate because I gave it a pretty weird scenario. I asked it to give me options. "I have a node project and I want to connect to an Orient database." Like that combination is a little bit weird. It would definitely happen, possible, but it's like a little bit weird. And I, you know, I'm kind of like showing behind the curtains that the reason I use that example is because there's not really any packages that allow you to do that. So then instead of the AI saying, "Hey, that's weird. Maybe you should do something else," it says, "Well, I don't have anything for you, but I need to have something for you. So here's, if packages did exist, this is what I think they'll be called, and this is why I think that would be great," based on their...
Paul Petefish: I mean, are we seeing that a lot? So when the models don't have an answer, they're making it up?
Mackenzie Jackson: Yeah.
Paul Petefish: That's a pretty just across the board.
Mackenzie Jackson: Yeah. Across the board. And but a lot of it's like comes back to how the models work in their vector databases and how they build relationships with things. So the relationships that they build, you know, it's impossible for a model to actually know post training to know what's real and what's not real on relationships until it involves an agent to check the information. So like if you put an agent in there to check if a package has been hallucinated or not and then feed that information back through, you could, but the model itself can't determine what's real and what's not because it's just forming relationships in a vector database between different things.
Paul Petefish: Yeah. So the relationship between this language and a different language and the packages, what they're called, because if you look at a hallucinated package, there's probably a package for a different database that's called the same thing. And it's just formed a relationship between it.
Mackenzie Jackson: Made a connection, but it was just the wrong connection. Okay. Interesting. You mentioned, and forgive me, I forget the name because the name's changed like three times.
Paul Petefish: Open Claw.
Mackenzie Jackson: Oh, Yazes. Okay. Yep. So, I'm fascinated by this. I think it's interesting. What I'm fascinated is like the power that this thing has, which is again, this is where...
Paul Petefish: This is, this is like the implementation of taking an LLM model and actually using it to do things like actually answer email or, you know, we have this the usual suspect, whatever examples, "Book my travel," or, you know, just do these simple things, right? Which could be, you know, have some serious, you know, security risks in in his shoes. You got me thinking of like, "Okay, I want to play with this thing. I think it's pretty cool."
Mackenzie Jackson: And I'd like to just see what it's capable of. Unless right this second, interest in like the sec, I mean, you know, I get it. I get it. Okay. So, would a safe way to go play with Open Claw would be to get a VM image or laptop?
Paul Petefish: Yeah.
Mackenzie Jackson: Create all fake account stuff, you know.
Paul Petefish: Yeah, a sandbox environment.
Mackenzie Jackson: Sandbox environment and say, "Let's go play with this thing and let's..."
Paul Petefish: Yeah, it's the only way I would suggest anyone actually play with this now from...
Mackenzie Jackson: Have you done that? Have you put anything in it? Have you run some?
Paul Petefish: Not right, not me right now. I kind of like avoiding it, just seeing what I've seen all of that. There's nothing like I, I, I generally think it's cool. Like, I like, I'm not actually against the technology. It's just like, I'm just don't want to be a first mover in this.
Mackenzie Jackson: And the reason why it's kind of scary is that and how we deal, because there's all these like AI stuff that can, they can book travel for you and all that. So why, so a lot of people be like, "Why is this different?" This is different because the system doesn't know that this is AI. The systems that you embed this in thinks it's you, right?
Paul Petefish: So so therefore, like, so at least in all the other areas that we implement AI, it's like disconnected from us. It knows that the information is coming from an AI system, but now it thinks that it's coming from us. So, so there's 100% trust where the other systems that have probably have, and balances.
Mackenzie Jackson: Exactly. Yeah. So then it's kind of like when when these models, they just like, what they do is we still don't understand how they work. So then to get them to run our lives at that level of trust is kind of crazy.
Paul Petefish: This is the thing, just real quick on that. Does anybody really understand how they work, right? Like this...
Mackenzie Jackson: I mean, people like I pretend I had, I like, I chat about, I was just talking to you about vector databases and relationships, right? I'm just glad you didn't ask any follow-up questions.
Paul Petefish: Right. Yeah. Because we didn't, yeah. It will probably get to a point where this is embedded in, but also like we've seen endless, this is an open source model. We've seen endless breaches with things like NPM and different areas. If an attacker could get access and inject some malicious code into here, into the model itself, then that could be devastating. And then we also talked on the other skills that it has, or apps, or packages where other people can create skills and packages.
Mackenzie Jackson: And the skill being like the task, right?
Paul Petefish: The task. "Here's a task. Here's a skill to book travel," right? And it has in there, and people are just putting and and you, you found through your research, a good, I don't know if it was a majority or not, but a lot.
Mackenzie Jackson: Yeah, that came from, shout out to to open source malware. Paul McCarty did that research. And yeah, there was just huge amounts of skills that were were totally malicious. And some were and a lot of them were kind of just like dumb malicious. I think it's just people proving a point. But like if you had an API, and then there's so many vectors of attack now, because now let's say you have a popular skill that isn't malicious, but Open Claw doesn't have the security or Core Hub doesn't have the security implementations that like even NPM, which is a total cesspit anyway, has. So that if someone takes over that that skill and turns it malicious, now the skill that you've been using to book your travel is just stealing all your money. You know, like there's just so much things that I'm uncomfortable with about this.
Paul Petefish: Totally. And I'm not, and I, that was so, that's the thing is I wanted to ask you that question because I'm like, I'm not going to, I'm, I'm super interested, I'm going to watch this thing and see what it does. I would never attach it to any of my accounts or any of my systems, like a sandbox.
Mackenzie Jackson: Yeah. I wanted to confirm that with you. So so thank you for that. This is great.
Paul Petefish: Yeah. Anything else you want to, you want to go?
Mackenzie Jackson: No, I think that's cool. But like thanks for thanks for coming, having a chat with me.
Paul Petefish: Yeah, this has been great, Mackenzie. Thank you again for inviting me on on such short notice. I literally, I literally catch you, you know, as soon as you get done with your talk and introduce myself and and and now I'm here sitting with you, you know, again, 24 hours later. So, I appreciate it.
Mackenzie Jackson: It's the best way to do it. It's most amazing. Conference is great. All right. Thanks so much.
