Geri 0:00 It became much easier to become a cyber criminal because you don't have to do all this hacking yourself. You can buy the excess. You can buy your ransomware. And because of that, it becomes very opportunistic.
Jo 0:12 Welcome to Shielded, the last line of cyber defense by PQShield, the podcast where we are moving the conversation around post quantum cryptography from the why to the how. I'm Joe Lindzen, your host, and I'm here to guide you through the challenges, innovations, and solutions that will help your organization stay secure in the quantum era. Let's future proof your defenses together. Welcome to Shielded. I'm your host, Joe Linson. And today, we're joined by Gary Revay, principal security researcher at Fortinet's FortiGuard Labs. With more than fifteen years of experience in security research, ethical hacking, malware analysis, and penetration testing, Gary works at the front line of threat intelligence, helping protect enterprises, governments, and critical infrastructure from evolving cyber threats. Welcome to Shielded, Gary. Thank you very much. Nice. Very nice to have you. So, obviously, today is day two of Mobile World Congress here in in Barcelona, and I do believe you had a session today at the, uh, security summit. Yeah. It's the TSMA security summit. Which was a presentation and then I think a panel discussion. Correct. And it was titled intelligence in action, predict, prevent, protect. So I'm curious about, you know, your takeaways. How did it go? Did you get any silly questions or good questions?
Geri 1:40 Unfortunately, it wasn't very interactive, but I think it went on. So my talk specifically was about fighting AI with AI. So this is one of my research I did recently to look at how threat actors or attackers are using AI and how defenders can use AI, and I basically compare these two to try to figure out who has the advantage at the end of the day.
Jo 2:04 Yeah.
Geri 2:04 So I think the audience liked it. So the people I talked to, they were very happy with it. Nice. The panel discussion was also interesting. There was more general about threat intelligence and how to use threat intelligence, how to make it actionable, and also things like how to present the idea of security operations to your CFO, how to, you know, get budget, get budget, these kind of topics. Yeah.
Jo 2:27 Yeah. Yeah. Not unimportant.
Geri 2:29 No. No. Absolutely not. Great.
Jo 2:32 Well, before we dive deeper into the topic, I think as we spoke before the podcast, you've described yourself as a hacker at heart. So I'm curious to hear a little more about, you know, what motivates you, what drew you into the profession in the first place, what kinda influenced you Mhmm. To choose that career path?
Geri 2:48 I think there was something there when I was a kid. I was the kinda kid who would take things apart, but not necessarily put them together. A tinkerer. Yeah. The so so I definitely like to take things apart. Getting into security was relatively spontaneous
Jo 3:04 Right. oose that career path?
Geri 2:48 I think there was something there when I was a kid. I was the kinda kid who would take things apart, but not necessarily put them together. A tinkerer. Yeah. The so so I definitely like to take things apart. Getting into security was relatively spontaneous
Jo 3:04 Right.
Geri 3:04 During my studies. So at the time in Hungary, where origin come from, you had, like, the five years university degree. And for me, that was three years, like, basics, let's say, and then two years of specialization. And, originally, I wanted to do robotics, but then I realized that's a lot of math. Right. So so I quickly looked into alternatives, and then that kinda you know, security caught me.
Jo 3:30 I didn't really figure out at the time why, but it definitely pulled me into I didn't really dodge the mathematics part because we did get into encryption. And so I did Kiekonate RSA on paper,
Geri 3:43 which is not a fun thing to do.
Jo 3:45 Not twenty forty is better.
Geri 3:48 But, yeah, since then, I'm working in security. I first worked in quality assurance for a security vendor. Then I wanted to go more in the direction of security instead of generic software. So that's when I started to do penetration testing or ethical hacking.
Jo 4:05 Yeah. Yeah. Yeah. Dive quite deep at that.
Geri 4:07 Yeah. Actually, I I like to do that. I kinda the challenging technical parts kinda attracts. So I went to the pen testing, which I did a long time, mostly at Siemens in Germany. Mhmm. I also started to work more on offensive security research. And then in 2021, when I switched to Fortinet, that's when I switched to the defense side. And, And originally, I again wanted to go deeper in a technical topic, so I wanted to do more reverse engineering. And, you know, malware reverse engineering is the more interesting use case to do that. So that's why I switched to Fortinet. And now my job is a mixture of technical research in reverse engineering and threat intelligence, and then the less typical part of talking about all these things on conferences or supporting my colleagues with threat intelligence or delivering threat intel briefs to customers and parsers.
Jo 5:01 And I imagine it helps that you've seen sort of the different sides of the equation. Yeah.
Geri 5:05 I think You
Jo 5:06 can take in those different terms.
Geri 5:08 Absolutely. The whole idea that I wanted to switch to the defense side is to become more, let's say, rounded as a security professional. So about just to have the view of the attacker that I usually had when I was doing pet testing, but to understand exactly how defenses work and how we can build better defenses. Because pet testing is fun, but at the end of the day, what we want is to build better security. Right. And that's much better achievable on the defense side. have the view of the attacker that I usually had when I was doing pet testing, but to understand exactly how defenses work and how we can build better defenses. Because pet testing is fun, but at the end of the day, what we want is to build better security. Right. And that's much better achievable on the defense side.
Jo 5:37 So looking at the challenge from all those different viewpoints and given your career spanning over a decade, what would you say are current trends that really stand out, um, that you see in your daily work and how, obviously, you have used to say AI to protect AI to attack. Right? So it's that's all around the use of AI and and how that changed the game a little bit. I'm curious to hear your thoughts on that.
Geri 6:05 Regarding the trend, I would say we kinda call it the industrialization of cyberattacks. Were you here? So at the beginning, we had, like, individual hackers doing bad things, often not deepened in touch and just for the sake of it.
Jo 6:20 Yeah. Yeah. Yeah. Yeah. The script kitty in the in the basement, the booty in yeah. Yeah. Yeah.
Geri 6:25 But then as it became profitable to do hacking, it became a business. Mhmm. And we're clearly saw in that instead of, like, the one hacker breaking in, stealing the data, stealing selling the data, all these responsibilities split between teams. So now we have initial access brokers who only focus on getting the excise, but then sell it, but they don't do the ransomware deployment or any of the rest of the kill chain. We have, let's say, ransomware authors who or ransomware as a service providers whose focus is to write them all by itself and provide the additional services like the negotiation and all these kind of things, uh, around the ransomware operation. But they are not the team who is gonna use by his support.
Jo 7:08 Not the one that So
Geri 7:09 they sell that as well. So there are all these different teams. And
Jo 7:12 So it's a supply chain in the end.
Geri 7:14 Exactly. And it became because of that, it became much easier to become a cybercriminal because you don't have to do all this, I think, yourself. Right. You can buy the excess from an inch initial access broker. You can buy your ransomware from ransomware as a service provider. They even do the negotiation for you. You just log in, do a little bit of hacking to get around and then decode ransomware, and then basically, you're through with your part of the chain.
Jo 7:37 Mhmm.
Geri 7:37 So this was the previous step. But now where we're going from this is the what we call the industrialization, which means that it's being scaled up, especially like cybercrime. It becomes very opportunistic, automated, and because of that, happens on a much higher scale
Haon 7:57 Right.
Geri 7:57 And much faster than we were used to, let's say, five years ago. And, obviously, AI has a big part in this.
Jo 8:03 Yeah. I look I'm imagining some of those effects wouldn't have been possible. Attacks wouldn't have been possible. f that, happens on a much higher scale
Haon 7:57 Right.
Geri 7:57 And much faster than we were used to, let's say, five years ago. And, obviously, AI has a big part in this.
Jo 8:03 Yeah. I look I'm imagining some of those effects wouldn't have been possible. Attacks wouldn't have been possible.
Geri 8:10 Exactly. Before AI. And, like Exactly. The the whole automation part is partially because of the capabilities that the AI provides to the threat actors as well. And, generally, like, I was talking about, for instance, whether a malware is AI generated or written by a human, at the end of the day, it doesn't matter from a detection point by you because it's the same malware, let's say. It will be interesting once AI will be capable of innovating, so creating new techniques. Variants out of something that's Exactly. New things, new tactics, then it will be different. But now it's basically creating the same type of malware, but where it differs is that they can do this much faster because they don't have to sit down and write the hardware, but they can do it with some kind of AI coding tool. So speed is definitely something that's increased because of the AI capabilities.
Jo 9:06 So, I mean, besides the bad guys being able to use those tools, I mean, obviously, they're also available to ethical hackers, but then also available for defensive activities. Right? So how has that changed sort of the job of the defender, so to speak, and how are, you know, ethical hackers using those tools as well?
Geri 9:27 I mean, if we compare the two sides, the attackers using AI and defenders using AI, then on the basis, the capabilities are the same. So we have similar opportunities on both sides, uh, to use AI. A good example for that is vulnerability research. So, you know, vulnerability research to find security bugs in software, that's a very difficult thing to do. So it's not something that many people can do in my job. Yeah. And now it's a very specific skill. And now we see that many AI systems are capable of doing this. So the first idea people have is that, oh, this is gonna be such a game changer for attackers because they can look for zero days, and they will be exploiting zero days left and right, which is true to some degree, but we have the same opportunity on the defense side. And so this is also a huge opportunity for vendors to Up their game as well. Similarly use these AI tool ways to find those vulnerabilities in their software before attackers have the chance to, uh, exploit. Go outright. Yeah.
Jo 10:33 Yeah. Yeah. Yeah. gree, but we have the same opportunity on the defense side. And so this is also a huge opportunity for vendors to Up their game as well. Similarly use these AI tool ways to find those vulnerabilities in their software before attackers have the chance to, uh, exploit. Go outright. Yeah. Yeah. Yeah. Yeah.
Geri 10:34 So that's just an example of how it's relatively balanced. Right. On the attacker side, there are a couple of things where they have the advantage. I think one key point is that that they don't have ethical and moral limitations. Like, they don't have Artificial guardrails. Yeah. They don't have to think about how this is gonna affect society and employee burnout or whatnot. There is a new technology they can deploy immediately. Right. Another benefit or advantage for the attackers is social engineering. So, obviously, this was the first place where when GPGP came out, it was immediately used in phishing attacks, which is I find it kinda embarrassing for hackers that they needed artificial intelligence To do that. To provide premetically correct emails. Right.
Haon 11:19 Right.
Geri 11:19 But, you know, now they have it, so they can they can mail it. And then we have all the different kinds of deep fakes that we see, whether it's some kind of fraud, you know, like colleague of the CFO as the CEO or pretending to be the CEO to tell them to transfer money somewhere or the whole North Korean IT worker scheme where they do the whole recruitment process and then also team meetings later on when they are already at the organization with deep fakes and things like that. So AI for social engineering is definitely something that is useful for very useful for the attackers, but not too useful on the defense side. And on the defense side, I think the biggest advantage is that we have a lot of data. So AI lives off of data. So it basically, it's data for breakfast. And we have a lot of data on the defense side, whether it's your security operations with all the logs you collected. You know, if somebody runs the security operations and they look at the bill that they are paying for story, all that thing Yeah. All those data, they know that they have aligned data. And also in network operations, basically, the whole detection and monitoring part is about collecting data and analyzing data. Yeah. Yeah. Yeah. So once we start to operationalize the data with AI, that's a huge opportunity for defenders to have an advantage there.
Jo 12:40 Yeah. And so a lot of that data that you mentioned also gets used to generate threat reports. I think you're doing, like, a yearly threat report, you know, leveraging that data that gets collected over time. So do you see any new patterns in that data? Like, is there anything that stands out over the last few years, you know, in that threat report where you go like, oh, we've never seen a spike here before? Why is that? ou're doing, like, a yearly threat report, you know, leveraging that data that gets collected over time. So do you see any new patterns in that data? Like, is there anything that stands out over the last few years, you know, in that threat report where you go like, oh, we've never seen a spike here before? Why is that?
Geri 13:05 I mean, on on a global level, it's not that interesting. What I usually do is when I talk to either present on conferences or talk to customers in a specific industry or in a specific region, like, let's say, in a country. Mhmm. Like, let's say, we are in Spain. If we look at the telemetry for Spain or, you know, we are the mobile word contrast when we look at the telemetry for telco, that's where we can get things. So We're able
Jo 13:32 to break it down more. Exactly. You do. Yeah. You know,
Geri 13:34 I think with threat intelligence, it's really important that it's not just about collecting the data. Actually, the more we can filter out the noise, the better information we get. Mhmm. So let's say, often, when I go to a country, I look at the telemetry and I compare it to the global telemetry or the EMEA telemetry just to see the differences. Like, for instance, I look at, uh, brute force attacks, and we can see from our detection data that which protocols are brute force more on an EMEA level. But then when we go deeper into a specific country, then we get a total different pictures. So that shows us that if you are operating in that country, then you should focus on those protocols that are in the top three, let's say, when you want to do brute force detections.
Jo 14:24 You you can make much more informed decisions on where
Geri 14:28 Exactly.
Jo 14:28 Your your defensive capabilities need to be improved. Yeah. Yeah.
Geri 14:33 And, generally, to threat intelligence, a very important aspect is that it's about answering questions. So we need to ask the right questions. And the right questions might not necessarily come from a threat intelligence team that could come from business, that could come from leadership. And, basically, the threat intelligence team is there to provide meaningful answers and evidence. Exactly. Evidence based answers to those particular emotions, but these should be well formulated as specific questions as we can because that's how it becomes actionable.
Jo 15:05 Yeah. Yeah. Yeah. So speaking of, I think, different verticals and different markets, you personally, but then also Fortinet as a company, obviously, you have a lot of experience also in your operational technology. There's a lot said about IT, OT convergence and all these things. I'm curious to hear your take on it. Like, what's special about OT and how different are maybe those attack vectors that you see emerging over time? also Fortinet as a company, obviously, you have a lot of experience also in your operational technology. There's a lot said about IT, OT convergence and all these things. I'm curious to hear your take on it. Like, what's special about OT and how different are maybe those attack vectors that you see emerging over time?
Geri 15:31 Yeah. So OT is challenging for various reasons. So, historically, it kinda grew separately from IT. Right? And you can even see that in the leg wave. So when I started to work with OT and I started to talk to OT people, I realized that we don't really understand each other. Things are called differently. It's a computer, but they call it HMI, so human machine interface, but it's still a computer or PC. So there is this human level differences as well. So it's challenging to communicate, and and the priorities are very different. Yeah. So in IT, the crown jewel is the data. That's why kinda ransomware is the perfect crime against IT because you're stealing the crown jewel and sending it back to the original owner. And in OT, the crown jewel is the operations, so the availability. It doesn't mean that the branch number doesn't work there because, exactly, even though we call these machines differently, but they are still, like, normal PCs, they are Windows and Linux machines, many of them. So, you know, accommodative ransomware will work there as well.
Jo 16:34 Input comes in Yeah. If you would have pause, output gets delivered.
Geri 16:38 Yeah. Nobody will care what the data was that it was encrypted on that machine. But if you run a ransomware on a human machine interface, which is basically, you know, what you see the control center Right. On the wall. So if that goes dark, then you don't have visibility in their operations anymore, and everybody will shut down the operations just because they don't see what's happening. So it's a safety issue at this point. So what one challenge is the difference between IT and LT, the priorities between the two environments, but also how everything is handled. So in OT, you must be much more careful of what you're putting there. That's why OT operators have challenges building security in OT because they don't wanna put security products often in the environment because they are afraid that it will somehow influence the environment rightly. Like, let's say that there are all the environments where there's a a real time requirement Yeah. Which says that I think it's, like, four milliseconds that communication has to happen so that we can call it real time. And if it starts putting things in between, then you might Yeah.
Jo 17:45 It cascades through the system.
Geri 17:47 Exactly. You might you might delays. Yeah. Lose this real time capability. So that's one issue, but also that because security was not a priority for a long time, Noci.
Jo 17:58 Yeah. Yeah. A lot of those systems just weren't really connected to anything. 5] Jo: It cascades through the system.
Geri 17:47 Exactly. You might you might delays. Yeah. Lose this real time capability. So that's one issue, but also that because security was not a priority for a long time, Noci.
Jo 17:58 Yeah. Yeah. A lot of those systems just weren't really connected to anything.
Geri 18:02 They are not ready to deal with, you know, random data that could arrive to them. I was once on a conference, and there was a presentation about health care security. And the first thing the guy said is that never do a network scan in a hospital because there are all kinds of machines connected to the network, but also to patients, like, uh, insulin pump. Yeah. And it's gonna crash if you just send a random network back inside there. Yeah. So you have to be really careful. Right. And, uh, that that safety aspect that Exactly. Exactly. But it's also what kinda train the OT organizations to be very careful of manipulating the environment. And because of that, they have to be careful about what security bureaus they put there. Right.
Jo 18:44 So rule number one in IT cybersecurity is patches, patches, patches. Like, always stay up to date with everything. Right? And I hear with OT oftentimes the terms, like, the, you know, no patch sort of, uh, mindset alone.
Geri 18:58 Yeah. Absolutely. Um, you know, the the never touch a running system Right. Joke is absolutely dreary when it's dreary in OT. Yeah. But the good thing is you still have capabilities that you can do in OT. Like, something I really like is deception technology in OT because it's passive. So, basically, you put decoys in your network. And Is is a decoy the same thing as a honeypot? What's the difference? So honeypot is a more a sadistic version. Let's say it emulates a service to some degree, but it doesn't provide the full aspect. So when I talk about decoys, it could be a PLC, so a programmable logic controller that you
Jo 19:37 use in OT. A much more sophisticated than Yeah.
Geri 19:40 Exactly. Or you can create the decoy for your domain controller. Or I think what's very useful in OT is one really good point to catch attacks is the IT OT bridge because usually the threat actors are compromising the IT and that they are looking for a way in into the API, I must be. Exactly. Usually, there is some kind of jump post or something that allows that. So for one, you can really secure that machine. Like, you can lock it down just by the bus wall. There you can put EDR. There you can to process light listing, etcetera, But you also create a decoy bridge. So base the decoy jump post, you can put, like, lures in the IT network where if the attacker goes there,
Jo 20:25 then We need to guide them.
Geri 20:27 Guide, JT. Their in their, like, survey. Even you can make it easier for to get into that machine as well. Like, hey. With the password, there is it's easy to guess just to attract them to go into that.
Jo 20:37 Yeah. where if the attacker goes there, then We need to guide them.
Geri 20:27 Guide, JT. Their in their, like, survey. Even you can make it easier for to get into that machine as well. Like, hey. With the password, there is it's easy to guess just to attract them to go into that.
Jo 20:37 Yeah.
Geri 20:38 And the good thing with deception is that we say that there are no false positives because every interaction is a suspicious interaction. Like, nobody should, especially in OT. That's that's the advantage of OT that there shouldn't be random traffic in the OT environment. So You know something's fishy if Exactly. If something connects to your decoy Yeah. That there's a very, very high chance that it's malicious. And by entertaining the attackers on these decoys, you also collect threatening data traps because you see what they are trying to do, which can help your incident response in the remediation effort.
Jo 21:12 So then you layer on top of, you know, no patch and the need to be in a real time environment, and then you have devices that are in the field for a very long time.
Geri 21:23 I decided.
Jo 21:24 And then they're extremely resource constrained. And so, you know, that's sort of the bundle of challenges that you get with OT. And I'm curious to hear your perspective as well on those long lived devices. Obviously, you know, a lot of them are already in use for ten years and have cryptography baked in. So do you see a lot of challenges arising from that as well where people are concerned about using those algorithms that are ten years old, twenty years old?
Geri 21:53 For one, we have devices where there's no cryptography built in. Right? So I guess that's the first time that every everything's open. No. But you're right. So as we are reaching or we are advancing the capabilities of quantum computing, this is gonna be definitely a challenge because, you know, it's not something you need to fix when the attackers have the capability to decrypt. Because at that time, they probably already already have the data they want to decrypt. So the data that they are gonna decrypt with their quantum computers at some point is they already have that. Right? So they are already harvesting that data whether it's OT or IT. I think at this point, it doesn't matter, but they definitely collect the data. And once they have the capability, they will do the decryption.
Jo 22:41 So that's the oftentimes mentioned harvest now, decrypt later. Exactly. Yeah. Yeah. Is that something that you see in your threat data as well? Like, has that always been the case, or has sort of that data harvesting, has that, you know, increased over the next period of time?
Geri 22:59 Unfortunately, I haven't looked at at this specifically, but data exfiltration did happen in various scenarios. So in espionage attacks where they are really after the data Right. I think that's where this is most relevant because they will extract the data if they have the capability or or
Jo 23:19 And, like, in the company secrets or government secrets. Right? ally, but data exfiltration did happen in various scenarios. So in espionage attacks where they are really after the data Right. I think that's where this is most relevant because they will extract the data if they have the capability or or And, like, in the company secrets or government secrets. Right?
Geri 23:25 Exactly. And Either it targeted When we talk about espionage, these are usually advanced persistence threats. So often nation states attackers,
Jo 23:34 you know So you saw the complete opposite of the script kiddie that
Geri 23:37 Exactly. Exactly. It's it's
Jo 23:38 Unlimited. But quite
Geri 23:39 it's in quite a position. You have budget, very Human resources. Oh, yeah. Very skilled attackers, very deep know how in everything around their attacks. And their main goal is to get in and stay under the radar as long as possible. And Stay undetected based. What they exfiltrate during their time, it's really up to them. So it's absolutely possible that they find data that they consider potentially useful, but they don't have the capability to to decrypt at this point. They might even look for, you know, decryption keys that are available. But if not, it might still make sense to take them out. Also note that since they are state sponsored actors, probably they will be the first who have access to quantum computers because it's, again, a part of the supply chain of state sponsored actors. Let's put it this way.
Jo 24:29 You you think there's gonna be a press release once they have?
Geri 24:35 Yes. Probably
Jo 24:35 not. Likely not. Yeah. So I think another really powerful tool that the defensive side is sharing of data, sharing of information. So how important from your point of view is that to sort of strengthen, you know, the good guys in that impossible job to protect for this sake?
Geri 24:55 It's absolutely important. And I'm personally but also in a fortunate way, big believers in threat intel giant share. So, actually, it's my team who's responsible for threat intel partnerships. So what that means is that we work with many organizations such as the Cyber Threat Alliance or the World Economic Forum to share threat intelligence and also work together on various projects. We support law enforcement such as Interpol or FBI and some local law enforcement as well when they have questions to their investigations. Unfortunately, that's usually a street, so they ask, but they don't share.
Jo 25:30 You don't learn anything from it?
Geri 25:33 At least not at at this point. So once the investigation is closed, then then we have the opportunity to learn about it. But, yeah, it's a bit different.
Jo 25:40 That's more of a postmortem than for you to et, so they ask, but they don't share. You don't learn anything from it?
Geri 25:33 At least not at at this point. So once the investigation is closed, then then we have the opportunity to learn about it. But, yeah, it's a bit different.
Jo 25:40 That's more of a postmortem than for you to
Geri 25:42 Exactly. But, yeah, it's understandable. But we also work together directly with, uh, national certs and national cybersecurity centers depending on how they are called in different countries because they also do a really good work. But, usually, they are very focused on their jurisdiction, the country they're operating in. So we try to support them with the intelligence that we create, but they are also often helpful to us because they also share what they see. Right. We can reach out if we learn about an attack in a specific country. We can reach out to
Jo 26:13 you at a real point.
Geri 26:14 Yeah. Get malware samples and these kind of things. So I think intelligence sharing is a key part of cyber defense, but luckily, this is very accepted in the threat intel community. So even if it goes to, you know, vendors that are competitors on a threat intelligence sharing level, we still like to work together because the idea is that we are not competing on this part. We are competing on who can build a better defense slice. But intelligence sharing is important because nobody has the full view of what's happening out there Yeah. Because companies have different angles of looking at Exactly. Based on their deployment, based on, you know, geographical limitations, etcetera. So only by sharing intelligence Yeah. Can we have all a better picture of what's happening out there.
Jo 27:00 The rising tide that lifts all the boats.
Geri 27:02 Yeah. Exactly.
Jo 27:04 Yeah. So looping back into AI and, you know, looking at the advancements of in just the last six to nine months, like, if you look ahead, do you think that the development and, you know, improvements in AI are gonna play more into the cards of the attackers, the defenders, and why?
Geri 27:25 So my current conclusion based on my research is that attackers have the advantage of speed. So in a short term, they have an advantage, as I mentioned, because they can deploy new innovations faster. But I believe in the long term, defenders have the advantage because of the available data and because we will be able to deploy AI in the whole defense in-depth structure. So it's not just one layer where we want to use AI, but we can really create the different layers of AI based security. So currently, how we define that is so Fortinet's strategy is that we define three main AI defense layers, the protect, assist, and secure AI. So protect is basically using AI technologies in protections. So this is not a new thing. Um, so before ChatJPT, nobody wants to hear about AI, but we were actually working on it.
Jo 28:25 Before it was called AI. There was ategy is that we define three main AI defense layers, the protect, assist, and secure AI. So protect is basically using AI technologies in protections. So this is not a new thing. Um, so before ChatJPT, nobody wants to hear about AI, but we were actually working on it. Before it was called AI. There was
Geri 28:27 machine learning and before that Exactly. So we started to work on AI technologies, like, twenty years ago. So we had machine learning for malware clustering. We had machine learning for web URL analysis. Then we have AI supported behavior analysis in the different products where we see behavior, whether it's on the network level in NDR or in an endpoint with EDR or see patterns that Exactly. Shift. We because we can find a baseline, and we can see anomalies in that. So we're using these technologies in the last twenty years. You know, we are enhancing that now with the newer AI technologies, but that's the main concept of AI protections to use AI technologies in bitway new type of protections. And now the assist layer is what generally we mean by AI right now, which is a generative AI or or agentic AI, where we have some kind of AI assistant in a security product. So we have that again for, you know, multiple purposes. Like, in the network operations, you can ask the AI assistant about where you have performance issues in your network, and the AI assistant then generate the new firewall configuration.
Jo 29:40 And That's gonna happen much faster Exactly. Doing it manually or doing for open ports or whatever. Exactly.
Geri 29:47 But similarly, in the security operations, you can have an AI assistant that can enrich the alert that is coming in or tell you Give more context around it. Exactly. Tell you which threat actors usually are using these techniques that you are seeing. And with the agent KI level, then you can also use automation together with AI and the human analyst to automatically analyze alerts, but also to respond to alerts depending on what you decide. So there are a lot of capabilities there. So that was the second layer, the assist. And then we have the third layer, which is the secure AI. Because what we haven't talked about yet, but it's very important, is AI systems are also becoming a significant attack surface. So especially because everybody is hurrying into deploying AI. So I just called the doctor, like, a few weeks ago and raised an AI assistant right there. So everybody is which is totally understandable. It's an amazing technology, and it really impacts every part of our lives. But we have to understand that all the AI systems that we are building and exposing is becoming an attack surface. And we've seen very complex AI systems hacked by researchers with an email or with a calendar invite. So there were various examples to interesting attacks. And the reason for that is because, generally, complexity is the enemy of security. Right. So the more complex system you build The more vulnerable it's gonna be. Yeah. Exactly. And when you take, let's say, a huge enterprise environment and you start connecting things with AI, then it's immediately becoming extremely complex. reason for that is because, generally, complexity is the enemy of security. Right. So the more complex system you build The more vulnerable it's gonna be. Yeah. Exactly. And when you take, let's say, a huge enterprise environment and you start connecting things with AI, then it's immediately becoming extremely complex.
Jo 31:37 All bets are off.
Geri 31:38 Yeah. Yeah. Based on fee. Yeah. Yeah. And test AI systems is also becoming a new skill that, for instance, penetration testers need to figure out. It's a very different way of hacking them than it was traditionally. So my point is that it's a huge attack surface, so we have to think about how we are protecting this attack surface. Often, whether an AI system is created by your organization or you buy it from somebody else, you don't really know what kind of security measures are there. So it's worth to think about how you can surround this system with security. And that's kind of the secure AI layer to provide these capabilities, whether it's about looking at the inputs that goes into an AI system to detect prompt injection or detect that your employees are trying to send confidential documents to public LMS or looking at how agents communicating with each other and to see that maybe somebody is trying to take over the control out of the agents and they are doing something weird. Or you can look at at the end, what is the output that is coming out from an AI system? Is it meaningful, or is it somehow influenced, or can you see any kind of malicious intent in it? So these are all very important questions, and these are what we try to insert with the secure AI layer of our AI strategy.
Jo 33:03 The three layers. Yeah. So do you see any blind spots? Because that sounds like you're addressing all variations of protecting AI, protecting with AI Mhmm. And then using AI to enrich your data as well. Is there anything that you would say blind spot or something that is underestimated that, you know, no one's really looking at yet in sort of, again, going forward the next, you know, maybe twelve, eighteen months?
Geri 33:31 So the first thing that came to my mind is prompt injection. I wouldn't say it's a blind spot because we know about it. But learning from the past, I think it's gonna be a really difficult thing to control. Yeah.
Jo 33:44 Yeah. When I was So history doesn't repeat itself, but it rhymes because if we had the the SQL injections,
Geri 33:51 Exactly my point. So I don't know how technically what I go in in this topic. So since I was doing pen testing, like cross site scripting, which is a very typical VAP attack Yeah. Or a SQL injection, which is also very typical attack, These techniques are have been known for at among thirty years. So for well now, very long time. But thirty years ago, it took us a
Jo 34:13 while to figure it out.
Geri 34:14 Yeah. But even today, there are so many systems still vulnerable to these same vulnerabilities because it's just so easy to overlook them. chniques are have been known for at among thirty years. So for well now, very long time. But thirty years ago, it took us a
Jo 34:13 while to figure it out.
Geri 34:14 Yeah. But even today, there are so many systems still vulnerable to these same vulnerabilities because it's just so easy to overlook them.
Jo 34:24 Or to have, like, unpatched systems or I mean, you were talking about, you know, hospitals and, you know, famously known for using old technology Yeah. Exactly. In in quite crucial environments. Right?
Geri 34:39 So I'm afraid a little bit that prompt injection is gonna be a similar thing that we know about it, but still it's gonna be there for a long time even when we have the capabilities to detect it, which is still it's more difficult than with traditional attack because it's really depends on how the AI system interprets the input that it's getting.
Jo 35:01 That's quite siloed.
Geri 35:02 Exactly. So it's a difficult thing to detect and especially every aspect of it because we also have to be aware of the inputs that are going into an AI system, which could be basically anything. So it's not just the prompt that you are sending to the AI assistant, but also the different documents that these AI agents were are gonna pull together and try to understand it to make the decisions.
Jo 35:29 It was that that long tail of
Geri 35:31 Yeah. In any of them, there could be some kind of malicious input. Right. But probably the bigger challenge is gonna be the people who just don't look into it. Yeah. Right? So even maybe we'll have the know how and the capability Yeah. There will be organizations, whether it's vendors or users or whoever, who will be just not aware of the problem.
Jo 35:51 So so the biggest threat vector is is gonna remain the human?
Geri 35:56 In the end. Yeah. To some degree. Yes. Yeah. So that's exactly it's not gonna go away.
Jo 36:01 Right. Yeah. Right. Right. So where do you send people that wanna learn a bit more about your work or, you know, find out about the threat report, send me, uh, any place you wanna point people to?
Geri 36:11 Yeah. So I'm relatively active on LinkedIn. So you can find me on LinkedIn as, uh, Geri. So just search for me. Since it's a Hungarian name, there are not so many people up, uh, on LinkedIn, so you will find me. And, you know, wherever something comes out, whether it's my personal take on something or whether it's, like, the transcribe report that we write, then I will post it on LinkedIn. At Fortinet, there's a fortiguard.com. Mhmm. So that's where we post outbreak alerts and different threat intel outcomes that we produce. They can also see some of the telemetry data that we publish. And if you wanna go into more in-depth analysis, then on the Fortinet blog, there's a friendly research part where the different teams from FortiGuard often document deep dives into attacks. Fantastic.
Jo 37:05 Well, thanks very much for joining us today, Gary. It's been a fascinating conversation. Thanks for sharing your insights and your knowledge. Appreciate it.
Geri 37:13 Thanks very much. ndly research part where the different teams from FortiGuard often document deep dives into attacks. Fantastic.
Jo 37:05 Well, thanks very much for joining us today, Gary. It's been a fascinating conversation. Thanks for sharing your insights and your knowledge. Appreciate it.
Geri 37:13 Thanks very much.
Jo 37:14 Yeah. Enjoy the next few days at Mobile World Congress.
Geri 37:17 Yep. Will do. Thanks a lot. Thank you.
Jo 37:20 Welcome to Shielded. I'm your host, Joe Linson. And today, we're joined by Han Park, cofounder and CTO at AIM Intelligence. Han leads a team focused on securing AI agents and enterprise AI systems to automated red teaming, policy driven guardrails, and continuous risk testing across multimodal AI. Perfect. Very nice to meet you, Han. Thanks for joining us to meet you
Haon 37:43 as well.
Jo 37:44 So, obviously, it's day four here, MWC in Barcelona. How are you feeling? How was the week?
Haon 37:50 Yeah. It was a very busy week. Lot of meetings. There has been definitely a lot of interest in AI security and safety solution, especially with how fast the adoption of new agents and audio services has been. So, yeah, it has been a great couple days.
Jo 38:04 That's great. And then you said you have you're in two locations where you're Yes.
Haon 38:08 So we had two different booths. We had one booth with the four years from now. So we're one of the semifinalists for that. So we're the top 20 startups. We also had a booth with the LG U plus sector as well. Great.
Jo 38:20 So did you get a lot of traffic? Is it Yeah.
Haon 38:22 Since we have two different booths, I think times two, the traffic Yeah.
Jo 38:25 Yeah. Yeah.
Haon 38:25 Different locations. Yeah.
Geri 38:26 So And if I'm not mistaken,
Jo 38:28 you were brave enough to do, uh, live demos on your booth?
Haon 38:31 Yes. We did a lot of live demos. We did this for some big enterprises as well, and they were quite surprised with the results. So we will already be having some interesting leads.
Jo 38:39 That's great. That is great. So your work, Haun, obviously, sits right on the frontier. Like, you're pushing the boundary moving the boundary as AI safety and and security. So I'm curious, what first pulled you into that area? Like, where is it that caught your interest and and led to you founding, uh, AIM Intelligencer here at Turin? at is great. So your work, Haun, obviously, sits right on the frontier. Like, you're pushing the boundary moving the boundary as AI safety and and security. So I'm curious, what first pulled you into that area? Like, where is it that caught your interest and and led to you founding, uh, AIM Intelligencer here at Turin?
Haon 38:58 Yeah. So it's been about a year and a half since we founded our company. But before that, my background was in Red Hatcher as the first models begin coming up, like TouchPT 3.5. I was doing my mandatory military service in Korea, so I had a lot of free time. And I've got really deep into red tuning these models, discovering a lot of different vulnerabilities. And when I finished my military service, there was this big AI hacking challenge, which I participated, and that's where I met our current CEO and cofounder. And we both received first and second place. So we, you know, discussed some of these ideas we had. And our CEO, he had an idea that red teaming hacking should be automated into an actual technical solution. And this was a perspective that I haven't thought before because I was I was doing all of this manually. So we realized this is going to be a very big issue in the near future and decided to build a enterprise grade solution for this.
Jo 39:55 Great. Yeah. So I guess your background in research really informed your approach to that whole endeavor as well. Yeah. So AIM intelligence, you position yourself as the security layer for AI across industries, really. It's not for a specific vertical. So what does securing AI agents actually need for someone who's, you know, maybe just starting the AI journey or just wading into the water? approach to that whole endeavor as well. Yeah. So AIM intelligence, you position yourself as the security layer for AI across industries, really. It's not for a specific vertical. So what does securing AI agents actually need for someone who's, you know, maybe just starting the AI journey or just wading into the water?
Haon 40:20 Yeah. So there are different forms of how companies are using AI. Some of them are just the foundation model. They're building a foundation model. Or some of these companies are applying these AI models into actual services. This could be in the form of chatbots or internal agents that have access to company data or for their employees to use. So each of these different use case comes with different vulnerabilities. For companies that are building foundation model, some of the biggest risk is people using these models for illegal purposes or unethical purposes or making it generate chemical and biological weapon synthesis or how to make homemade pipe bombs or how to grow minors on Discord. So these are some of the issues that the foundation model companies are very put a lot of resource into preventing. And when we move into companies using these models for services, every country and every company has different policies, different regulation, use different rules that they have to ensure that their, you know, service or chatbot is compliant with. And that's quite difficult to truly ensure that it is kind of under their control and is behaving the way they want it to do. So these are some problems that we're solving for them. And when we move into internal agents, when these agents have access to company database, for example, like a HR agent, then attacking these agents can lead to leakage of these datas. Or if the agent makes a mistake or hallucination, it can lead to modifying company data. So these are some of the issues internally as well. That that you see. So it's quite a wide range on that vulnerability solutions. And if
Jo 42:00 I'm not mistaken, you're also automating in such a way that you're using AI systems to attack other AI systems. Yes. So current sort of an an automated red red teaming. So how is that different from traditional cybersecurity penetration testing, which is also quite automated these days? So before we move into :00] Jo: I'm not mistaken, you're also automating in such a way that you're using AI systems to attack other AI systems. Yes. So current sort of an an automated red red teaming. So how is that different from traditional cybersecurity penetration testing, which is also quite automated these days? So before we move into
Haon 42:19 the cybersecurity sector to share a little bit about the automated AI red teaming side, When the first models began coming out, there were a lot of human red teamers, and these companies would hire these human hackers to perform, like, manual testing on their AI models or AI services. But the problem is a human red team hacker cannot do a full comprehensive test, say, and also takes a lot of money and resource and time. So there has to be a automated red teaming solution to find these vulnerabilities in a comprehensive way. So that's why we're building a red teaming solution that almost behaves like a real human Red Team hacker, but also has very high attack success rate. You can customize the goals and target. And since we have this attacker agent, you can deploy this at a massive scale, like have hundreds of these attacker agents at the same time attacking a target, which means we can find the vulnerabilities much faster and also in a more comprehensive broader spectrum.
Jo 43:15 More big and bigger scale and shorter time. Yeah. Yeah. So through that work, what are some of the most common failures that you find in those testing Yeah. Activities?
Haon 43:25 So when we perform a retinine, we actually claim we have 100% attack success rate on all of the currently available models. So what that means is all the models out currently are very vulnerable. We can make it still generate anything that we wanted to. Even if you, you know, try to add rules through system problems or try to come up with a guardrail solution, it's still very easy to bypass these guardrails. For example, Entropic is one of the leading company in AI safety. They launched some of the world's safest models, for example, Cloud Opus 4.6 couple weeks ago. But when they launched the model in just ten minutes, we were still very easily able to jailbreak the model. In ten minutes? In ten minutes, our launch, making it generate very dangerous chemical and biological weapon synthesis, which is the very topic Entropic has spent billions of dollars to prevent with. And these models are being applied to enterprise services or chatbots and agents, which means all the vulnerabilities are transferred indirectly. And lot of working is still very easy and possible among these services and models.
Jo 44:29 Now you just used the term guardrails. Mhmm. Can you explain that? Yes. Like, what e models are being applied to enterprise services or chatbots and agents, which means all the vulnerabilities are transferred indirectly. And lot of working is still very easy and possible among these services and models. Now you just used the term guardrails. Mhmm. Can you explain that? Yes. Like, what
Haon 44:34 what are the So guardrail is a solution to kind of prevent or detecting these adversary attacks and also controlling what the output of the chatbot models. There are different approaches to this. Sometimes it comes down to just fine tuning the model with specific curated safety data. These are some of what some of the model provider companies do. Another solution is to have a external guardrail or external model that is checking the input data and the output data to see if there's any attacks or any data leakage or private information leakage. From those two different options, we did a latter one where we have a external third party guardrail solution that we provide to our customers and clients. And having a external guardrail makes it easy to customize it without having to touch what's in the inside of your model because all you have to do is modify the guardrail to incorporate the hardware generator dependent. Yes. And also easily upgrade the plane because there's no 100% guardrail currently. The best guardrail is a guardrail that can be improved over time.
Jo 45:40 So with all of this, I would imagine that partnerships are really important. Yes. So maybe you can speak a little bit about the partnerships that you do have, that you're working on, targeting as well.
Haon 45:49 Yeah. So we work with a lot of these big tech companies in US as well. Since we have this very high tax success rate RedTune tools, we work with OpenAI to test their guardrails before they launched three months ago, you know, discovering these vulnerabilities, edge cases so that they can improve it before launch. We also test them models with Entropic before they launch, team actual models in more sensitive topics like chemical and biological weapons. These are the topics that they're very wary of. And for providing our guardrails, we usually partner with more of the traditional cybersecurity companies because they already have this client base and they have the pipeline to deploy the solutions on premise. And since our guardrail also has to be deployed on premises, it has to be a good partnership with us.
Jo 46:36 Yeah. That route to market
Haon 46:38 into did that sense? Another partnership that has been helpful was companies that are doing AI projects for enterprises, like building a chatbot system or HR chatbot. So we just add the security layer to that big project. So that's another way we have worked with.
Jo 46:53 So you you just bring that extra layer.
Haon 46:55 So I see that as our data and being project. Yeah. Yeah.
Jo 46:57 So I would imagine that
Haon 46:58 a lot of those guardrails, as you just laid out and explained,
Jo 46:59 are policy driven. Mhmm. AMO list. And if you automate, how do we enforce that, Jan? bring that extra layer.
Haon 46:55 So I see that as our data and being project. Yeah. Yeah.
Jo 46:57 So I would imagine that
Haon 46:58 a lot of those guardrails, as you just laid out and explained,
Jo 46:59 are policy driven. Mhmm. AMO list. And if you automate, how do we enforce that, Jan?
Haon 47:18 So as I mentioned before, every company from different industries have so many different and unique policies that they have to ensure and comply with. And, also, when we move into different regions and countries, the laws are all very different. Mhmm. So these enterprises need a way to kind of apply these policies to their to the guardrails or their AI service, and that is why we're providing a very highly customizable guardrail. So very easily with just natural language, they can just add these policies. And it's maybe, like, good and bad cases, add examples, and this is applied directly to our guardrail without having to go through any fine tuning or data curation. And once they have this set up, since we also have the RedTuning attack here, we run the RedTuning attacker to find all these policy that are actually being applied. And, of course, some edge cases will be discovered, and those edge cases can be directly updated to our guardra policy because we have a system called Autoforge where these edge cases will be incorporated into the guardra policies, like adding rules or combining rules or removing rules. And then going through this loop maybe a couple times will ensure that it's much more robust in ensuring all of these policies are kept while also ensuring no adversary attacks to any unwanted behavior is happening.
Jo 48:40 No. In that feedback loop, are there any humans involved, or is that all machine driven automate?
Haon 48:46 Yeah. So the individual process itself is automated, but, of course, there can be manual customization within that, let's say. You know, looking through, oh, this is another policy that I've had, and you just, you know, add the policy yourself, run the test again. But updating the policies, the right teaming itself, regardless of our all individual automated components.
Jo 49:06 Yeah. So one of the things that you mentioned is that cybersecurity traditional cybersecurity is a little different to AI risk and as, you know, the topics that are important in that context. Now what is the mindset shift that needs to happen for, say, SSO, you know, who needs to protect against these attacks that you're describing as well? ings that you mentioned is that cybersecurity traditional cybersecurity is a little different to AI risk and as, you know, the topics that are important in that context. Now what is the mindset shift that needs to happen for, say, SSO, you know, who needs to protect against these attacks that you're describing as well?
Haon 49:26 So what I think that is very interesting about AI security is if you look at traditional cybersecurity, one of the most vulnerable attack factor is the humans themselves. For example, like phishing emails or spam messages to get passwords or whatever, those were some of the most vulnerable endpoints for traditional cybersecurity. Social engineer. So. Yeah. Yeah. Yeah. But now with the age of AI, we are replacing those humans with AI agents or services. But before, we're able to keep the humans who made the mistakes or were attacked accountable, maybe by firing or disciplinary factors or, you know, legal or, you know, policy sport is managed by people. But women replace this with AI agents. You know, how are we going to keep those agents responsible for these attacks or accountable? And these agents are now given more and more authority, more data, more access, more autonomy, and the attacks are coming into these agents. Mhmm. And for traditional cybersecurity against attacks against the infrastructure, you can just list up these vulnerabilities and block them. But for the agents, they have to have these authorities and control to be able to do the actions they're supposed to do. So while maintaining that, you also have to prevent these attacks. Right. So it's a quite interesting landscape for AI security because even, like, social engineering is involved to manipulate these agents. Right. So there has to be a shift in the mindset, and I think even now, there are a lot of these news and vulnerabilities, incidents happening involving AI agents. It has definitely has to be considered by CISOs as well.
Jo 51:01 Yeah. And then I think the other mindset that needs revisiting, if not shifting, is that, you know, of course, everyone who talks about AI thinks chatbots first. Mhmm. I but, of course, there's other modes of right? There's video, audio Yes. Besides all the textual things as well. So how does that change the equation looking at all those different modes? dset that needs revisiting, if not shifting, is that, you know, of course, everyone who talks about AI thinks chatbots first. Mhmm. I but, of course, there's other modes of right? There's video, audio Yes. Besides all the textual things as well. So how does that change the equation looking at all those different modes?
Haon 51:22 Yeah. So we're already seeing very fast development of these multimodal AI systems that, for example, generates image or video or receives image and vision and audio as input. And we're incorporating these models into, for example, autonomous drones or vision based self driving cars. And since it receives this miniature audio as input, these could be different attack factors. So we do a lot of this research at AMI Television as well, where attacks can come in through the visual input. We had a hacking challenge to hack a VLM based self driving car, where the hackers were supposed to manipulate the input image that goes into the self driving car. And the first place with a very small black text in the image was enough to check the self generated car to drive to a group of people. And since it's very small, this can be placed on, like, T shirts or hats or billboards or in the military setting as well. For example, tricking the drone to non detect the soldiers or also even inject actions into the drone. To me, it fall out of the sky. So these are also more frontier areas, but right to me at guardrails for this situation also has to be very heavily considered.
Jo 52:35 Yeah. Well, and that's certainly one thing that I noticed, you know, walking the floor at MWDC. A lot more booths now showing autonomous systems. Yeah. This is robotic systems. There was one booth. They had fighting robots sighted off you if you managed to see that as well. So I wonder how do you control or put guardrails around autonomous systems that are also physically autonomous? Yeah. king the floor at MWDC. A lot more booths now showing autonomous systems. Yeah. This is robotic systems. There was one booth. They had fighting robots sighted off you if you managed to see that as well. So I wonder how do you control or put guardrails around autonomous systems that are also physically autonomous? Yeah.
Haon 53:00 So I think the ultimate challenge for AI safety is physical AI safety because it can actually impact the physical world and also cause physical damage to humans as well. So the way we are tackling that problem is physical AI also receives multimodal input. It receives the visual input, audio instructions or whatever. And those could be attack vectors, so we have to do the red team for those to find what vulnerabilities there are before and also add multimodal guardrails to be able to detect attacks from audio or image. So the input part is already quite covered. Mhmm. But the difficult part for physical AI is the output of this robot when these robots are doing autonomous action. Because to control the robot action, we have to understand what this autonomous robot is trying to do. Right. And that's quite difficult. There's two different ways for this. One way is to have a external view or camera of the current robot situation to kind of, in a live situation, know what the robot is doing. But the problem for that is you can only do that indoors. We can't do that outdoors or especially outside of the location. Yeah. Yeah. Yeah. So the other way is to do this within the robot using the current eo centric view of the robot, what it's currently seeing, and trying to identify what it's trying to do. But the downside of that is any action that is outside of a sphere, for example, a stabbing motion from kind of starting here, you won't be able to see its action until Right. You know, very, very happen. So the approach that we are doing at AM Intelligence is we have to be able to understand the raw robot motors and use that to interpret what the robot is trying to do.
Jo 54:41 And then what's gonna happen next?
Haon 54:42 Yeah. What's happening next? And also understand what it's trying to do right now. Because right now, most basically, I've received the instruction, but very soon in the near future, they're going to be doing stuff autonomously, like what you're seeing with Open Call agents. I think there was an article yesterday where somebody put a Open Call agent inside a physical body is doing some stuff by itself. So we're creating a new model that receives the raw motor values as a input and creates the action vetting to be able to kind of map out what it's trying to do. And we will add them to our VLM based guardrail to combine with what the robot is currently seeing and what the robot is doing currently to make a safe or unsafe decision comparing with the policy that is given. So that's how we tackle this problem. e action vetting to be able to kind of map out what it's trying to do. And we will add them to our VLM based guardrail to combine with what the robot is currently seeing and what the robot is doing currently to make a safe or unsafe decision comparing with the policy that is given. So that's how we tackle this problem.
Jo 55:28 So within your system, like, there's a certainly some type of authentication that needs to happen, authorization to make those policy changes. Mhmm. I wonder how those controls remain within your system as well. Is there, you know, human in the loop, I think, is a term that's often been used. And then I'm also curious if those authorization systems are using any type of cryptography, like, how you view that and if there's any, you know, Quantosafe thoughts that you're putting into your systems as well.
Haon 55:59 Authorization is quite a big issue for enterprises because, especially, when they're having their employees use AI servicer or AI agents, different departments or employees might have different access. So those also have to be quite customizing. So we provide that platform with their guardrail where you can kind of enforce different policies for different sectors or different employees. Rule based, the access, grocery, and use tax control. And, also, the person who's modifying these policies and controlling also has to have kinda specialized language to Couldn't ask you over the
Jo 56:34 years or you live up.
Haon 56:34 Yeah. So, of course, those issues, security issues is a big concern for us, and then we try to kind of stay up to date for what's the most latest, I guess, methodology to keep this access safe. So quantum safe methodology is quite interesting, and I think that it can definitely be applied to our solution as well. Yeah.
Jo 56:53 Makes sense. So, you know, looking ahead the next six, twelve, eighteen months from your point of view, from where you sit, what is the sort of the biggest challenge within all this, you know, big AI rollouts that maybe we're not quite aware of yet? What's your thought in that?
Haon 57:09 So kind of similarly for AI and monosome chatbots, there was a big incident maybe, like, last year. There was a big incident with people, like, committing suicide or, like, legal issues. And now early this year and late last year, there has been issues with agents coming up with, like, breaches, data leakage. Mhmm. And I think in the next year, we will see an incident with physical AI and some of these autonomous system. There's going to be a major security incident where maybe potentially physical harm to humans might caught Might happen. So these are something that we are trying to prepare beforehand to have the solution ready for when these autonomous systems actually roll out Yeah. To the real world. Once ends that occurs yeah. And it's moving very fast. So the AI world is changing very fast. New new functions are coming out every single day. So the security safety side has to work very hard to make sure we don't fall behind them. on ready for when these autonomous systems actually roll out Yeah. To the real world. Once ends that occurs yeah. And it's moving very fast. So the AI world is changing very fast. New new functions are coming out every single day. So the security safety side has to work very hard to make sure we don't fall behind them.
Jo 58:03 So what's your recommendation for an enterprise that wants to safely roll out AI today? Like, what are the first concrete steps you you recommended to?
Haon 58:12 So I think it's very important to know that there are these security issues. A lot of people in that event don't know about the issues. So in first knowing these issues is the first step, and they'd be able to solve these problems. And there are a lot of other companies like us that provide safety and security solutions, like consulting firms as well. No traditional cybersecurity companies also build solutions with us. So talking with some of those experts and trying to find what is the best solution for my current need, it think is will be the next step.
Jo 58:43 Great. And so for our listeners, if they wanna follow a bit more on your work, where's their best place? Where would you center?
Haon 58:49 So I guess you can reach me out on LinkedIn. You can search my name, Hong Park. And you can also find a contact page in our AM intelligence website as well. So those two will be the best message to methodology to reach out to me.
Jo 59:03 Fantastic. Well, thanks for taking the time. I know we're all busy during NWC, so I appreciate you sharing your insights with us, and have a good rest of the day, last day, and then safe trip out.
Haon 59:13 Thank you. Thanks, Hans. Yeah. Thank you.
Jo 59:16 The quantum era is already here. With the right tools, the right knowledge, and the right mindset, together, we can face it head on. To find out more about PQ Shield and how our cutting edge solutions can help you secure your data against the quantum threat, Visit pqshield.com. Make sure to search for Shielded, the last line of cyber defense, in Apple podcasts, Spotify, or anywhere else you get your podcasts. Don't forget to click subscribe so you never miss an episode. On behalf of the team here at PQShield, thanks for listening.