Transcript
We don't want to replace anything. We want to become faster, faster and more dynamic. We have to stop believing that only the big ones are hitting. Everyone is on target now. Through AI, attacks become faster, they are scaled higher. You could say they become more dynamic. And that's exactly how we have to become more dynamic in our defense. Hello and welcome to the Human Firewall Podcast. We are Gundi and Christian. And today we're going to do another deep dive into a very special topic that we would like to take a closer look at together. Today it's about dynamic defense. Because we see that behind cyberattacks there is increasingly more and more AI. This makes attacks faster, they become harder to detect and also incredibly flexible. And of course the question arises for us how we can use this technology to defend ourselves against these attacks. That sounds totally logical now. Attackers use AI, so do we. And in fact, we will also go in this direction today. But maybe we have to take a small step back beforehand. Because in our everyday life we are actually dealing with AI every day. Sometimes consciously or unconsciously. I think if we asked people, do you use AI? Then they would say no. The unconscious users of AI. I think we all use AI somewhere. In this respect, it is logical that we say we also need it in the field of cybersecurity. Excitingly, I believe that we have all come into contact with AI in relation to cyberattacks. Whether it is phishing mails that have been written, orchestrated or sent. Or even more complex stories. But one thing we have to be clear about. There are almost no more cyberattacks that are not AI-facilitated to a certain extent. Whether this is a very small bit now or whether they have very strong support through AI, I think is still a bit open. But, or we can discuss it. But from my point of view, attackers have been early adopters and heavy users at the moment. Exactly, yes. You said many attacks are supported by AI. And I would say that's even an understatement. We can even say that cybercriminals have usually perfected AI. So they are real heavy users, as you said. And we recently had Thomas Zollner with us on the podcast. He is Chief Information Security Officer and Global Head of Security at TDK Electronics. And he told us what is so dangerous for him about cyberattacks supported by AI. Let's listen to it. AI-driven cyberattacks. So that's the biggest risk for me at all. That the quality of the attacks is getting better and better. That the differences between a fake mail and a normal mail just disappear. You no longer know the difference. Here it is really about having technical solutions, maybe. But also to develop a human sensitivity. Is it real or is it not real? That will be very difficult in the near future. Even more. If we summarize it a bit. I think that describes very well what a lot of people out there are experiencing right now. Cyberattacks are becoming faster and faster through AI. More and more personalized. And also more scalable. That's a very exciting point for me. Also the last one right now. Scalability means, in the end, that the business case is getting better. That means I don't always have to attack TDK Electronics with 100,000 people. I can also attack Iron Ede or someone. Because if I can orchestrate that very quickly, then the business case is also good if I get little money out of it. And that means we have to stop believing that only the big ones hit. Everyone is now in the spotlight. And there is a second point that is also very important to me. What he also says. The possibility of detection is more difficult. Because the quality is simply increasing. I think we have already made an episode here where we discussed that the times of the right-click error in phishing mails or the seven fingers in the video call are slowly over. These things just get extremely good. And to undermine that in terms of numbers, whoever has read the State of Social Engineering has now taken out two or three numbers. 83% of the IT security managers there have experienced at least one case of AI-based social engineering in 2025. That's an increase of 33% compared to the previous year. 30% points, exactly. Exactly. And if we look at that in more detail, voice cloning is always an interesting case for me, because it really shows how well AI is becoming and how effective attacks can be. If we look at that, we see an increase from 16% in 2024 to 30% in 2025. So it has doubled what we have in these attacks, which were relatively exotic some time ago. And from what we see now that they are a little more widespread and that they also show a little bit how good AI usage is. So if we also look at how they are made, what is learned, what is imitated, what is orchestrated, what is then called, when it is called, what the content is, that is already relatively sophisticated and not without reason also relatively successful. Yes, totally. So we have two aspects there. One is that the threat is increasing. We are seeing an increase in the share of AI-supported attacks. But what I find a bit optimistic now, that was a survey of people in leadership who have seen it. That is, we also see a certain sensitization that the person is noticeable. So we are not completely blind to it. That's true for me too. If I want to pour a little water on your wine, if you look at it cleanly, then we now have the number of incidents that were recognized. That's true. So it could well be that we have a much higher dark number. We don't know how many are still falling through the cracks. Exactly. Without wanting to make the case now, there is more in it. For me, the crucial point is actually, there was a time when it was science fiction. Something like that will happen. But in the meantime, these are real cases that we are experiencing and that are getting better, that can be orchestrated incredibly quickly. When we think of tools that we also use here for the podcast, which sometimes take my nuzzles out of here, I believe that within three seconds, the industry is worth learning my voice. And I'm honest now, maybe you won't even notice when you listen to this, but from time to time I nuzzle so extremely that Gundi asks in the aftermath of the recording, what did you actually want to say? And then I pass that on to our producer and say, hey, that's what I wanted to say. And then I say that too. And I can't say exactly where my original voice ends up, where the clone starts. That's a great use case. But they are incredibly fast, incredibly good. And unfortunately, in addition to the good purpose, there is also the user who does not have the good interest behind it. And unfortunately, there is also this bad use, or bad use, how should I call it, harmful use, and it just goes incredibly fast. So three seconds would be like, I'm calling you and I'm wrongly connected. Then I would probably still have enough voice samples from you. Yes, totally right. And it's crazy how quickly this technology is getting better. Which may not quite work today. It can get really dangerous in a short time. So yes, to summarize again what we found. So the threat is getting more. We see more AI-based attacks. And that is also seen somewhere. And the BSI, the Federal Office for Security in Information Technology, has given an assessment of where this threat comes from and how it actually appears in the threat landscape. And I think they have observed something very interesting. Namely, that AI is primarily used for attacks that have always existed. So to simplify and accelerate classic, established techniques. So to scale up again. Now you can imagine a little bit of what used to be totally wide-ranging and unprecedented phishing, which somehow worked a little bit for everyone, but not really for anyone. The watering can somewhere. Exactly, the watering can. This is now being used for targeted spear phishing, but not on individual people, but because it is supported by AI, you can still use it on many people in a high-scaled way. For example, you can imagine, to give an example, that attackers use AI for research, to look at, for example, what is actually happening in the companies that we want to attack right now. Do they have projects that they are working on? Do they remove jobs? For example, something like that. Or is there anything in the social media that is a current topic right now? And then they build in such information in phishing emails to make them a bit more realistic. Of course, this can also have something to do with the writing style, that the writing style is somehow imitated by predecessors. And that can be used in a wide variety of ways today to carry out even more realistic attacks. And that is also interesting, not just via email, but via all kinds of channels, which can then also be combined with each other. That is an important keyword, from personalization to multi-channel attacks. And that's something I've almost become a victim of myself. I think I've already told this here. But I would like to unpack it again briefly. And I didn't even have it on the screen that it could happen to me in any way. I was at an event, and that's what I have more often today, and then someone wrote, hey, I heard the talk, I thought the talk was cool, and that's a nice feedback, thank you very much. And it went back and forth a bit. So two or three simple communications, I thought that was great, I would be happy if we had an exchange, then I said, sure, we can do that. And then there was a message, by the way, we have an event in our company and I would like to perform here too. And I'll send you my email for that. Then I got an email with the thank you again for this great talk. It was very comforting, which is also psychologically good, because you fell a bit into your defense barriers. The classic techniques, they stay the same. And to be honest, I was also receptive. It was nice to hear that. And then came the motto, we're planning this here in the company and I put a document on it where the information about this event is in it. The speaker info doc. And then I actually thought, oh, that's cool. And I didn't look at this email like a phishing email, like the classic email from strangers where you look at it skeptically, but it felt to me like the email from someone I already know. I've already talked to him. And he likes what I say, he's like me, he thinks the same way. So I know that all this is unrealistic and probably now for the neutral listener thinks, oh my God, it's for Christian. Who is that exactly? To adapt my answer well. And then I noticed that he only had seven contacts at the time. One was me. And also from the position, there was no one who invites for an event. And I had a colleague look at it at the time. At this point, greetings to Lukas. And he thought, are you crazy? No way. I reported it and then it came directly to me, for God's sake. And the exciting thing for me then was, I passed it on in an internal channel. And in two different or three different languages. So obviously it was also a bot who basically read out, someone presents somewhere. And usually when you're at a big fair, you don't even know if the person who answers you was really a viewer or listener or not. And that's why this multi-channel story was so, yes, almost successful for me. Because it made the second email look like an email from someone you know. Yes. So there is an ideal contact that somehow suggests that it works. And if we now think about topics like voice cloning, where top-down processes in the brain, where we can prime it a bit, that is, if I write you a message, Kundi, I'm your CFO, I'll get in touch right away. Then you will expect to hear the CFO on the phone. And even if the cloning is not perfect, this pre-delivered message will make the overall structure more plausible for you and help you hear it more. And now you can say, for God's sake, we have to give up because AI is superpowerful and does it so well. That's definitely not the case, otherwise we wouldn't be on the podcast here. On the contrary, you can protect yourself very well from it. How do you protect yourself? That is the psychological factor again, which we will discuss in a moment. But I would like to give a short example of what impressed me very much. And that was also from Thomas Zollner when he was here on the podcast. He told a story about an employee who was attacked with such a CFO fraud and caused a learning situation from the attack. And let's listen in for a moment. Thomas Zollner had an employee in Spain who was attacked in such a way that I assumed he was our president of the GDK. Okay. She recorded it. She was there so quickly that she said, please call us back on the fixed network. Yes. And she called and recorded the phone call. And so we were able to present it to our employees and show how to react properly. So it also showed that it is already well implemented in our company, I would say. Our employees are already aware of this. Yes, totally impressive. First of all, to switch so quickly and then also to have that in mind. Hey, okay, I can record that. Would you have had that? I would never have come up with the idea. I don't think so. I would have been incredibly proud to recognize that at all. And then to make this additional jump and to say, I'm going to use this now, I'm going to turn the stick around and show my colleagues what is actually possible there. So incredibly great. Great story that also shows that we can really tackle this human risk topic to human strength, human firewall. Totally. Impressive. Yes, exactly. So deepfakes and voice cloning are an application case for a very specific attack. But if you look at it on a higher level, then you can see that cybercriminal AI still uses it in a different way. Namely, to basically automate attacks. And I think that's very, very scary, I have to say, a little bit. There was a case in September last year, so 2025, which has become very famous, where cybercriminal AI was used to automatically attack more than 30 organizations. You can imagine it that way. Actually, such an AI has security guidelines that should prevent this. The cybercriminals have these security guidelines and have managed to automate 80 to 90 percent of their cyber attack. That means they had to control something again and again in between. But so much of it can be done completely autonomously by the AI. And if you just take that as an example, which is already possible now, then you can't imagine how much cyberattacks can be scaled up in the future and how many companies can be attacked by relatively few people at the same time. I would like to add that it doesn't need the geniuses anymore, the experts. I remember the Amazon hack in the Times of India, where it turned out that the attacker was not so incredibly skilled in IT, but simply used various AI tools and orchestrated the attack with them. I found the headline very exciting, and with the result you get a fried over it because you are not so capable. But that shows that it doesn't need the big resources anymore, if I can scale it via the tools, not the great expertise. And, and that is also important, that I can attack incredibly fast when I have automated it. That means we sometimes see the speed of attacks in a kind of double way to become more efficient. I need fewer people and these people also need to be able to do less. So, of course, the question arises directly, how can we actually protect ourselves from this? So you said the threat has changed through AI. Yes. And that means for us that we also have to change our defense. You can say they become more dynamic and in the same way we also have to become more dynamic in our defense and use AI dynamically. I think that's a very, very big point. You introduced when the attackers use AI, how can we use AI to be able to defend ourselves better? And maybe we can get a little more concrete there. If we think about how such an attack usually works, then we see, if we have an attack that happens in a larger scale, we notice that an e-mail comes in several times and we alarm the SOC and the alarm sirens go up, the pulse rises, then we need a few hours, then we have blocked the e-mail, flagged the domain and say, okay, incident closed. But attackers have sometimes also automated, as soon as they recognize that something is happening, they can adapt to the old tactic and can continue to attack. That is, we are simply in a mode where we can hardly close it. If we say now, hey, we want to do a training so that people understand what is happening there and protect themselves, then, classically, after such a, if we have just recognized that there is the e-mail, then it takes about three to ten days on average to build an IT team or have a 20-man team. I have to pull people off activities, at least against myself, to build that and we have a average survey that says that after the attack it takes about 19 days, i.e. between two and three weeks somewhere, that employees really trained that. That has the psychological disadvantage, we live in a time of such information overload and waste. That is, if that happened 19 days ago and I train something now, then it is still such a vague memory, but dissolved by so many other stories. That means, on the one hand, that the training is not so effective, the relevance of the user is not so high, there is not so much hanging and then I can think about how big the behavioral change really is. And there is a second point to it, the classic, the attack happens and then the management comes and we say, hey, we're just planning the training. It would be something completely different, and I think we'll talk about it later, if we were to be able to react immediately. If we can say AI-based, hey, we're just seeing this e-mail coming in and that's exactly what we're doing training, and now. That is, we save all the big resources, we save the 19 days in between and if the management knocks, we can say, hey, we've already done the training, look at your mailbox, there's something there. We can accelerate in this area. And that means, on the one hand, the real-time danger detection by AI. So we have to detect in real time, just improve detection, what's happening, so that the time window for the attack is reduced. From my point of view, it needs this dynamic defensive ability, so I have to be able to detect as quickly as possible, also react much faster and say, now I have the training, now I'm doing something. That means it's about automation and also, in the end, for me, it's also about learning and adapting. That means, if attacks, as we are just developing, we have to do that too. And with the help of AI, security systems can learn new data, build new defense strategies and become more adaptive. And then I think at the end of the day, we have a kind of, I don't want to say weapon equality, but then we have at least, then we are also the heavy user, like the attackers. AI is not a replacement for things, but simply a support. And then I think we are very well positioned. Yes, totally. So it's basically like a race, where our opponents, the cybercriminals, now have a new support. And if we don't use them ourselves, then of course we lag behind very quickly. Exactly. So of course it all sounds great when you think about it like that. Okay, we're going to use AI now, we're going to be super fast, we're going to automate everything. But in practice, of course, that also has certain challenges that we don't want to hide now. So of course, you first think about that the data quality with which an AI is trained must always be correct. So you know the rule, right? If I put garbage in somewhere, garbage comes out somewhere. And the same goes for AI. That means we have to have AI often like a black box where you can't really look inside. And if I use AI to make decisions, especially for defense applications, then it should be transparent. That means I should know how a decision is made to be able to understand what is happening. And best of all, of course, or actually super important, always have a person in the loop. You just said that we don't want to replace anything. We just want to get faster, faster and more dynamic in order to grow from this flood of attacks. And of course, there are always technical problems. So is an AI system compatible with what I've already set up? That's always a difficulty to integrate into already existing security systems. And this compatibility is not only a technical one for me, but also an internal one. Does the AI fit how it works in my idea of security? Does that work? I believe that this dynamic defense is a very good strategy in terms of cybersecurity. But, and I don't want to underline that at this point, there are still people who are still the decisive link in the chain for me. If we also look at the numbers, they have remained relatively constant and the numbers have still run by the factor of human somewhere. If you're listening right now, could you explain to me why this fluctuation of 20% in the reports is created between Verizon Breach Report and Allianz and so on? I would be very interested in that. My hypothesis is that the perpetrators are not listed in the Verizon Breach Report. But at least there is a certain fluctuation. Nevertheless, and that is the exciting thing for me, there are totally exciting things about this attack and how quickly you can react. I took a look at this Recreate the Attack feature and find it incredibly interesting. There are a whole range of use cases for it. I can basically just do a screenshot of the current phishing mail, load it into the system and immediately have the training, i.e. the phishing simulation. That means instead of 19 days, I need 3, 4, 5 minutes and can react. There is another use case for this. He uses it, if he sees a real mail that has been stored somewhere in another company, then he packs it into the system to see how high the click rate is with us, to figure out how sure we are. Do we have to do more? Are we good? And to have a kind of real danger point. That is really an email. It was successful here and there. And with us, it may not be clicked anymore in the context of cyber security, if you don't devalue the toy, but rather in a positive direction. It's a measuring tool, actually. I think it's great. And the use cases go on for me. I can say in the supply chain when someone is attacked, hey, I'll react right away and make myself safe here. Long story short, I think the use of AI would not only allow us to have this weapon equality with the attackers, but also with cyber security. And I also believe that it is not unimportant for the perception of cyber security that we show people, we are fast, we are modern, we can do that too. That is, for example, when the management comes and asks when we react to these threats, we can say, we have been doing this for 10 days or so. I think it's a very good conversation to have. Totally. We don't just make a statement when something happens, but also say, hey, that's a big problem right now, but we're a little ahead of the times. We're already dealing with it. Yes. Wow, that was mega exciting. We talked about a lot of topics. Let's just briefly summarize what we found out and what is so important. Exactly. We said, okay, AI makes cyber attacks more often, so it's just a threat that we have to adapt to. And best of all, by developing a dynamic defense, by using AI for cyber defense ourselves. Exactly. But of course, that does not replace the sensitization of our people. But it can improve or make it even more effective, for example, via this Recreate Attack feature, where you can use AI to create new simulation emails. Did I summarize that well? Do you have anything else to add? No, it was almost as if an AI had summarized it. I think the important point that we have in there, adaptive defense, or this dynamic defense, is ultimately something that I think many will already do, to whatever degree, like attackers to a certain degree use AI or not. I think what we tried to do today is to keep in mind that AI is not invulnerable, quite the opposite, AI is sometimes designed to make mistakes. That means it's about using the technology, but also about my understanding of how I can use it well to be safe. And that also shows that, in my opinion, it will continue to be the most exciting field in every company, because you have to be ahead of time in terms of office. So I don't think there is an area where you can be ahead of time and become even faster. That's why it's still exciting for me and it won't be the last episode, I think, that we have on this topic. I think so too, yes. Then there is not much more to say at this point than thank you very much for being here today. We hope you enjoyed the topic. If so, please leave us a five-star rating. We are happy if you also subscribe and tell people that we exist, so that we can continue to get these messages out. Thank you very much.
