Very much using this first. Okay. Thank you very much. And thank you very much Alexei for that perfect handover, basically, because that's what I'm gonna focus on.
Like, how is AI gonna be utilized or how can AI gonna be utilized when not protecting us, but more attacking us. And especially in the focus of social engineering and fishing at techs, Alexei said it already, it's more like a hypothetical yeah. Thought experiment. But on the other hand, we also have some, some very solid cases or at least one solid case from this summer where AI was most likely used in order to conduct a social engineering attack. My name is Nicholas Salman. I'm one of the managing directors of so safe and so safe.
We are like totally focusing on the human side of cybersecurity, putting the human in the middle of everything and trying to, yeah.
So to say, increase the human intelligence before. Yeah. Before the artificial intelligence is gonna kick in and I'm a psychologist by trade.
So I'm, I'm also having a more or less like exotic profile for the area of, of cyber security, but also for the area of artificial intelligence, but regarding the current threat landscape. Yeah. The human factor is, is a really important thing. So hackers nowadays also have to be good psychologists and not only have a tech focus.
So, and this is, this is actually the first, the first picture. It's a very, it's a cliche.
So the, the hackers sitting in front of lines of code looking for a zero day exploit, it can use, he can, he or she can use to actually enter systems of companies. But if we are looking at the current threat landscape, this is but nothing but, but the cliche, because nine out of 10 kill chains, nine, nine out of 10 HES on companies are starting with a human.
So starting with a, with an employee. And we also see that these kind of like social engineering and, and, and human based attacks are getting much more complex.
So like three quarters of, of every company's all around the world are experiencing some kind, some form of social engineering attack attempt every year. But if I, yeah, Alexei already told us the whole business email compromise thing, which is also a part of some form of social engineering. I think currently every kind of company is experiencing at least some tech attempt. So social engineering is at the center of, of most more complex attacks when, when looking at companies.
And it's also the fact that 90, over 90% of all attacks start currently with a phishing email because it's, it's so easy. It's, it's basically a very efficient way to get into the systems of companies and hackers nowadays are very professionalized.
They're organized. So they're acting according to business case. So when you're acting according to, to a business case, you are basically using the most simple way into a system. And that's currently is the user, especially in, in, in times of like cloud applications, et cetera.
It's, it's very easy to enter a system of a company just via, via, via email. And we also see that these attacks are getting much more sophisticated. Hackers are putting much more time and money into the attacks. So we see a huge search, huge rise in basically spearing attacks are more targeted attacks where a certain victim inside of a company is getting targeted. There's gonna there's happening. There's information gathering happening upfront.
And yeah, we see, we see a search in those kinds of attacks, which also demonstrates that there's a lot of money involved because hackers are taking more time to, to conduct more complex attacks basically.
And that's also the reason why we see fishing and social engineering attacks in all kinds of segments, all kinds of sizes of companies. A couple years ago, it was limited to the large ones, the, the really, really profitable victims. But nowadays we also see also Sping attacks in very small companies.
We, for example, have like 50, 50 employee companies that are targeted by a Sping campaign. And we also have the very prominent ones, 2016. We had the presenter campaign of Hillary Clinton. Yeah. Basically attacked and successfully attacked by a fishing campaign. It was a comparably simple attack basically based on a Google mail warning, your, your account has been compromised, sent out to the com to the campaign leader and fact very conveniently. It also included a link to actually yeah, renew the password and he clicked on it. And basically all these internal emails were leaked afterwards.
Then we had a very prominent case here in Germany, Leon I've most of you probably have heard of, it was an example of a very, very lucrative C X O, or CFO fraud. Basically a fraudster sent out an email spoofed, the email address of the, of the CFO, and basically asked an employee for a wire transfer of 40 million euros, which then was executed. And basically, yeah.
That, that money is lost. There's currently, I don't think there's a lot of cyber insurances that would cover such a case because it's basically also yeah. A problem of the internal processes that there's no for eye principle, etcetera. So very lucrative and very, very costly for companies if it's, if it's successful. And the case of Leoni basically was a simple attack, probably a couple of hours of preparation.
Then we also have smaller organizations also case from Germany, the Lucas hospital in NOIs couple years ago, was hacked by a very broad fishing campaign based on a malware or ransomware in that case.
And you can see here, it's not only money that is at stake. It's also in certain cases or increasingly cases. It's also the lives of people that is at stake. Basically here. The consequence was that the Lucas hospital had to shut down the emergency room for, up for, for a week basically, which is especially sad because they were yeah. Pioneers in, in digitalization, in the healthcare sector.
Yeah. And, and basically they had a huge threat or, or exposure because of that. So you can see it's, it's, it's, it's really effective and it's, it can be costly, but it can also be really dangerous. And if we look at the schemes of those social engineering and fishing attacks, we see currently, then we can also see or realize that these schemes are not really new. Basically you all have heard probably of the, of the Nigerian scam.
It's, it's the, it's the label for nowadays really prevalent form of advanced fee scam. So the fraudster is gonna ask you if you can transfer a couple thousand of euros because he is sitting somewhere, but he's a rich heir and has huge millions at a bank account. He cannot unfreeze and he needs some money. And if you give it to him, then you will be participating in that huge fortune that will be released afterwards. And that kind of scam is prevalent.
A couple of hundreds years ago already is in the can be traced back to a fraud called the Spanish prisoner, which was already prevalent in the 18th century. Not via email, not via fax, but in this case via regular mail or me like real person messengers. So we see like over those 200 years ago, we haven't learned a lot, basically parallel to that.
The, the technological, yeah. Innovation took also place. So we have still the same schemes, still the same tactics frauds are using, but we have much more evolved technology. So I was already talking about spearing attacks. It's only one form of complex attacks that are conducted. Basically the fraud is looking up information about the victim from social media, especially like, for example, your position information in, in social media, in like professional networks, like LinkedIn or sing, it's a, it's a perfect resource for, for information gearing. And then yeah.
Basically designing an attack, tailored to you as a person, which is gonna be very, very convincing. Another one is a so-called double barrel where it's, it's a chain of various messages where the first messages are used to legitimate. The second one. So the first message would probably be from your CFO, just asking, Hey, are you at the office a really, really not malicious intent in that email, but the second one is then probably gonna ask for a wire transfer of a certain amount of money.
So the first one, you should keep that in mind because I'm gonna return to that later is really, yeah, not really, not really special and not really malicious asking just for a normal normal thing. And we also see maybe I think, not yet, but we see more complex forms of fishing in an automatized space. So automation also is, is being used for fishing attacks. Emoted is currently probably the most widespread malware around and the, the, the really cool thing about emote is that it's so versatile and that it's first changing its code constantly.
So spam filters and antivirus software is really for, for certain period of time, not able to, to trace it down. And it's using something called dynamite fishing, which is really cool. It basically scans the outlook account of, of a victim it's called outlook harvesting and generating a network of people that this victim was actually in contact with.
But even one step further, it also, it also screens the content of that emails that, that, that were sent around like legitimate emails and generates emails based on that content and sends, sends it out to, to, to a further victim.
So this is basically spearing in an automated way. Some people call it AR artificial intelligence, but it's not based on, on current ML, deep learning models.
It's more, more or less like based on, on like a keyword analysis, but you can see it's, it's some form of automation that is used during phishing attacks. And it's really, really, really effective. I currently wouldn't wouldn't know any company that is not being attacked at least as an attempt by a emote based campaign. So that's also the reason why current spam filters and technological barriers are struggling to detect those emails. In most cases, like in, in cases of emote, they are not even carrying a malicious payload.
So they're just carrying like an ex Excel file with a, with a macro in it. So spans are really struggling. If we look at the, at the numbers currently there's around like 160 million fishing emails. This is not spam. Just fishing emails send out every day worldwide and 90% are gonna be are, are being filtered out.
So that's, I mean, huge sum, but still 10% are getting through the filters. And that's a very constant number. And after that, we see that half of those males are being opened and half of those are actually being interacted with, so with a, with a really, really dangerous consequence. And the reason is because the reason basically is that employees are not really able to spot a really, really, really well made fishing email from a legitimate one. So you could see where, where should we actually optimize?
I mean, the, the current ML models are probably gonna optimize on the, on the left part, but I think there's still some way to go, but it's much easier to pull that lever on the ride, educating people.
And that's gonna be hard thing because there's also gonna be AI applications that are focusing on those, on those people and are carrying out people basic text. But one step back first we have a lot of AI and ML applications that are there to help us basically.
I mean, that's the main reason why we are yeah. Advancing in, in, in AI and especially in that field of human computer interface, AI should help us. So with examples from, from text to speech and NLP, I mean, that's a very old example. You probably all know it it's like Google duplex, it's that AI avatar that can really interact with people. It's a voice based bot interact with people in a really, really, really realistic way. And we are gonna see a lot of those applications in the future where we have voice interfaces in, in various areas of our lives that are really, really yeah.
Difficult to distinguish between those ones and, and real people. The problem with that is which channels can we actually trust in the future? Because if we have highly realistic voice bots, highly realistic, yeah.
Deep face, deep, fake based interfaces, or, or like, like facial avatars, which channel can we trust? So there's another example from the Texas speech field. It's a lie bird. You might also know it's a startup that has had just been acquired by script. They operate a podcast editing tool basically, but lie bird was up, I think three months ago was open to the public and it had a really, really cool feature or really cool, main, main, main feature. It could create voice bots that imitate voices like really realistically.
And when they started 2016, you had to had to train it with around like five hours of training input. So voice files that were indexed.
So voice files and text in, in parallel tagged, like the voice files had to be tagged by, by the words that were spoken actually. And you had to train it like around five hours in order to have a really, really realistic voice mimicking bot. It came up to one minute.
So, and you're gonna hear an example, like further, I trained it with like 15 minutes and it's already pretty convincing. This is how it looked now it's close to the public, but I mean, the model is still being trained, but only with the voices of people that are actually using it as a, as a part of script. But a couple, a couple months ago, you could still like train it with voice files, basically based on public people on, on celebrities, et cetera. And that's also what they did.
And they sent out this promotional video back
Then, hi, everybody this night, I am happy to share with you a small announcement about a cool start up called library. They launched today, their website where you can create a digital copy of your voice. They only need you to record one minute of audio. This is just the beginning and they're working hard to improve their results. They want to use this technology to change the life of everyone that lost their voice to a disease. By helping them recover this part of their identities, let's help them achieve this goal.
You can help them testing their software and sharing this video with your friends.
Yeah. So we can see it's a really cool example because I mean, you could technically hear that it's a, it's a digital manufactured voice, but I think it's, it's, it's really, it really sounded like Obama. And the thing is there's, there's some like, like beneficial applications to it, but obviously there's also some yeah. Criminal applications to it. There's another example from the field of text to speech and, and deep fake video synthesis.
Also a couple of months ago, the, the, the national news agency of China launched this AI news anchor. I also wanted show this example quickly,
Everyone in many English AR official intelligence anchor. This is my very first day in zing one's agency, my voice and appearance or model on J the development of the media industry calls for continuous innovation and deep integration with the international advanced technologies.
Okay.
It's based on a, on a real news anchor, but it was said that this person basically was like digitally created and you could on, yeah, you could basically enter text and the avatar would actually transform that into yeah. This what we just saw. It's not really clear if it's really the case or if it was pre pre rendered, but applications like this will be soon out there, another and last one, and that's, that's something I, I just implied, some of these applications are take like huge times of rendering.
So in, in the area of defects, we say, basically, it's, these, these models are either really easily implemented. Like you probably have heard of that Z app where you can put your face into, into movies, into movie actors faces that's really easy takes, takes not a long time, but it's not really convincing if it has to be convincing. It takes a huge time of rendering. So it's not possible doing that like in real time, but there's also a research group at Stanford for, and, and the MPI MPI in south Germany. And they developed a model that actually could do that in real time.
And basically this also use an example for that,
For both the source and target actor, we are able to manipulate YouTube videos in real time. Here, we demonstrate our method in a live setup on the right a source actor is captured with a standard webcam. This input drives the animation of the face and the video shown on the monitor to the left
This, yeah, you can see it's, it's always presidents, us presidents being used. This is actually a really old video.
So the, the current state of technology is, is, is much further advanced and you don't need much fantasy to actually think about criminal applications of, of those like beneficial applications. And one is actually using, especially the, the, the voice mimicking one in C CXO or CFO fraud applications. Another one is extortion. That's also something we we've seen in the world that videos are being created with people that are doing things they don't wanna see anybody them doing, and then basically blackmail them with them and also know your customer fraud.
You all probably have, have opened a bank account using ID now or some other identification method. And also this is pretty obvious that these technologies could be used to basically circumvent those identification methods or actually use, use it to, to commit identity theft.
But if you look into that seeks all fraud example, you remember, I just talked about the double barrel attack. This is something we presented at the BSE Congress, the, the office of, of security and information science of the German government.
Like this may we presented this hypothetical attack method, basically saying that an integrator could re really, really easily train a voice bot based on a CEO's voice call, several top level managers of a company in order to pre legitimate an email sent afterwards with a malicious intent. So, I mean, not much fantasy involved. We also laid out how to actually conduct it. And it was not a lot of work basically to, to, to, to create a setup like this. And you could say, okay, so, so how to train a voice board like this?
We had to look at the CEO of Zen, ER, and if you look, if you just look what's what's, what's out there in, in, in basically voice resources, you have like 32,000 videos where he's speaking somehow. So you could basically generate a perfect training material for that kind of voice bot to train a voice bot based on Joe ha's voice. And that would probably work like this. And this is me training that model for 15 minutes,
Who is, sorry, very briefly. We are voting right now. You have to help me with an extremely urgent matter. I'll send you an email with further information right away.
I have to go, thank you.
Okay. Took me 40, 40 minutes in total to, to set this up.
And if, if I had trained the model for a couple of hours more, it would be totally convincing. And you could also like change the tonality, change the pace of, of the speech. And then after what, obviously I would've sent an email telling my colleague to do that quick wire transfer to one of our suppliers. And he wouldn't actually think about it because he received this call before.
And the, the first legitimation call doesn't even have to be really interactive. Doesn't have to involve a voice bot that reacts to, to the voice, because it would probably only be a voice message on his mailbox with some like airplane noise in, in the back and saying, okay, I'm just boarding a plane. I'm gonna send you more details by email. And as I said, we presented this tech form at the BC conference.
I conference is a hypothetical example. But as a matter of fact, we've seen this summer, the first news report offer possible attack like this.
There was a UK energy company being attacked possibly by an AI AI based voice bot. It's still not really clear if it was really a voice bot because we've seen in the wild, we've seen a text like this also just using actors. But the insurance of that, of that UK energy company really claims that it was, was a voice based or voice bot based attack and a voice bot based on an AI model, mimicking the voice of the CEO of the parent company, actually of that UK energy company. So what can we do?
I mean, the first thing is rely on technical methods. There's a lot of funding. There's a lot of stuff going on. The DARPA is giving out $140 million in order to create AI that can basically detect AI generated fakes there's there's there's models.
There's there's one that's actually gonna be published in December for the next presidential campaign in the United States, trained on all presidential candidates that can actually, yeah. Spot really well, a deep, fake generated face.
But again, I mean, as, as in all AI applications that are trained to detect other AI applications, they're always one step behind because I mean, it's a, it's basically an arms raise, so there will be technical methods. Another one was presented at, at blackhead this year. There's gonna be a AI model based on the hearing system of mice, like real mice, because apparently you can train mice to de to differentiate between fake voice of yeah. Fake voices and real voices. And they are putting that now in a, in an AI model to basically also detect fake voices.
The other thing is education Alexei just said it, this is no science fiction.
This could happen already today. It's still unclear if it's happening today because also the simple attacks are so, so effective hackers are pretty lazy people.
So they, they, they will stick to that simple attacks until they're forced to move over to the technology that's already there. So spread the word. That's also one thing we are, we are doing like really intensively educate people, educate the, the broader population that this is no science fiction that in the future, their boss can call them and behind it, it's it, it might be a voice bot. And another thing is like really systematically. And that's what we are doing at soul safe, creating employee awareness tools to basically, yeah. Confront people with those kinds of attacks.
We are, for example, also building a voice bot that will actually call the, the employees of our, of our clients and present them to an attack like this, to be one step ahead of the hackers. And that's actually, yeah, the only truth that's, that's really important because that's gonna be the case also in the future.
Also, when we have AI models, we think that that, that hackers are still gonna focus on the, the people, because Bruce Schneider already said a couple of years ago, amateurs hack systems and professional professionals, hack people. And I think that's still gonna be the case for the next couple of years. Thank you very much.
Thank.
