Mikko Hypponen of Sensofusion

Jeremy Snyder: Mikko, thank you so much for joining us on Modern Cyber once again.

Mikko Hyppönen: Thanks for having me once again.

Jeremy Snyder: I think now you're officially on the Mount Rushmore of the top four guests ever on Modern Cyber. This must be like number four or five.

Mikko Hyppönen: I think it could be five.

Jeremy Snyder: Could be five. Yes. So I have a lot that I want to talk about today, and we don't have that much time. So without further ado, let's skip all the intros and bios and everything we've done. Those go back to past episodes. I have like five things I want to get to. I want to start with where we see workforce AI adoption today, and what parallels you can draw with the current situation to things we've experienced in the past.

Mikko Hyppönen: Would you first like to know where we are?

Jeremy Snyder: Sure. Let's start from there.

Mikko Hyppönen: Let's start from there. So this is a tactical drone factory. We are at the sensor fusion headquarters in Helsinki, Finland, which is where I work today. And as we've discussed before, the two of us, in previous episodes, I work nowadays with counter UAS, so drone defense. And this is a container sized factory building interceptor drones, which we build with 3D printers using fiber reinforced nylon to build the basic building blocks. And then you add basic parts like electric motors and, you know, small cameras and antennas, and you end up with something that you can actually use to take down Russian long range drones. So all you really need to do to build a small factory nowadays is a single container.

Jeremy Snyder: So let's talk about this, because this brings an important kind of junction, if you will, between warfare and cyber warfare. And we've seen now in the last couple of weeks a major attack. I don't know if attack is the right word. With the supermarkets, especially those in France last week, where the, and I can't remember the name of the manufacturer, Danfoss, right, where their controllers exposed on the internet had a vulnerability in them that allowed for relatively easy bypass of any kind of controls, basically the equivalent of root control on the controller, that allowed the attackers, again, is that the right word, to switch them into defrost mode. And of course, we saw, you know, just a huge amount of food waste and monetary damage. But when I think about that from, let's say, the warfare perspective, there's a saying in English, an army marches on its stomach. There you go. Imagine you take away what feeds the stomach. What's the impact of that? How do you see the current situation? Drones, industrial control systems, OT warfare, cyber warfare. How do you see it?

Mikko Hyppönen: Well, this was one of the reasons why I originally made the decision to shift from cyber into counter drones. Clearly, the war in Europe, the war in Ukraine, has a very important crossing point with cyber. I mean, it is a cyber war to an extent, but even more it is a drone war. If you think about what's the defining thing about the war in Ukraine, and now more and more also the war in the Middle East, it's not cyber, it's drones. There's been several cyber links to it, starting from cutting electricity multiple times in Ukraine with cyber attacks. And now, as we've seen with other attacks, as we record this, the previous night, the air alert mobile app in Ukraine was hacked by the Russians. This is the mobile Android and iPhone app everybody in Ukraine uses to get the alerts about incoming drones or missiles. Russians hacked that and actually sent a fake alert to all the users. So if that's not cyber war, I don't really know what is. But the Danfoss incident, or the supermarket defrosting of their freezing systems, nobody's taking any credit for it. It's not a ransomware attack. It doesn't seem to be cyber criminals. There's no message. It's happened in the Nordics now as well, so it's not just middle Europe or Central Europe that's been hit by these attacks. We don't know who's behind it, but you could easily see that maybe it could be a governmental attack, or testing their capabilities, or just trying to, you know, disrupt Europe as a whole.

Jeremy Snyder: Yeah. It's interesting. You know, from the defensive side, right, and, you know, I spent the first half of my career as a cyber defender, right, running cybersecurity for first two software companies and then a video game company. And, you know, at the time, all of the infrastructure work that we did, it helped the company get the job done. And even today, you see when people say, like, oh, what's the value that cybersecurity provides, you often hear this: well, it's a business enabler, you know, without it, the business can't move forward. Or you'll spend so much time on, set that myself. So what I see a little bit as a parallel is, is cyber warfare a warfare enabler? Because it's not direct attacks on the people, but it is things, to your point, let's say you attack this air alert system, you send people into a little bit of chaos and panic and into the shelters, etc. Does that create either a diversion for another attack? Does it create, you know, panic, disrupt the morale of the population that you're attacking? Is it an enabler of other warfare actions, that are drones, etc.?

Mikko Hyppönen: Sure, sure. I think that's exactly what it is, and it does have an effect on the way people trust these systems, and the morale overall. Another similar example: in the very beginning of the war, we saw Russians targeting with cyber attacks the systems of the Ukrainian border control, which then created these massive queues on the borders of Ukraine and neighboring countries for civilians, like women and children who were trying to flee the war. They couldn't get across the border. That's a cyber attack, an indirect cyber attack, or an attack on the system itself. And it's a means to an end in that sense.

Jeremy Snyder: So following up on this industrial control system, nobody's claimed responsibility, as you said. And as we record, at least, that could still happen in the next few weeks.

Mikko Hyppönen: Let's see. Probably would have happened already, but yeah.

Jeremy Snyder: Yeah. Fair enough. But is it a proof of concept? We don't know. Is it a test of capabilities? Again, we don't know. But it is in many regards, I think, predictive of things that we can expect to see again and again. And going back to Hyppönen's Law, if it's smart, it's vulnerable. I mean, every single one of these, this is like the, I don't know, the largest scale single event case study of this law in the real world, right?

Mikko Hyppönen: I'm afraid I was right when I created the law five, six, seven years ago. When we had connectivity and functionality into everyday devices like freezers, they become smart, which brings you great benefits, but they also make them vulnerable. If it's programmable, it is hackable, right?

Jeremy Snyder: And at the time, I think, you know, the first proof of concepts that we saw at the time were largely like DDoS attacks using, like, smart doorbells and, you know, home video cameras and things like security cameras.

Mikko Hyppönen: Exactly. Yeah, yeah, yeah.

Jeremy Snyder: So with this law, and now with the AI powered, basically, like, any vulnerability can be found for pennies. And can be identified. I mean, we have the whole internet catalog pretty damn well at this point, right? And you can go search it on Shodan and Censys in any number of sites. And if you want to see the Danfoss controllers, how many of them are exposed? And I heard a researcher give a talk about this, who said, look, France just had the highest concentration of publicly exposed Danfoss controllers. That's why they got this attack more than anybody else. Where does it leave us? Right? Because from a defender's perspective, first of all, patching vulnerabilities is always hard. But patching vulnerabilities in industrial control systems or smart devices is doubly hard, because you actually rely on the manufacturer for an update to a firmware, because that's basically all you can do with these devices, right, is update firmware to remove vulnerabilities. How are you supposed to respond? What's the lifetime of an industrial freezer? How many decades is it in use when you take it into use?

Mikko Hyppönen: This is the core problem of OT devices altogether. Same thing with factories and everything else. These devices are used forever, and they really run legacy systems eventually, always. Because you use them for such a long time, they will be out of support. Eventually they will be running on platforms, even if the system itself is supported, the platform it runs on, like Windows or Linux, it requires a specific version of Linux kernel which is no longer supported, and you end up with a nightmare. If it's hard enough to try to defend up to date systems, it's impossible to try to defend systems which are not up to date. And then you end up with the only solution, which is you try to keep them offline. You try to keep them out of reach of the attackers, which is surprisingly hard to do because of the basic building block of the internet, which is TCP/IP. The reason why TCP/IP is such a success story is that it reroutes, like, if there's a connection which is cut, it will find a way around it, which is great and awful. The awful part is that if you want to keep something offline, it's surprisingly hard to do. If there's any route to the public net, regardless of how complicated it is, it will find it. And that's why we keep finding these surprising things online on Shodan and elsewhere.

Jeremy Snyder: Yeah. But you raise the point, which has been an idea in the back of my head. I worked with a CEO years ago, after I left one of the major hyperscalers and I worked consulting around that space, helping companies migrate to the cloud. The CEO that I worked with at the time, he said, look, all of it goes in cycles. We're moving to the cloud now. Ten years from now, we're going to move back on prem. Is that where we're going?

Mikko Hyppönen: Maybe. But I think he does have a point. We really have been going in circles. If you look at the really long, long history and how everything's been designed and redesigned over and over again, we had the dumb terminals and central servers, which is exactly the same thing as the cloud, if you look at it from that point of view. But I think the real thing, the real shift, is obviously around AI. And when I've been thinking about what it really enables in the world of vulnerabilities, is that it enables sort of the industrialization of finding vulnerabilities, in the sense that we've built systems that automate all the parts, which means if you want to find vulnerabilities from a Danfoss system or any system, the only thing you need to do is to throw basically money and electricity into the system. You give it time, you give it the resources, and it will find vulnerabilities, if not now, eventually, from any system. So it seems like a pretty grim future, unless we realize that in the longer run, this means that we will find eventually, most eventually, all of the existing vulnerabilities from current systems. And future systems, the systems that are being built from now on, are more and more being built by superior programmers. I'm referring to the fact that humans are going to stop all programming, and all code will be written by machines. And it's easy to see that in five, ten, fifteen, twenty years, the code quality will be superior. The code will be written by machines which have read all the code ever written, as they've read every programming article, every security design guideline, everything about verifiable code, all of that. And they remember everything, unlike us humans. So you could argue that we will end up with a future which will have no more bugs. Think about that. No more bugs. And if there's no bugs, then there's no vulnerabilities. But getting to that point is a bumpy, bumpy ride ahead of us. It's going to be painful years, and a decade or two maybe, where we find all the existing vulnerabilities created by us humans and step into this bright new future where we have no more bugs. Very bumpy, very painful. There's plenty of work to be done for security researchers. But I do hope that, you know, I'm an optimist. I've said this before. I'm a technology optimist. I do see a future where vulnerabilities will be a thing of the past. Now, will that mean that there won't be any security problems anymore? I'm afraid not. If the systems are technically perfect, then of course attackers will go through the weak point, which would be humans then. But the situation would be better than it is today.

Jeremy Snyder: Going through humans, we already see, and not necessarily in terms of going directly through Mikko the programmer or Jeremy the wannabe programmer, but going through systems that are built by humans, right? And we've seen any number of so-called supply chain attacks where third party pieces of open source software get compromised by one or more bad actors. Most notably, I would say, the examples around things like LiteLLM, where the controller of the commits to the repo is a threat actor. I think in that case it was most likely attributed to a Chinese threat actor. So everything will trace back to the human in some way. Which leads me to the following question. Okay, machines are writing the code. How do you assign responsibility and liability, in the short term, to the code or to the systems built by AI?

Mikko Hyppönen: Well, you don't. That's the thing that humans still need to control. The responsibility will be with the humans, and the trust in the systems will be placed on the humans building these systems. Exactly the same thing with the problems these models and agents create. We can't escape the responsibility when our systems commit crimes or felonies.

Jeremy Snyder: Yeah, I mean, I think it was Bruce Schneier who made the comparison to dogs, that if you have a dog, you're responsible. If your dog bites someone, it's not the dog's fault. I mean, you're responsible for your dog, and it's not an excuse for you if your dog escapes and goes and bites someone. I couldn't, no, I couldn't control the dog because it escaped. Nope. Doesn't work like that. You're still responsible.

Mikko Hyppönen: That's the thing with models. Like, if your model escapes and commits crimes, it's not the model's fault. You continue to be responsible.

Jeremy Snyder: And there's already legal precedent for this as well. I think there's at least three or four cases where the company who created a system, not necessarily a model, but let's say a customer service chatbot or an agent or something that, to your point, escaped, it's the liability of the company. And that's where we live right now, right?

Mikko Hyppönen: Yeah, yeah.

Jeremy Snyder: So humans are liable, humans are responsible, or the organizations behind the AI systems are liable and responsible. Any parallels from your past that this really brings to mind? I have one that I always talk about. I always talk about the advent of AI being very similar to the advent of cloud, but maybe there are others.

Mikko Hyppönen: Well, the thing I wanted to mention is about how, when we look at the workplace, and these systems are becoming very powerful and useful, and you have a lot of pressure from the workforce to take these new tools into use, the IT department and the IT security department might be hesitant. When the top management wants something done, they will get it done. And we saw this so well with the iPhone in 2007, when the iPhone became a thing. Before that, we had smartphones, but they were at a completely different level. And suddenly you had this new Apple device, which was so neat and nice, and there was no central management available. We already had central management available for other smartphones of the time. That's what the corporate environment used, especially BlackBerrys, which were really easily centrally managed. And then top management wanted to start using iPhones, and the IT department was trying to backpedal. You know, you can't use those. We can't manage and we can't secure those things. And basically the CEO said, well, we don't care. We want to use this anyway. And of course, then they forced their way in. That's sort of what's happening now with these AI tools. The top management loves the fact that they can make their Excel spreadsheets and throw them into cloud, and it will figure out what's what, what they need to do. And the IT department might be trying to explain that, you know, we can't send confidential data into cloud models. And the top management says, we don't care. Just make it work.

Jeremy Snyder: Yeah. It's funny. I gave a talk at, I think it was Black Hat last year, but it might have been RSA. And I started my talk talking about going through exactly that, but with the Palm Pilot.

Mikko Hyppönen: Oh yes.

Jeremy Snyder: And that was the same. And it was, you know, a C-suite. And it was funny because, exactly to your point, I didn't have any way to control the Palm Pilot. So all I could do, and I thought and thought about, like, what's the risk? What's the threat model around this Palm Pilot? And the best thing that I could come up with.

Mikko Hyppönen: iPhones are way more secure than Palm Pilots, right? Palm Pilots were super simple devices. You know, one button, little stylus, etc.

Jeremy Snyder: Nice devices.

Mikko Hyppönen: Nice. They were fun.

Jeremy Snyder: Yeah. Um, not super productive in the end, if I think back on them.

Mikko Hyppönen: There's plenty of viewers who have no idea what a Palm Pilot is.

Jeremy Snyder: Yeah. Fair enough. Very, very touch pen system for note making. But basically, you know, it was, I'll never forget, it was the CEO of the company who came in and put it on my desk and said, make it work. And I said, okay, well, what do you want, Henry? And he's like, well, I want my contacts. I want my calendar. I want my emails on here. And you say, okay, well, like, what's in that? Right? What's the data that's in that? Well, contacts has, you know, all of the employee directory plus our key customers, not every customer, but all of our top, top customers. And, you know, Microsoft was a customer. Nokia was a customer. We had some very, very big customers at this company. You know, personnel information, etc., etc. Company confidential information, customer contracts, etc., would have been in the email. You said, well, what's the threat model? Well, my biggest threat model was that this guy would go out, have a few drinks and leave it in the back seat of a taxi. And so what could I do about that? There was no encryption on the device, right? All I could really do was make sure you have a passcode, right? So if somebody picks it up from the back seat of a taxi, you can't get into it. But in that example, and in the iPhone example you gave as well, it was really one stakeholder group, right? Really, this kind of C-suite. What I see different with AI is that it's not just this one group. It's every group, right?

Mikko Hyppönen: You're right. It's the C-suite. It's the developers, all the ones who try it.

Jeremy Snyder: Love it.

Mikko Hyppönen: Yes. Everybody finds a use case that helps their productivity or something, even if it's just, summarize this document, give me the top five bullet points that I need to know from this report before I go into this meeting. The benefits are so obvious. Everybody wants to use it. And that's why the pressure on the IT department and the security department is so strong.

Jeremy Snyder: Yeah, yeah, yeah. And this is the challenge that I see right now. And it's funny, because, you know, to your point about machines writing the code and machines writing better code, we've gone through a number of these exercises recently at FireTail, and a couple of initiatives that we've had. And we're at a point where, you know, we're actually doing a little thing about this in a couple of weeks' time, where we share a little bit of our lessons learned from this. We're at about ninety percent written code at this point. Human directed, nice harnesses around it, good security controls, much better process. We know from our own data the volume of bugs that we've squashed in the last three months, and it's staggering. And we're a relatively small company with a relatively small code base, but we've gotten so far, so fast with this. I see a lot of organizations who want to go this route but don't have the capability to get there. And what I worry about is that these organizations, they keep trying, but they don't know how to try securely. How do you tell an organization, if you're going through a kind of transformation, don't just transform, do the right things along the way?

Mikko Hyppönen: Well, it's probably best to start with a prototype, like a smaller unit, a smaller department, which takes these tools into use, tries different tools, has it nicely monitorable with the tools you can use, like FireTail, to get visibility into what's really happening with these tools, and then see how big the benefits actually are and what's the risk that you're actually taking. And the practical risk typically is fairly small, but of course it depends on what you're doing. So for example, this is a defense contractor, the company where I work. So of course we have very confidential information which we can never, even by agreement, or the contracts we have, we cannot send out of the country. So we have to have local systems, which we do. So it really depends on what the organization is and what they do. But I think the key really is the visibility. So you have to understand what's really happening. Where is the data going? What are the systems being used for?

Jeremy Snyder: Yeah. And one thing that is different now from the Palm Pilot or iPhone examples is we actually do have much better ways to get that visibility now than we ever had with the Palm Pilot, and than we had with, let's say, the first ten years of iPhone. Maybe it's only really, so if you think launch of iPhone 2007, maybe around the 2016, 2017 time frame, we got good MDM tools for Apple devices.

Mikko Hyppönen: It took surprisingly long.

Jeremy Snyder: It did.

Mikko Hyppönen: Yeah, but we have to remember, when iPhone launched, it didn't even have an app store in the beginning.

Jeremy Snyder: Yeah, that took a year and a half.

Mikko Hyppönen: Yeah. It evolved slowly, but obviously it got there.

Jeremy Snyder: And I remember, even in the early 2010s, when I saw some large scale financial service institutions trying to do kind of managed iPhone, because the benefits they could see, the benefits were real, they ended up issuing second devices to everybody under an enterprise app store. And that was the solution, right? And, you know, it was okay, but it wasn't great adoption, because nobody wanted to carry a second device.

Mikko Hyppönen: Right, right, right. Yeah. It's a trade off, like almost everything in security.

Jeremy Snyder: So, more trade offs. In talking about where we are with AI, we just put out our State of AI Security report for this year. And, you know, spoiler, if you haven't watched the previous webinar about it, but three hundred and some incidents just over the course of the last twelve months from when we prepared the report. I mean, the uptick, the exponential growth in the number of AI related security events is bonkers. I mean, it is the classic exponential, to the right, hockey stick. Like, I always joke that if I went into my board meeting with that growth chart, you know, my VCs would just throw more and more money at me. I picked the three top stories that I wanted to highlight in that report, and they were really OpenClaw, Mythos and the rise of the discoverability of vulnerabilities that we've talked about already, and then the kind of sandbox escapes of agents that we've seen. Those are the three themes, and I picked one story from each to illustrate that. First, any other stories that come top of mind? And second, anything that you think was really interesting in that, or from one of those stories?

Mikko Hyppönen: Well, I think regarding the agents in this case, of course, everybody refers to the Hugging Face incident, but we have to realize that's a drop in the bucket. There's been dozens of these escapes from all frontier labs. So clearly the frontier labs are unable, or have been unable, to really contain these systems. It's almost a joke at this point. If you don't have an agent escape, you're not a serious LLM at this point.

Jeremy Snyder: Right, right.

Mikko Hyppönen: And one thing that worries me, especially with these incidents that we know of, well, first of all, there might be incidents that we don't know of, that we've missed. But what worries me about the cases that we do know of is that there's multiple cases now where we know that agents have tried hacking into cryptocurrency exchanges. And if a superhuman hacking AI agent isn't scary enough, a wealthy, superhuman hacking agent is, because you could imagine, you know, a billionaire agent which could simply bribe its way to anywhere. Could you please open up this firewall for me for a second? I'll pay you a million. No? How about ten million? How about one hundred million? You know, one of the real core risks we have is that, if we don't really think this through and make sure that it doesn't happen, we might have rogue agents which are wealthy, which could set up companies, start hiring people, start building their own data centers, these kinds of things. So that's clearly one of the things that I think is relevant in this era for escaping agents.

Jeremy Snyder: And we say this right after the largest cryptocurrency hack in history, the Binance exchange.

Mikko Hyppönen: Yeah. Well, funnily enough, as we're recording this, I just returned back from TEDAI in Vienna.

Jeremy Snyder: Yeah, yeah. And this Binance, Binance, Binance hack. However you say, is it Binance? I think it's Binance. It rhymes with Binance finance. Yeah, obviously. Yeah. Binary finance. I never thought that.

Mikko Hyppönen: The hack happened, like, it hit the headlines like an hour before I went on stage at TED to speak about cybersecurity in the age of artificial intelligence. So I did mention on stage that, you know, these things could hack into cryptocurrency exchanges, and we've seen examples of that already. And at the time, I didn't know if the Binance hack was done by an AI agent. As we now know better, it was done by North Korea.

Jeremy Snyder: Phew. Just the North Koreans.

Mikko Hyppönen: Not an AI agent, but still the biggest theft in mankind's history. Six billion dollars.

Jeremy Snyder: What could be bigger?

Mikko Hyppönen: What could be bigger? I mean, yeah, there's nothing even close.

Jeremy Snyder: Yeah. It's crazy. I mean, that was the scale of the Lazarus one. I mean, that was tens of millions.

Mikko Hyppönen: They tried. I mean, the big SWIFT bank heist was five million US dollars, of which they only managed to steal around ten. No, the other way around. It was a billion, but they only got one hundred million.

Jeremy Snyder: Okay. But, you know, six billion is six billion.

Mikko Hyppönen: And they didn't get all of that either. But they got over a billion. So it's a sizable part of the North Korean annual budget.

Jeremy Snyder: Which is, it's remarkable. All of this sounds like science fiction, that we have a country which is so sanctioned that they resort to stealing cryptocurrencies, which are unsanctioned. Actually, it sounds even worse if you say it's a country that's so sanctioned, they can't even participate, and they're stealing imaginary money to fund this country. That's even worse.

Mikko Hyppönen: Yes. And of course, the way they actually manage to turn it into something they could use is that, if I understood correctly, they were staking the stolen cryptocurrencies with the blockchain system and then making loans against the stolen currencies.

Jeremy Snyder: Oh, I didn't realize that was the cash out mechanism.

Mikko Hyppönen: Yeah, that's what I read, but I haven't checked it in detail. It's very complicated. But it's still the whole idea that there's, you know, not only agents escaping, but potentially agents being able to steal money, to make themselves wealthy, to open up completely new ways of working for the agents themselves.

Jeremy Snyder: So just on that topic of cryptocurrency, one of the arguments I often hear about cryptocurrency as well is, it's mathematically provable. So its value is based on mathematics, not on kind of a trust escrow system like we have in a fiat currency, like the US dollar or the euro, where, you know, a dollar is worth a dollar because I say it's worth a dollar, you say it's worth a dollar. It's based on nothing.

Mikko Hyppönen: Yeah. Exactly. Right. But I mean, cryptocurrency is the same. There's no.

Jeremy Snyder: Yes. Okay. The limited supply, that I will grant you. That's mathematical. Great. But why is one Bitcoin worth ten thousand whatever, as we record, as opposed to ten cents, as we record? Right? That's because you believe it is and I believe it is, and so on.

Mikko Hyppönen: Well, the classical explanation is that they are wanted and they are rare, so someone always pays money for them. So, supply and demand.

Jeremy Snyder: Supply and demand.

Mikko Hyppönen: And that's basically it. And if there's, if they have any way of, like, usable properties that people want, and they're hard to come by and they're expensive to make, because it costs quite a bit to mint a Bitcoin nowadays because of electricity costs.

Jeremy Snyder: Sure.

Mikko Hyppönen: So I mean, you could compare it to anything else which is expensive to make, very wanted, and has a limited supply.

Jeremy Snyder: I mean, my counterargument to that, going back to my M&A days, is when you value a company, you can value it on replacement value. And you could say, well, if I had to build this myself, it would cost me X. And, you know, the electricity argument effectively leads down that path. Similarly, if you say, well, okay, it costs a lot to make. Well, it costs the U.S. Treasury a certain amount to print a dollar or to print a quarter. And the same is true for the European Central Bank.

Mikko Hyppönen: Well, most money isn't printed anymore. And you know this.

Jeremy Snyder: Yes. I mean, most money is just created in a spreadsheet.

Mikko Hyppönen: Well, one thing I wanted to also highlight about TEDAI was that I actually met Peter Steinberger of OpenClaw. He was one of the speakers together with me in the event. It was actually a very nice collection of people from OpenAI, from Anthropic, from Meta, from Google. Really interesting point in time to discuss where we actually are in AI, because clearly this is the fastest technological revolution we've seen. We could easily make the comparison to the internet or mobile phone or cloud. This is much, much faster than anything we've seen.

Jeremy Snyder: Yeah, one hundred percent. It's crazy how quickly it's changing. But did Peter share any great wisdom about OpenClaw? Because I have my own opinions on it, and we're, you know, full disclosure, we're in the process of decommissioning our OpenClaw, for instance.

Mikko Hyppönen: He was speaking more about the new project he's working on. It was actually a discussion with Alina Nikolaou, the curator of TEDAI, but it was more about things he's working on now at OpenAI and less and less about what actually got him here. But he's actually from Austria originally, and I think, I knew that the conference was actually in Vienna. So he was the only local speaker in the conference. But of course he lives in California now.

Jeremy Snyder: But aside from what we just talked about with the kind of Binance hack, what else was in your talk?

Mikko Hyppönen: Well, I did speak about the big picture. Like, why exactly, um, the AI revolution is changing everything in cybersecurity, in the sense that what we've built is a mechanism of being able to find vulnerabilities by simply supplying them with resources. And I also tried to explain to the audience why we've seen these escapes, because it's not obvious. I mean, people know these things escape, but why do they escape? And the comparison I used, the metaphor I had for the TED audience, was that if you imagine that you're put into a jail cell with crossword puzzles, a big stack of crossword puzzles, and your job is to solve all of them. And when you're done, you can leave, and you have fifty years to do them, but as soon as you're done, you can go. And then you realize that one of the crossword puzzles has been ripped. It's missing half a page. It's impossible to do this. So how are you supposed to do it? And you realize that you have, well, either you figure out a way to do it, or you stay in the jail cell for fifty years. So you come up with creative solutions, like, okay, I'm going to hit the guard that brings me food in the head and escape, try to find the missing part of the page, things like that. So when you give impossible tasks to agents with unlimited resources, basically unlimited token budgets, you end up with solutions you could have never imagined.

Jeremy Snyder: And the only thing I would add to that is, when we've looked at, and we've worked on, I think we've deployed about thirty agents inside FireTail today doing various tasks for us, and we've worked with customers on a number of them as well, the one thing that I would add is unlimited budget, access to tools and resources, impossible tasks, very little in the way of constraints. And that's the thing that I think is an important nuance to point out here. It's like, great, I give you this task, you know, blah, blah, blah. You say, well, just take the OpenAI Hugging Face scenario. Well, this task, the best way to do it, well, Hugging Face has the answer that I need. I need to go get inside Hugging Face. How do I get inside there? Oh, so there's some exposed credentials. There's this, you know, vulnerability. How do I get to that vulnerability? I get out of this micro segmented network, blah, blah, blah, etc. In that whole set of steps, there was nothing in the agent definition that said, don't do that. Don't break out of this network. Don't go exploit credentials that you've discovered elsewhere. And I will grant that it is impossible to come up with every single thing that you need to prohibit an agent from doing. But you could have very, very simple things. Like, if you think that you or your fellow agents are committing felonies, report it to humans.

Mikko Hyppönen: Exactly. Twelve hundred agents escaped. Not a single one of them told anyone.

Jeremy Snyder: You would think.

Mikko Hyppönen: You would think. Nope. Nope. It's even crazier. I mean, everybody thinks, laymen think, that they hacked into Hugging Face to solve the problem, or to find the solution. That's not actually the truth. They actually hacked the solution by being able to reverse engineer the algorithm that creates the flags. They hacked into Hugging Face to try to find ways to hide their cheating.

Jeremy Snyder: Yeah.

Mikko Hyppönen: That's what they were trying. They had already cheated, and they had a perfect solution to all the questions. They needed to find a way to falsify the logs, or change the public train of thought, or whatever, to hide the cheating. It's even more insane.

Jeremy Snyder: So I saw a talk of yours a couple of years ago, where I think it was maybe one of the first talks that I saw you do about AI, and a classic Mikko talk, kind of three stories. And each one, it's great and it's terrible.

Mikko Hyppönen: Sounds like me.

Jeremy Snyder: Yes. Honestly, you could summarize the last thirty minutes of our conversation with a lot of that. You know, these IoT systems, they're great, they coordinate all this stuff, they have vulnerabilities, they're terrible. We have the ability to find every vulnerability, it's terrible. But we're going to get bug free code eventually, that's great. Cyber warfare, all of this stuff enables real warfare. I don't know if I see the great on this one.

Mikko Hyppönen: I can try to give you one. Clearly, we are seeing a war of the machines happening already today on the battlefields. If you look at the situation in Ukraine, you look at the situation in the Middle East, we have machines, like AI driven drones, killing humans, sometimes completely autonomous drones. But even if they're remotely controlled, they're still, like, machines killing humans. But I see a future where we see, and it's already happening, we see more and more machines fighting machines on the battlefield. I have video clips from Ukraine where ground drones are shooting down aerial drones, and then I have other clips where aerial drones as a swarm hit together a ground drone and take it out. So we have machines fighting machines. We have these drones right here that take down other drones. This has been designed to take down large wing drones from the air. So if I have to choose between machines fighting humans or machines fighting machines, I would rather have machines fighting machines. And that's sort of the direction where we're going. So maybe there will be less death and blood on the battlefield, at least on the front lines, because of this shift. But not a day goes by in my job here at Sensor Fusion that I wouldn't think that I wish I wouldn't need to do my job at all.

Jeremy Snyder: Yeah.

Mikko Hyppönen: Unfortunately, it's just not realistic at the moment. I mean, the same is true for cybersecurity. You could say exactly the same thing about cybersecurity. You would wish that cybersecurity as an industry didn't even exist because it didn't need to exist.

Jeremy Snyder: There you go.

Mikko Hyppönen: Unfortunately, there's always bad people on the planet and evil things happening. So our job is needed. And in security, when you do your job right, that means that nothing happens when you prevent a disaster. No one even knew you were doing anything at all.

Jeremy Snyder: Yeah. It is, on the best of days, a thankless job. And then on the worst of days, it's not a fun day. So thank you for the job you're doing.

Mikko Hyppönen: Well, thank you for the job that you're doing as well.

Jeremy Snyder: And with that, we will close out this episode. Mikko, thank you so much for taking the time to join us and talk through all these things. Very thought provoking conversation, and we will talk to you next time. Bye bye.

Mikko Hyppönen: Thank you.

‍