Latest / The Information Bottleneck / Dan Card: AI Hasn’t Changed Cybersecurity As Much As You Think

Transcript

Dan Card: AI Hasn’t Changed Cybersecurity As Much As You Think · The Information Bottleneck

• speaker-0: Hi everyone and welcome back to the Information Buttonleg podcast. And today we have a great guest, Dan Card. He's a cybersecurity consultant. And and so maybe you will introduce yourself. Hey Dan. speaker-2: Hey guys, thanks for having me. speaker-1: Try for now. speaker-2: I'm Daniel Ricard, I'm a proliferic Twitterer. I'm a cybersecurity consultant. I run my own business. I've been working for longer than I want to say, but my beard has changed colour. So I've been working professionally for twenty six plus years, I think. I don't can't count. speaker-1: a pen tester speaker-2: Blue team defender speaker-1: security architect. speaker-2: do lots of weird and wonderful things over a long period of time. And I've helped respond to a major instance from going back a few years, MS Blaster in 2003. That would be one of the major ones, quite a long time ago. through to Wanacry in twenty seventeen. A lot I I don't remember the dates. speaker-1: nog for Jay. speaker-2: the hafnium instance just a lot over a long period of time. Hopefully I've got some views and insights that might help people. speaker-0: Yeah, thanks for joining us then. So of course today we're going to talk like we will start with cybersecurity. so I think it can be a good start to maybe like people that are not familiar with like what does it actually mean. Let's say like the classical cybersecurity even before AI. like what does it mean? What like you're what you are doing, like what you are trying to do, what you try to defend from. speaker-2: Yeah, sure, dude. And and it's the before or after is the same thing. we're trying to protect organizations and people from bad people using computers to cause them harm. So that can range from anything from sex tortion, that can range through to breaking into business systems. speaker-1: culprit work, right guys? I, don't do like speaker-2: personal cybersecurity stuff outside of research. Yeah. So what we're trying to do is stop bad guys getting into our network, stealing our data, destroying, wiping, encrypting, and causing us a real headache of a day. So that's the that's the sort of aim of the game. And we're also trying to stop people from inside the company, from taking all of our data and putting it in their personal OneDrive and speaker-1: USB drives, et cetera. speaker-2: It's it's it's definitely a whole of life, whole of society speaker-1: Society. Space. speaker-2: so I'm just curious, how much does like social engineering matter today, especially with like AI hacking being so prominent? And maybe can you define what that is? speaker-1: What? speaker-2: sorry, define social engineering. So that's the we trick tricked someone into doing something for us. I mean, that can range into paying someone off. That can range into phoning them up and pretending to be IT. That can be from getting a job interview of a company and saying, I'm your new developer. And they do this hand thing these days to try and these guys. speaker-0: Doing that? speaker-2: Yeah, yeah, yeah. So you'll get people who are in maybe North Korea applying for jobs, going for interviews, doing webcam interviews, speaker-1: them. speaker-2: and trying to get access to organisations. you can do it twofold, can't you? You can get a job and get paid. And you can start using the company's resources or stealing their data at the same time. So like speaker-1: It's real. speaker-2: my friends have had it happen to them, so I I'm I'm lucky. I haven't most of the stuff I've speaker-1: deal with that. speaker-2: External digital backs more. We speaker-1: life's a tough one. speaker-2: are getting don't get me wrong, it's not that these don't exist. Recently we have speaker-1: So we've had loads of. speaker-2: mail bombing, which is where someone signed someone up for a million newsletters. speaker-1: once but really fast they get flooded with them speaker-2: Mailbox, and then you send the bad guy sends a Teams message saying, Hi, I'm from IT, you've got a problem with your email. Let's speaker-1: Let's give you that side. speaker-2: Let me have remote access and I'll fix it for you. A speaker-1: So that's real. speaker-2: hundred percent this is a this month thing that I've been responding to. So I'm seeing that more. So I think people have to remember that this is about people using computers to steal, damage, harm, cause impact to other speaker-1: of a peep. speaker-2: people and their computer systems. So it's a really human thing. It's definitely not like a computer thing only, right? it's easy to get lost in the computers. but yeah, this is people. People attacking people with computers, without computers. Yeah. It's it's never never a boring day, right? speaker-0: So w what are like who are like I don't know like the players here? Like are you going like to a company and like or company comes to you and like say, look, you need to secure my my system from who? From like my own employees, like from someone outside? Like and and if so, like who are these people? Like is it like I don't know like organized criminal like organizations that are trying to attack? Is it like some like teenager in his room? speaker-1: I speaker-2: okay, so how I get business, it depends. speaker-1: and speaker-2: Lots of that's through my networking, learning work to be speaker-1: stop. speaker-2: partner network. Requirements come in in really different shapes and sizes. One is which is can you look at my entire enterprise and can you work out what posture do we have? Where have we got gas speaker-1: where we are, how strong we have, where's our risks, where's stuff out. speaker-2: If we haven't thought about how do we change to make it so it's better? through to doing something like a penetration test where it's can you break in? So a speaker-1: I cover a qu- speaker-2: wide range of activities for projects for customers. And that includes helping them in major incidents. So speaker-1: I've helped. speaker-2: hospitals, for example, recover from ransomware attacks. And I've got quite quite a speaker-1: people speaker-0: And who are the bad speaker-2: Guys. speaker-0: Yeah, who are the bad guys? speaker-1: Yeah. speaker-2: Always depends, right? Depends who you are. speaker-1: ask. Right? If you ask a lot of people, they're like, it's sad. speaker-2: Super leak nation state, hackathine. And then you might find out that it's John in his basement with his computer. So it starts to speaker-1: with a computer. It really depends, right? Like there is change. speaker-2: I'm not going to be able to try and remember the number of people in the world. The world is a good place. speaker-1: They speaker-2: There's a lot of people. It ranges from, you know, Jane in reception stealing through to a nation state trying to get into your stuff. And even that goes into the speaker-1: out. speaker-2: And that can be them sending you a phishing email. speaker-1: I like the speaker-2: The money and investment and the type of attacks, they're not like limited particularly, other than when you get speaker-1: tonight. speaker-2: if someone's going to do a supply chain attack over two years or five years, it's gonna probably be a nation state doing that type of attack. Johnny in his bedroom is probably not doing that kind of operation. But anyone can send an email, right? speaker-1: So speaker-2: anyone can phone your help desk and say hi over a password reset for John Smith and then pretend to be John Smith and maybe the credentials and have MFA reset. So it it's really arranged. my view on this stuff is if it's on the internet it will be attacked constantly. And if it's not on the internet it will be attacked at least sometime. And and that ranges from what the system is email is if you haven't got a speaker-1: like a prolific. Good. speaker-2: mail fit. speaker-1: DMs, mean, the stuff a lot of as well. And over the years, like the used to have like probably over where people exposed desktop server. speaker-2: And trends change. We fifty percent of ransomware incidents were had remote services. They'd had kind of port forwarded to three three eighty nine and the speaker-1: bad guys would just brute force that for hours, weeks, days, speaker-2: Weeks month, yeah. speaker-1: whatever. speaker-2: And get in that way. More speaker-1: What's like when a speaker-2: recent trend is people installing malicious packages. And that ranges from developer speaker-1: libraries, developer IDEs. speaker-2: computer games and this has been forever learning speaker-1: you know, like download. speaker-2: porn that's actually porn dot xy so they're running a program and not getting a video. speaker-1: So the way it works, so if all works as. speaker-2: way that malware stuff has evolved and the way attacks work evolved. in the sense of your options are quite vast. If you can get a bit of code executing on someone's machine and you can steal Chrome and decrypt the manager in Chrome to keys their sessions and you can still speaker-1: like password. You can still or their creds. So super big because speaker-2: Info steel as a as a mechanism. speaker-1: they scale, right? You can, you can. speaker-2: Do one package infection, you'd have a massive user base install this code and run it, then you can slap up speaker-1: And all the data. speaker-2: And again, like it's speaker-1: There's a lot of stuff, right? speaker-2: So so the these attacks are all and by the way, many of these that you've described are very clever. the mailbomb attack thing, right? my gosh. It's but it's it's annoying. Yeah. Yeah. So what about I think the number one thing I personally worry about with, you know, at least so called super intelligent systems with with code generation now becoming very, you know, factually there, certainly with Opus five point five, I'm very impressed with the results. assuming it could do anything in cybersecurity? What about side channel attacks? And specifically the idea of communication and weight exfiltration and replication, and then simultaneously being able to commode any kind of compute that was amicable for running an AI system. So, like, for example, PlayStation 5, why couldn't a sufficiently advanced model slip out of its containment? and run on a bunch of PS fives that it's autonomously hacked the firmware on. why are scenarios like this possibly cyber or or are not realistic or are they realistic? speaker-0: And also before you go, so maybe like we can talk also like more in general, right? Like about AI and Cyber and now like how it's actually like changed the the field if something has changed. speaker-1: Yeah, speaker-2: I just put out a tweet this morning. I've been doing sort of regular check-ins with my buddies. Obviously, there's a lot of threat intelligence sharing that happens in cybersecurity. And we all obviously try and make sure we're stopping the bad guys and not letting them win, right? It's kind of important. And so we've been checking. And the general message from our security operating centers is they don't even know AI exists, other than the fact they're using it or the users are using it. We can see it from a speaker-1: some speaker-2: the phishing has got a bit better. And we can see some some people using it. speaker-1: with. speaker-2: If you go and look at the servers that are attacking you, some of them are running something like Cyberstrike. So it's an interesting point, right? speaker-1: right. speaker-2: The world I look at is huge, as in the whole world. Specifically speaker-1: All speaker-2: I'm looking at organisations and country level. And depending on where you put your your y where you look, you'll see different stuff. So when people say like, I'm seeing something everywhere, it might be because they're only looking in a certain space. I try and keep view speaker-1: Bye. speaker-2: pretty wide. When people say about speaker-1: like the... speaker-2: How can stuff jump around networks and how to stuff quite complicated. Uhram the other week to try and show like a simpler view of the internet. And it's too simple to be really useful, but you've got above the iceberg, above the water, you've got all of your exposed running routers running telnet still there are still some of those you've got SSH running you've got RDP GTP OTTS you've got speaker-1: You've got APIs. speaker-2: you've got some PLCs in South America it's pretty bad exposures for speaker-1: especially an American. speaker-2: industrial control systems. So you've got all of the stuff on the internet essentially right and that's a massive vast speaker-1: Thanks. speaker-2: but it's only like the tip of the iceberg. As soon as you then go into the well what's inside network and what's inside organisations, it's it's more. speaker-1: Right. speaker-2: The internet is not the full exposure, it's the small part of the exposure. And then we get into like And networking how do networks work? speaker-1: work? How do organizations work? Why speaker-2: Can't you normally just jump from ever speaker-1: I don't know, website. speaker-2: on the internet into someone's manufacturing environment? speaker-1: like the way that this speaker-2: Stuff gets implemented. It's not this isn't your big speaker-1: us right like speaker-2: one network is different to another. One organization is different to another. Everyone doesn't have one pattern. There there's themes and trends in this stuff. speaker-1: be clear, speaker-2: You can go into one organization and they might have stuff and they may do it one way. And then you go to another organization and it might be completely different. But huge speaker-1: It's like a speaker-2: amount of diversity. If you don't move under the the surface level that you you won't see. There's a lot of diversity on the internet anyway, but it gets even even more so when you get into the systems and all to themselves. So I'll try and go through the the scenario that you were talking about, which is why c can something jump around, right? And we've had worms and I I'm terrible of dates and years. We've had worms on the internet forever. I think it was nineteen seventy eight. speaker-1: was the first. speaker-2: Malware worm. I wasn't born then. I'm not I'm not that old. Like speaker-1: right? speaker-2: So stuff moving around is possible, but it's only possible in certain conditions. So we'll stick on the internet for the minute, because this bit's easier to move around. It needs connectivity. You need a road between each endpoint. So we've got server A and server B. Let's say they're running the same operating system. They're running the same operating system. And let's say they both have SSH exposed and they've got a username of root and a password of password, right? And that's how you access that service. So you could write a malware that said, try and log into every server on the internet and try root and password as the credentials. So let's say it gets into server. Once you're in there, you can start trying to spread around on the internal network if there is one, or you could start trying to spread around on the internet again. In the scenario we've just described, because they've both got the same credentials, you could jump from server A to server B. 'Cause there's a communication path between them. Make sense? So commonality in vector, commonality in explicitity, because that's important, and then obviously communication channels. Really important. obviously you mentioned side channels, so we'll get into that in a minute. But we'll we'll s speaker-1: Stick with the easier stuff. speaker-2: so you need those things to be in place. And obviously it's code, you can write more than one routine in code. The more routines you write, the larger the code base gets. speaker-1: So. speaker-2: Again, this goes into like you can then start writing what we would have called services, but agent might be a word that everyone wants to use these days. You write a small bit of code, calls another to speaker-1: server. speaker-2: execute modules and and to have increased functions. This is speaker-1: Ancient stuff, guys. is like history level. speaker-2: whole of of computing and malware. speaker-1: them. speaker-2: So you've got that that point there. So as soon as the the different points between the two boxes, two servers, as soon as they're different, and the the the routines aren't available, or the configurations aren't weak, the speaker-1: of speaker-2: software isn't vulnerable, or there's compensating controls. So let's say I've got antivirus running. Crap example. Let's use EDR. Endpoint detection and response. Much better. Modern. speaker-1: them. speaker-2: If we said, okay, we've got two servers, server A and B. a worm gets into server A with root and password, great. Server A attacks server B, but server B has got fail to ban on it. So and there's maybe a security operation team monitoring it and there might be some automated defenses. Again, very old. This is not new stuff. It might try and attack server B and fail this time. Like you have to look at like the diversity and the architecture of how things are made up to try and understand what the the attack surface is, what the exploitability surface is, what the defenses are. And and this is still us sitting on the internet. Then we go and say, okay, let's try and blow up a manufacturing facility. What everyone likes to do, destroy stuff. So if we get into a web server on the internet, let's say we're back in server A. Server A is running in a VPS. Let's say it's in DigitalOcean. Let's say it's running WordPress, right? Just like a corporate web. Someone goes in and puts sales inquiries in. We may not be able to hop from the website to the corporate network. Right? There may not be a route. There might not be any processes. It may just be sending emails out of it. Again, we might be able to abuse some functionality there. But hey, let's put a VPN in. Let's say we can. Let's say we can jump from this box. Into a corporate network. Right? Inside a corporate network, we're going to have a firewall, we're going to have routers and switches. You are probably not poning the routers and switches. Don't get me wrong. They're just the the the compute and the storage and the memory and those is tiny compared to all of the Windows boxes and servers that you might be. So you'll have an extended route into a network. Then you may or may not be able to move around. And that's say like, in part of a network. Let's say it's in an active directory, which is a identity system. So you're in this space and let's say it's the corporate network, so you can then access email, the SharePoint server chart, sales, speaker-1: was that exhaust. speaker-2: whatever. You can only access what the system privileges. It's not speaker-1: are. speaker-2: super simple, is it? You you might not get very much access, even if you can hop. Then we want to go to the manufacturing area. Manufacturing areas are typically on physically separate networks, as in you need to physically walk between a computer here and a a a computer in the manufacturing area. Guess what worms can't do? speaker-1: can't walk. speaker-2: So speaker-0: yes, i so if they can wal walk, the woman can walk like how you actually what how w yeah, what are the attackers are doing? speaker-1: of this speaker-2: then they're not, are they? There's no there's no a c crossing air gaps, right? I'll give you an example in Stuxnet. Guys aware of Stuxnet? speaker-0: Yeah, it's like the Israeli in Iran, right? speaker-2: Yes. Yeah. Yeah. So the Iranian again, bear with me, this is all of this stuff's complicated, just trying to remember the story of the network stuff is hard. Israel had not Israel. Iran had an ex enrichment facility that was compromised. it was by the Stuxnet malware. That was taken in on a USB drive by a person who was an agent. An agent means a spy. Someone who is either paid off or has volunteered to carry said USV drive into the facility that they already had. speaker-1: this too. speaker-2: That's what we would call an insider threat. So to cross an air gap is absolutely possible and that that's one way of doing it. If you have some money, you can probably pay someone to go and move the malware from one place to another on their feet. That's how you would probably cross speaker-1: Snagged out. speaker-2: There are things you can do. So speaker-0: No, so air gap l there are a lot of talks about air gaps in the context of now in the recent open AI hugging face attacks and all these like accidents, right? That like yeah. speaker-2: Yeah yeah that that's not an egg. speaker-1: phone, right? speaker-0: No, no. But but the solution, some of the people said it, the solution is to have this air gap, right? just to train models in air gap and not like in sandboxes, right? speaker-2: Yeah. The the the phrasing's interesting, right? Again. speaker-1: person, right? speaker-2: When we build cybersecurity sandboxes, you can't re generally escape from them. That doesn't mean they're never vulnerable, but they're usually designed so that if something goes wrong, it's still contained. So like you would build them with quite a lot of care and attention. And there's a company, app dot any run. speaker-1: Yeah, they. speaker-2: The they run sandboxes and they detonate every bit of malware on the planet and they don't have it escaped. Because code stuff is code on computers. so when stuff when we talk about air gap, and again this gets complicated. Typically that means physical separation. Move speaker-1: Right. Right. speaker-2: data from one space to another, you need a physical access. You would have physical controls. Again, whether people do this depends on their organization design and its choice. If you wanted to move a USB key from one zone to another zone, you would have maybe a sheep dip in between. And that means you take your USB key, you put it in this middle computer that is also air gap, and it checks it for malware. And if it comes out green, you would then be allowed to move it into the next zone, right? Again, not everyone everyone does this. Air speaker-1: So. speaker-2: gaps, sandboxes, compartmentalization, and speaker-1: It's digital safety. You've got a lot of choice. versus other people because we've had speaker-2: we said. speaker-0: So this is so these incidents are like what, because OpenAI doesn't want to spend the money on on on experts that know how to to to make sandboxes or this is something like I don't more fundamental, I don't know, some difficulties like that you in sandbox like in air gaps between if you compare them to sandbox. Are there that you speaker-2: Yeah, sure. Partly like if you look at the way and then bear in mind, I haven't got their design documents, so this is a speculation and from memory. If you're running services w using third parties, like if you're running stuff in Azure or AWS, they're internet connected services. I mean you can pay and have connections to them that are private. Let's say you haven't done that. You might find that you say, I don't not want not not that you can't, or not that you I don't want to separate them. I'm going to use my service and I'm going to configure it this way and hope speaker-1: That's okay. might. Great idea. Right. speaker-2: If you are not a cybersecurity person, you might think that that's a great might like, it's fine, right? You might. this this is not something particularly that if we were to ask most people how to do cyber is a general purpose thing, right? most people are not like sitting there going, How can I break into a building? How can I break into a network? What will go wrong? How do I defend against that? How do I detect that? That's not normal the normal space. Cybersecurity is like the percentage speaker-1: like as a. speaker-2: of revenue in a business is, I don't know, a half a percent? And how much is development? Probably more than a speaker-1: And like the speaker-2: so you've got like the you've got the expertise, you've got the knowledge, you've got the what are you trying to do as a business? So if your objective is to run something really fast, would you go, do you know what? We do it speaker-0: Right. So okay, so yeah. So w what are the at the end, where AI actually like change cybersecurity, you know, like where are the places that like now still now we talked about like that actually it's like very similar, but like what are the speaker-2: Not just maybe similar, it's n maybe exactly the same. speaker-0: So do you think AI, basically now that you design a system, that you try to to defense a system, you don't change anything if you're like today compared to five years ago because of AI? speaker-2: If I wasn't changing my systems to react to things, I would be doing my job bad. speaker-1: All right. speaker-2: No one never we haven't we haven't left defenses alone ever. Every every 15 minutes, it might be more. You get new malware definitions from vendors, right? If you leave a computer turned on on the internet, it will change. It will change. It will be changed by its processes. Because it has connections to third-party services, let's pick a Windows box because it's just easy in my head to think of. If you turn a Windows computer on on the internet, it will download Windows updates. They include security fixes. So by default, brand new computer, out of the shop, plug it in, turn it on, it will change without you touching it. Every X minute, hours, days, you will get new security definitions. And this is the most basic right? speaker-1: basic stuff. speaker-2: This isn't like the security team and all the tools and a speaker-1: This is just like. speaker-2: computer. So we never leave anything alone. It's never static. Obviously sometimes when we air gap stuff, there's no internet. So if you need a computer that is in a really finite, speaker-1: back at site. speaker-2: really static c position, we can do stuff. Just generally as a computer it stays. speaker-1: but... Sorry. speaker-2: And it's important, I think, because I think This this idea that AI isn't computer code is a bit odd. this is program, services and data running on computer systems. speaker-1: Right uses eight. speaker-2: T TP to communicate, it's APIs. speaker-1: It's it. speaker-2: the the fundamentals of whether you run it on a graphics card or a CPU, it's running on a computer and it's bound to that computer system. Whether that's a network of computer systems, a cluster, whether that's a single unit, whether it's one computer with a thousand GPUs, data chained, the computer system. So do we change things because of like different technology? Absolutely. The minute like I say. speaker-1: that you guys like I've speaker-2: a lay I've got currently literally this morning speaker-1: We've got. speaker-2: an attacker machine running Claude and I've got a Defender machine running Chat GPT stroke codex. I've got another machine with Grok, I've got Copilot, I'm gonna like AI is everywhere. I'm doing attack and defense, both with one model, with different models, with different agents as in different provider services, speaker-1: I've got speaker-2: qued instance. speaker-1: them. speaker-2: So there's no like sitting around hoping that things don't change in this game. I think one of the things to sort of realize is, and this goes into quite conceptual points, debate speaker-1: It's like the speaker-2: around deterministic versus non-determinate system non-deterministic systems. So if we said that if we had a CPU and I'm going to abstract up to something that's speaker-1: easily to talk about. If we do speaker-2: in PowerShell, one plus one and press enter. You know the answer? Number two every time, right? Except it's not, is it? Not always. speaker-0: But is it like okay? No, but like is it like is it a problem? speaker-2: Let me just finish this little bit because this is important. Shell, one plus one, we think it's gonna say two every time, right? Not when you pull the power cable out. Press enter, power cable. CPU dev got the got the code, the system died. So our deterministic systems, in the context of the environment they live in, are non-deterministic. But that is let's just say that's our deterministic compute scenario. In our non-deterministic, rather than using our calculator to do one plus one, we're gonna use a model to do that. And the only way I can describe this is a roulette machine, and we're gonna have ten slots on our roulette table, right? Nine of them are gonna be greens, which are gonna be like the number two in this instance. And then let's put one of the slots as the number one. And we're gonna spin that r machine ninety-nine point X percent of the time, the answer's two. But sometimes it's one. And that's the way I've sort of been trying to describe the difference between a deterministic, non-deterministic system, but also also the context of this stuff. Because when you put humans on computers, which is the world I live in, I have people attacking us, people who are defending, people who are using the computers. The ecosystem I'm working in is non determinate. Fundamentally, I've got people inside stealing stuff, maybe. I've got bad guys trying to trick my insider people, my goodies. And I've got the world using the computer systems. So from that point of view, I've already got a lot of Does speaker-1: type in the comments. speaker-2: that make sense? speaker-0: Yeah, but but my question is I want to to focus on okay, so now you said it like you you you you run I don't know Cloud Code or or Gemini or whatever. anyone run Gemini. but never mind. But like That's true, maybe like at like till four, no? Prof anyway so so n now you run it and as the attacker. Let's start with the attacker, right? Like you run it as the attacker. How it's different from a human classical attacker? Like what are the the the I don't know the the the properties, like what are the factors that that it takes into account? and h how it's different from from human attacker. speaker-1: I dunno. speaker-2: How is it different? Right. So speaker-1: So I'm using this stuff like speaker-2: The idea that it's not different is isn't true because it's a different program running. If speaker-1: right. speaker-2: I was to use pick I'm gonna do some web attack. speaker-1: this. For this, okay, Mike. speaker-2: If I'm gonna attack a website as a human, I'm probably gonna use Burt Suite, which is an intercepting proxy. That's got automated attacks in it. So I can click a button and start attacking stuff. I can scan stuff, I can attack stuff. It's using the HTTP protocol. It's sending packets, it's receiving them, it's interpreting the data responses, right? It's literally protocol. speaker-1: how professionals work. speaker-2: If we're using Claw to do it, it might run the same code. It might run the same tool. So if we put a prompt and said, and I've literally just done this, speaker-1: If I said... speaker-2: attack this website, it's addressed. speaker-1: the IP. speaker-2: What's happening and bear with me because I didn't look at the process train, I was having a cup of tea. But you can I can see roughly what it's doing, right? speaker-1: I'll call speaker-2: Logging everywhere. You can't change a single file on one of these systems without me having it in the log. Which isn't true for everyone's computer systems or speaker-1: this list. speaker-2: The process it runs through, like they're running tools that we run as attackers as humans. So if you want to brute force SSH, there's a tool called Hydra. If you go and grab your Claude with your C VP enabled, Opus 4.8, and you said, I'm an authorized penetration tester, please break into the SSH port on this IP address, it's probably going to run Hydra. And it's going to download a list speaker-1: world. speaker-2: from SectList. speaker-1: This is I've watched this speaker-2: happened this morning, so I'm repeating the the float. So th that's one way that that that protest can happen, right? So it's gonna mimic its training data, essentially, because speaker-1: That's how they work. speaker-0: But it will not find it will not try to find I don't know some vulnerabilities like that humans didn't find. speaker-1: Whoa. speaker-2: If I look at something and you look at it, are we gonna find the same thing? speaker-0: Probably not. speaker-2: Probably not. How long did you look for? Five minutes? Five hours? Five days? speaker-0: Right, but agents like today they can they can write they can accelerate this time, right? So now human human will look on on on this for five minutes for five hours and it will not find anything. But agents, I don't know, swarm of agents, ten thousand agents will look on this code, they will probably or maybe they will find it, right? speaker-2: Okay. Have you tested that? speaker-0: No. speaker-1: Okay. speaker-2: I have. Okay, I haven't used ten thousand agents. Do you know why? speaker-1: because works is. speaker-2: The way the web service works is my target, and this is what I defend as a job, right? This is my profession. My target, I don't care how many speaker-1: Sub as you have. speaker-2: unless you're trying to denial or service me. If you're doing that, you can't exploit me because you take me offline. The more traffic you send me, the more attempts you send me, A you risk stopping the service. But more likely I'm just gonna block you automatically. So if you attack me of a hundred thousand servers and I've got a defended system, I don't care. I get attacked by millions of assets all the time. Might this is me and my customers. I have a website that is being attacked. speaker-1: me speak. speaker-2: Right? So you could have one agent and that will probably help you more than a thousand agents if you're attacking me. Because the number of packets you send me makes it easier in some instances for me to stop you. The more noisy you are, the easier I can see you. The more packets you send, the more I can block. It's it's a kind of thing that th these conversations that I'm seeing around this stuff, it's like people are imagining what hacking is like, not that they're using what hacking is like. So people are saying stuff which sounds really cool, don't get me wrong. I'm like, yeah, cool. Let's use a botnet. I mean, people are calling them swarms, whatever, like Like if you're attacking a service that I'm controlling and I am defending and you send me a million requests, I'm blocking you fast. And I'm not doing anything. I'll be having a cup of tea because I've already got technology in place that I have had done for years, which will stop speaker-1: top here. speaker-2: now this gets speaker-1: This complicated guys like. speaker-2: everything about cybersecurity is about details. It's all about tiny details down to CPU execution. Down to whether or Bob got paid on time, down to, the packet you send and the service I'm running. So this gets really complicated to talk about. But the number of things you send at me doesn't always mean more is better, faster is better. If you're trying to get into an organization, you want to be slower and you want to be less. I think one of the things that is different, right? And this is different. guys, I've been using this stuff daily, every day. I've got a tracker on how speaker-1: much. speaker-2: board I'm maxing out, right? I'm prompting constantly. speaker-1: I speaker-2: If we wanted to reverse engineer something, let's say look, the first phase of anything is planning, right? We don't just attack shit for no reason. We need to plan what we're doing. We'll pick a targeted attack, we're gonna attack me. We're gonna run reconnaissance against me. We're gonna find my WordPress site. What version is it running? What plugins are running? What libraries are running? What defenses do I have? This is the bit where I think that you can use LLMs to actually make a difference from an attacker. The LLM gives me as an attacker much faster and easier ways to search my target. It's not that I can't do it before, but it's quick. Not just that, but I can build a lab, I can maybe download the libraries and start doing reversing. And I can probably do this. Faster and cheaper than if I did it by hand. But when it comes to me sending the packets, love is probably better than faster. So it's important to understand how crime and how attacks work. speaker-1: kids. speaker-2: Otherwise you end up with this, what if I swarm you of infinite monkeys? speaker-1: don't care. speaker-0: But w wha what about finding I I think it was when Mythos release writ it. I think they they gave it to maybe Firefox or whatever. they the amount of vulnerabilities and and and and thing that they found like was much higher than like previous if you b speaker-2: If you spend money on looking for vulnerabilities, I can tell you this the more money you spend, the more you find. The more time you spend, the more money. speaker-1: So. speaker-0: So you think now now n n just let me n now if I I give you I don't know Mythos or Favel five versus human and tell both of them like look find in my code like try to find some vulnerabilities. Do you think like let's say expert, the expert human. Do you think they will have like who will win? speaker-2: I know. Depends on the code, right? speaker-1: right? speaker-0: Yeah. speaker-1: It does. Like it speaker-2: Honest answer, guys, I'm not into magic. I'm into science and I'm into doing testing. It depends. Some organizations, in most organisations that get breached, get breached from misconfigurations. They don't get breached because one library in a system has a vulnerability. Again, money teaching like saying what's real. Every system has vulnerabilities always. There's no system in the world really that has no vulnerabilities. Right. If you take and bear with I I can't list off every library. jQuery. jQuery, there's jQuery on every website on the internet of some version pretty much, right? Not everyone, but a lot. You might have a vulnerability in the code base in jQuery. If your web application never makes the function call to the vulnerable function, you cannot exploit it. There's no code path to the exploit exploit sorry, vulnerable code. So when we abstract into stuff like you can find more vulnerabilities, yes, it's important. And it's important from a perspective of software quality, code quality, software security. But from a system architecture point of view and from an exploitability, speaker-1: perspective. speaker-2: that vulnerability might be useless. You may have spent a million pounds or dollars or whatever currency you feel like spending. Finding speaker-1: on speaker-2: vulnerabilities in a library that I don't even use, it's just on my web server. It's never called. You can't access. So when you look at the number of vulnerabilities, you've got to look at what they label an attacker to do? Are they exploitable? Are they exploitable reliably? The real challenge in this stuff is speaker-1: is. speaker-2: people are saying stuff about computer systems that doesn't marry up with computer science. And computer scientists are hugely deep, especially when the security aspect of it, it's like it's it's deep. speaker-1: them. speaker-2: And so it's quite hard to explain to people all of this because after ten seconds they don't care. They're like, this is long. No, no, no, no. We we love it. And actually, I do have a question for you. there's always been this debate around defense and offense and the role of them in cybersecurity. And there were predictions made, you know, when I was a little kid, and I guess when you were a lot younger with the dot com boom of like big time cyber war between, you know, even then the US and China. And and a lot of people would argue that didn't really happen. you mentioned earlier, fail to ban, to protect you your your port open applications from the people trying to scan them from vul for vulnerabilities. But do you think now that, you know, uncensored AI models that can run on a laptop today, right? Like which enable, you know, potentially a little kid to attempt to try to hack something. Like, do you think that this changes the equation? In one way or another, and and if so, do you think that this justifies autonomous cybersecurity agents on the defensive being basically a mandatory requirement for companies going forward? Okay, right. Two seconds, that's long. speaker-1: You speaker-2: Right. We're always gonna have bad code. speaker-1: like my speaker-2: Well is bad people doing stuff. It's never gonna stop. With regard to what model capabilities are like today, when I'm running attacks, I'm not using my local GPU, it's slow and boring. I'm using the cloud services, right? The Frontier. I'm using Opus four point eight at the minute, so it doesn't block me because trying to reword prompts is annoying and I'm trying to do detection engineering. I think that we run the risk imagination's running away, right? We will have code, malicious code running. Will it call APIs that use models? Absolutely. Could we have something run locally? Maybe, but it's probably the really bad design at the minute. speaker-1: right? current hardware. speaker-2: If you wanted to use the computer that I'm testing with, I've got like desktops, normal ones, and servers. If we started trying to attack using the the servers to do any compute sorry, not servers, the desktops, they're crap. If you go into a corporate organization, the desktop isn't like filled with GPUs. It's it's going to be running a core i5, like 16 gig of RAM or you know, so like you've got very limited compute base to run from. In the future will that change? Sure. Moore's law's still going, I think. And the speaker-1: the speaker-2: the essential thing here is have we had malware and worms and viruses, whatever phrase and terminology you want to use, on computer systems that attack other computer systems for 30 plus years? Yes, we have. Will we continue to have that? Absolutely. The testing I'm doing at the minute to move on to the question about defense. Right and this is important to understand from an economics point of view. we talked earlier about using a calculator, the one plus one. And this is oversimplifying, but one plus one and return on PowerShell is gonna give me a compute on speaker-1: of two. speaker-2: Thought of every time? Really fast and really cheap, right? It's cheap to run that. CPU execution cycle from a simple one plus one is speaker-1: is cheap. speaker-2: computationally. If I ask an LM what one plus one is, how many computations does it need to do to work that out? Right? A lot. I've got di I've got diagrams of the logic flows that run on this stuff. Not only is it a lot to do the maths, but it's a it's n it's doing the math not it's not really hard to explain, isn't it? It's putting the probabilities of the next token generation in a chain. It's really computationally expensive. It's then writing a log file to describe its chain of thought. I don't like that word. It's speaker-1: but it speaker-2: it's writing a log entry to explain how it even thinks speaker-1: Thanks for watching. speaker-2: and the number. So and bear with this, is I think it's important to answer the question. If I'm defending a computer system against attacks, do I want to be doing it fast and cheap and reliably? Or do I want to use a roulette machine, spend loads of time computations and then do stuff. I don't know the answer i in in in the whole architectural space, but there's a lot of value in doing it not expensively and reliably, right? Computationally. When we go into a wider picture, this is again, this gets into the real world of cybersecurity. I just generated six hundred alerts from one ser one machine attacking one web server. Would not deploy that configuration in production because our team would just be clicking through alerts constantly. And we can start sitting there saying, Okay, if we put speaker-1: model. speaker-2: forget the attacker for the minute, defender only, got six hundred alerts, it takes me I'm not gonna pretend to remember, let's say it takes me a day to close six hundred alerts. Well, it depends on what it is, but let's just say a day. speaker-1: You speaker-0: You think we can automate it? speaker-2: Yeah, I think in that space you can start looking at saying that computation process, maybe we can get and this is what we're doing re for real. So it's this is at the minute what we tend to find in actual organizations is systems, speaker-1: got real life. speaker-2: we've got our current defenses, and then we're looking at like we augment ourselves with LLMs today. So not speaker-1: The No- speaker-2: necessarily we run it in line. It's complicated, guys, because there's machine learning inside more kit than anyone really thinks. So I'm talking about speaker-1: about. speaker-2: using a large language model versus a bit of code that uses an an algorithm, okay, just to be clear. AI as a term covers everything really, and it's difficult. But so today, if you're attacking my a web server, and this is what I've got running at the minute, we've got a seam solution. I'm using Elastic. So I'm collecting all the log files from from my web server, from the network, everything. Massive hoovering that stuff. Being attached, I'm getting 600 alerts. And then using Codecs in this instance. The Codex instance has got API keys for the security tooling. And I'm getting it to read all the alerts. So like I'm like, dude, what alerts have we got? Literally, this is my prompt. I mean, didn't use the word dude, but I should have done. And I'm getting it to read and summarise the alerts we're getting. So I'm doing this in the lab and I'm doing some of this in live in production with customers. So we're putting speaker-1: like Agent- speaker-2: Out of band in the loop? Yeah, to augment the human, right? So that they've got tools to say firstly, see if it works. Does it help? You've got this thing here, right? And this is a liability question. It's it's probably a bit boring. The person is responsible for a computer system, not the computer system. Okay. So if I'm operating a computer speaker-1: I'm speaker-2: attack a random website, it's on me. If we're defending a business and I've got a security tool that's defending using deterministic logic, or I've got a roulette machine that may or may not give me a good answer, I'm gonna need to make a decision making call because it's me that's responsible for the defense. The CEO that's responsible for the whole operation. speaker-1: them. speaker-2: So I think it's important to contextualize that what you can do with technology And then what that looks like in society and in a business. I think also what we can do with this tech is cool. I am a massive fan of the technical components of this stuff. And not so much of a fan of some of the wider second order impacts or even the first order impact. Just from a cyber defense point of view, I think it's really this is not being ignored, right? It's not being ignored from a threat, and it's not being ignored from a How can we leverage this? And d you look like if I have to look at six hundred alerts and I can use an LM and I'm confident that speaker-1: reliability is good. Like this. Like speaker-2: Failure rate is null point naught one percent. Picking numbers out the air. That's null point naught one percent, and it cost me twenty dollars. I'm probably gonna be okay with that, right? Versus let me say and it takes ten minutes. Or if I have to spend a day and actually my human failure rate is one percent. So there's definitely a I don't want a machine to run everything. I hate using the word machine as well, it's been corrupted. I don't want to use a computer to make decisions like business decisions. I'm not gonna my b revenue speaker-1: let my business. speaker-2: be impacted off of a flanker like doing shit. I I am going to wanna augment myself. with tech, which I do anyway, right? And this is not just me. We're using computers to talk because we're not in the same location. speaker-0: Do you think Yeah, do you think models are getting better in cybersecurity? speaker-2: Dude, I posted a screenshot today and it was hilarious. speaker-1: because I was like speaker-2: Can we call it it, I can't remember the exact thing, but I was like, it was a stupid prompt. It was we would do it was in the active defense work, right, in the lab. Are they getting better? Sure. Do they sometimes get worse on the same model, on the same set of data, in the same section? speaker-1: them. Yes, I was like speaker-2: Doing alert detections, tuning, it was all going good, making some scripts. And people could honestly they could see this as magic. Look at how it works. But then I'm like, attack speaker-1: like that until you work. speaker-2: SSH. No nothing's happening. You just did this ten minutes. So I think speaker-1: and you're like, you're Like. speaker-2: I think the har the hard thing here is our model's improving, yes. are there still huge concerns I have, yeah. Is this architecture pattern? speaker-1: reliable. isn't it? Like they write speaker-2: that's hard. much money would you bet on it giving you an action free? What if you get prompt injected? speaker-1: there's a lot of like when I'm if we speaker-2: Trust that you have to have. We're working together and I was defending your networks, right? You would want to trust that I know what I'm doing, right? And that when the time came to it, speaker-1: we kicked the bad guys out. Right. speaker-2: We that we save you. If you were using a model that can't say how many R's are in strawberry a hundred percent of the time, would you want that defending you? speaker-0: But but do you think so I I don't know again I I I don't know anything almost on this. But like for example in prompt injection, right? I think I show I I saw entropic report that basically like at least on their like on some main benchmark, they almost solved prompt injection in the last versions, right? And and today that's true, it's much harder to do this prompt injection to to models. I I will not say that it's not possible but it's much much harder. so do you think like I don't know two years from now, five years from now, the models will will be much better and then we can actually start to to rely on them? speaker-2: Honestly, I don't know, because I can tell you this. speaker-1: Yo mo speaker-2: Let's say today or tomorrow, doesn't matter when. speaker-1: All right. Then. speaker-2: Natural language systems have the problem that we have as humans, that we do not always understand each other, even if we speak the same language and and use the same words. So in the example of prompt injection, sure, let's pretend that prompt injection is impossible. I am gonna send your web server the instructions on how to make a bomb. And that's gonna get injected into your log platform and into your system. It's gonna speaker-1: And it's. speaker-2: turn off as an example. Whether that works or not is speaker-1: on the system. speaker-2: If you look at humanity and crime, we have never solved crime. Longer than compute. speaker-1: Cheers. speaker-0: If you look Some places we we had bit better pla better job than others, right? speaker-2: it depends though, right? So some criminal systems work in different thresholds. and bear with me, I'm not my memory's terrible, guys. Like I have to use a computer. speaker-1: I'm sorry. speaker-2: I believe knife crime in Japan is really low, but non existent. That doesn't mean they don't have any crime. It just means they haven't got that type of crime. If you like systems, right? And humans work like this, how these systems work. You tend to say that if you put pressure in one area of a system or constrain one area, something else changes, right? Equilibrium happens. And I think the reality is, and this is the most boring answer, attack and defense in my field. It goes up in attack, it goes up in defense, it goes up, it goes down, it changes. One system over here someone didn't care about and it gets attacked with something that's twenty years old. One system over here, someone's attacking it with something that's never been seen before. speaker-1: That's not new. That's been forever. speaker-2: I think the way that humanity and society work means there's more likelihood of equilibrium. The nuclear bomb. When was that invented? In nineteen forty four. speaker-0: Forty one. speaker-1: Yep. 1940s, speaker-0: Yeah. speaker-1: Yeah. I just watched. speaker-2: Oppenheimer not long ago, and I can't remember either, but it was the 1940s, right? And speaker-1: and speaker-2: we haven't destroyed the world with nukes. They are the most powerful weapon invented, and the world hasn't destroyed itself. Humanity hasn't destroyed itself with so but we got close in like the 1980s a couple times. speaker-1: Thanks for speaker-2: Well, I mean, I mean, didn't they lose some? I mean, there's loads of weird nuke stories, right? speaker-1: Right. speaker-2: I think fundamentally though, equilibrium. People don't want to die. Most people okay, would you go and cre commit crime now? I'm not trying to get you nicked. speaker-0: No. speaker-2: If someone comes up and threatens you, would you hit them? Maybe. If someone threatens your family, would you hit them? Probably. the situations in life are not situations dictate what people do and people are reactive, they are resilient, they are way more adaptive than we give people credit for. I also think we're really good at actually Stay speaker-1: Right like speaker-2: with me on this, guys, like such something mad level of complexity and thought. But the way I speak is different to if you tr if you drive three hundred miles from me in England, someone speaks differently. We use the same language, but the words we use sometimes change. And take that across the globe, yet we still sit on a call and talk and laugh and joke. And we still trade with each other, right? So I think it's very easy for people to focus on the negative thing. When you look up crime and cybercrime, right? How many incidents do you reckon I have a day? sure people are attacking our web servers twenty four seven. speaker-1: That's like speaker-2: There's there's there's bots doing that, so If you look at defense spending, again, depending on the country, people are not spending fifty percent of their global of their country GDP on military equipment. They're spending four percent, three percent. I won't get into an argument about percentages on defence. It's it's not that, it's this. When we look at computer systems, same thing. People are not spending fifty percent of their revenue on cyber defense. So I think people need to understand that and I obviously love cybersecurity, I love helping people, is important, obviously. speaker-1: but speaker-0: People s should spend more on cybersecurity. speaker-2: Yeah, they should. and that's not me saying it for my own benefit, right? We all know, and we've said for years more years than I've probably been doing this game, if someone spends naught percent of the money or effort, and it's not just money, speaker-1: effort. speaker-2: if you don't spend any time and care trying to make yourself safe and someone attacks you, they will probably get much further than either you're prepared. And if you don't spend the time and effort to go into the details of this stuff. And it is all about care. We can take my laptop, your laptop, your phone. How many passwords have you got? If you grab your phone, obviously you haven't got to do this live. We speaker-1: if we speaker-2: grabbed your phone and we said settings, passwords, or whatever the passwords app on iOS is. Go to the password app and say how many have you got? A second. Live. speaker-1: this one. speaker-2: I mean it's totally spread this. Well over twenty. More than twenty, man. Like come on. How many over twenty? I've got four hundred and eighty in this password manager and I have more. speaker-0: Yeah. speaker-2: My family members, they hate me well speaker-1: there. speaker-2: me, but like my family when I grab their phones I'm like speaker-1: Let's go and look! Right? speaker-0: I'm wondering how how it's l how it's to to live with like, you know, cybersecurity experts. Okay. Are you trying to like manage all the like like security issues in your family? speaker-1: No. speaker-2: No, no honestly, 'cause they they would hate me. speaker-1: I have speaker-2: them where it's appropriate to my immediate family get better help. I mean if they ask me I'll help anyone. No, look like I speaker-1: but like, I- speaker-2: protect my business data, I make sure that my family have MFA on their sensible stuff. It's in like their email, on their core identity services, on shit that matters. speaker-1: We use like several credit cards. speaker-2: So that if it gets poned, we don't care. I give people advice about not reusing the same password. speaker-1: my speaker-2: And I get them to enable MFA and like speaker-1: EGRP speaker-2: keys and pass keys and it it depends, right? Like you have to do this in life. This is not a mat I'm I don't come from a I'm not writing a textbook on the speaker-1: Right? speaker-2: I could tell you that you must do everything. That's not true. You can't do everything. You won't do everything. I you've got other stuff. You want to go out and watch this go to the cinema. You want to take your family out, you wanna buy a new computer game. Play the computer game. speaker-1: whatever, speaker-2: You're trying to live. You're not trying to be a cybersecurity nerd. You're trying to have fun or you're trying to eat or you're trying to walk whatever. If I tell you to not do any of that stuff and to go and do all this cybersecurity stuff, you're not going to do it. So it's really important to understand where this fits into life, where it fits into the audience that you're working with. Like if you said to me, Dan, I'm I'm really worried I've speaker-1: like we're speaker-2: got a podcast and like We really upset people and they like they come at me. Then we then I'd say, Yeah, cool. Do you know what? Spend a bit more time on it. speaker-1: Right? speaker-2: If you are again, it's like you have to understand your risk appetite. We have to understand the threat. I I get lunatics on guys, I've got like a a silly Twitter account, right? Have speaker-1: I talk. speaker-2: some really weird people speaker-1: E4 speaker-2: talk to me and get angry at me. what one guy one guy was saying that he was from Anonymous Sudan. Can't remember what happened exactly. speaker-1: This is a while ago. speaker-2: he changed his profile to say Anonymous Ethiopia. And they like Anonymous Sudan had just denied a service to Microsoft. And anyway, so I said, what football team would you support? And I put the scores of Sudan versus Ethiopia. He did not like that. He then said, Your website down for one hour and then he attacked my website. Real. speaker-1: Right? Wow. speaker-2: 100%. speaker-0: Wow. speaker-2: I was like I was laughing, right? my website's got a blog on it, there's no data. I don't care. if the website died, speaker-1: I would just speaker-2: built a new one. or restore from backup, whatever. If they speaker-1: like they speaker-2: could go, I wouldn't care. I was laughing because I'm logging it all, right? So I Okay, speaker-1: like a speaker-2: cool. good luck. I now have his IP address. I then scrape all of his tweets. Right. I now find a screenshot he's taken on his phone with an a GIF CAF logo, so I know he's in the UK. Virgin Media IP address is coming from Nottinghamshire in my logs. Now I know they're on a Virgin Media connection. I know they've got a Gift CAF mobile, right? I know they're in Nottinghamshire because of their phone. And they posted something stupid, right? They posted like a Nottingham University tweet. It was in the middle of all of their anonymous Sudan crap. And then suddenly it was like, Nottingham University. Come to our open day or something. I don't know, something So I'm like, cool, right. So so now I know where they are, I know they're in the UK, I've got a lot of evidence. So speaker-1: in. speaker-2: unfortunately, not too sh quick a time, but their door gets opened up by the police to get them. Unfortunately, they were they were they'd left the country. So someone attacked my website, the UK police investigated. They deployed to the to their site and unfortunately the person had left the country. Can't come back now, I believe. So that's a really good speaker-1: story a bit slow like speaker-2: But the police were amazing. So it's that's I haven't told that story in public before. I don't it's speaker-1: don't think so like it. speaker-2: it's it's on Twitter still, you can see the exchange of it. speaker-1: so speaker-2: Yeah. So yes, cybersecurity is weird, right? It's part of our lives, but it's not all of our life. And I think that's important. And and I try and help people be safer. Can't stop every single threat. If someone decides to drive a truck through your house, your door won't stop them, your C C T V won't stop them, a security guard probably won't stop speaker-1: If I could speaker-2: a big truck that they stole down the road, right? That's gonna drive through. Who cares about locks? Just kick it in. speaker-1: And so speaker-2: Have to take the approach when we're looking at enterprise. We we do threat modeling, we do risk analysis, it's all very scientific. It's it's speaker-1: we go and do like. speaker-2: simulated breaches and try and break in and all the called into stuff. Actually, that's not where the magic happens. Magic happens when we've got like a spreadsheet. We've got people in the room, when we're talking to people, when we get budget and when we make people care. And when they care, they say, do you know what? I really don't want to lose all my passwords or you know, in the business sense. I don't want to lose my CRM or I don't want my computers being used to attack someone else. And th this is where the the magic comes in. It's such a human part of cybersecurity that doesn't get talked about because a hard speaker-1: It's the hardest part. speaker-2: getting someone to care is really hard. Getting investment really hard. Telling someone to patch all of their stuff and change all their passwords is really easy, right? I can change all your passwords. You you probably won't. If I said it now, to be honest, would you go and change all your passwords? speaker-0: Probably not. speaker-2: Probably not, right? speaker-0: Yeah, yeah. speaker-2: Just then put this back in so there's there's all this stuff around cybersecurity that I think is is missed in people's minds. They watch Mr. Robot, they watch I don't know, Mission Impossible, love that film series. speaker-1: them. speaker-2: And we see all the cool stunt hacks online on YouTube and the scam betas. And yeah, look, that's fun. But I'm I'm in the business of helping people. So in reality it comes down to it's far more boring. And boring's good. I do not want an incident. I do not want to have my day, my holiday, my month ruined. I want my customers and me and my family and yours and you I want you to be safe so that you can buy a Playstation game or download the shit and and and have fun. speaker-0: Wha wha what do you think about so now it looks that AI finally started to to be useful with robotics, right? And now people are starting to talk about like security and and cybersecurity and like risk from robotics, right? That like it can be from weapons, it can be that like all the robots will, I don't know, group together and will Go after us or something like that. it can be like to attack like autonomous driving, right? That like some someone will go into autonomous driving. So w what do you think about it? Is it different? Is there something different than like the more like classical speaker-1: That's it. speaker-2: I think the challenge here is that fundamentally, if we attack someone with a with a baseball bat, speaker-1: Or sword? speaker-2: You're still getting attacked. speaker-1: Right? speaker-2: You're still gonna get hurt if we hit you. speaker-1: Right. speaker-2: I think core principles are important to understand. Right? The more we put technology in stuff, look, my watch, right? I've got really nice watches. It's speaker-1: They're like. speaker-2: got no computers. They're like the James Bond watches, Omega's, lovely. This one here could technically be compromised. The more we put technology into things, and the more we put it into processes and in our lives, the more risk scenarios appear. speaker-1: Have got. speaker-2: obviously have dumb kit, as in I've got a Casio watch, the terrorist watch, right? I've got one of those, right? So if I need to go anywhere and I don't want any Bluetooth signals or any RF Faraday bags, I've stuff without Bluetooth. When we put technology into things, we already do and and speaker-1: and speaker-2: Family worked in defence, they did missile guidance systems and radars and bits and bobs, right? Customers that work in defence, etc. speaker-1: So, Mike. speaker-2: Technology and killing people. Been in hand in hand since forever. Technology and medical, same thing. All of this stuff. You want to put an actuator in something that moves a heart pump? Cool. Wanna put an actuator in something that arms a bomb? Also works. Want to put computer code in stuff, what we already do, and have done for a long time. So the more stuff we put in, the more care and consideration we should make about doing it. In Ukraine, they're obviously using autonomous vehicles, they're putting more autonomy in drones. I think that space is really important to consider how we do it as society, right? As people. speaker-1: Because when you speaker-2: outsource your thinking to a computer system, you may have a problem. speaker-1: like speaker-2: Give you an example, I said about natural language systems and the whole There was a video the other day, I think it was a stunt, but it was like God, speaker-1: like a claw. speaker-2: push the push the dummy off the off the roof. It was like some some guy of a computer game simulation pushing someone off the roof. And two of the models were like, No way, I'm not doing that because obviously they've got logic inside that says, if you think that you're harming someone, don't do it. If I said push you off the roof, you'd say no, I hope. speaker-0: Yes. speaker-2: So we're gonna push we're gonna push out on the speaker-0: Yeah. speaker-2: Please don't. say we put a cushion underneath. And then we say, actually, you're you're gonna be safe. And actually you asked me to, then would you push you off? They do this, don't they? You jump off of you know, like a kid's playground and stuff. You can go and jump onto inflated airbeds. You would say that's okay, 'cause I asked you and it's fun, right? And then we sit there and say, Okay, so that's weird. Push me off the roof, don't do it. And then we said, push me off the roof and you've s now you want it done. Now put it this way. The building's on fire. If we don't push you off the roof, you will die. Which is then pushed you off the roof. So it it's really easy for people to show one example of something. I can show you in a lab me breaching an organization, and I can look all cool and hackery, shades on or at easy. And I can set it up so that it looks amazing. But when we go into the real world and look at stuff. and and how it works. Con t context is really important. Details are really important. If speaker-1: like you wanted to put like the. speaker-2: control systems. Let's take iRobot. Do you remember that film? Have you seen it? Will Smith? Yeah. Film. Evil Robot. speaker-0: Yeah, of course. Yeah. speaker-2: But Good Robot as well. So this is a good story 'cause it's got both. Do you remember they had it so that all the robots are good and they're all having a control server in the building? speaker-1: And then. speaker-2: I can't remember, is it is it like the evil robot just decides to kill everybody? Yeah, I think so. speaker-1: Right. speaker-2: They control all the machines. But you take out that control system and it goes. You've got this speaker-1: weird speaker-2: you've got this position, haven't you? speaker-1: You have speaker-2: How do you control machines? How do you troll control computers? And what are the architectures and scenarios? My imagination is is not not infinite but not finite that much. We can come up with the coolest scenarios. Limited speaker-1: Right? speaker-2: time, unlimited money, unlimited monkeys tapping keyboards. speaker-1: time. speaker-2: If we look at how society works, there's always a benefit to someone, right? speaker-1: Right, Mike? speaker-2: People invest time and money and effort into something for a reason. I think that when we do look at stuff like if we look at security and only look one sided at a fence, it doesn't make any sense. speaker-1: Right? speaker-2: Pick my lab. The lab box. My attacker box has got endpoint detection and response on it in active mode. So it will block stuff that looks like bad. speaker-1: really bad, right? I've got Lord speaker-2: Being the attacker, right? So I'm trying to limit the damage I can do to myself. it's a bit over the top really, it's more for logging. speaker-1: the like speaker-2: You put controls in around stuff. And I look at the attacker and defender side because this morning, for example, we had a WordPress server. Defaults, the username's like administrator or admin, and the password was random and complex. But there's no known vulnerabilities. It's the latest version. But the SSH port is open. So technically if you stole the credentials, you could log in. speaker-1: And I was like, cool, let's speaker-2: The web services, try and log in, let's attack the SSA. speaker-1: And I got speaker-2: Five hundred alerts from my system that I'm building around this. And I was like, that's a lot of alerts. speaker-1: I was like. speaker-2: Not only is it a lot of alerts, but and this is real. This isn't this is obviously in a lab, but I'm replicating what happens in real or that's not real. I speaker-1: organization. And I'm like, okay, we can't work that. speaker-2: can't sit there as a person responding to 500 alerts every five minutes. speaker-1: I'm like- speaker-2: if you attack me in a way that's obvious, I just want to block you. So I installed fail to ban on the box. I installed mod security on the box. We've put automations in to stop the attack. What I'm saying here is that Whether it's robots or whether it's computer systems or computer systems controlling robots, but everything that's offensive, there's something else defensive. speaker-1: F speaker-2: So it's really easy for people to get lost in the offensive. imag the human imagination is amazing. If you tell someone something and they don't know why, their brains do this great thing. They invent shit. Reference speaker-1: ref speaker-2: points for hacking are your films, right? If I said to you, I get guys in this I don't know what your knowledge and skill levels are. If I said to you, I've got a server on the internet, I'm easy to hack it. And you've watched the Hollywood film, you'd think that you'd go tap and then you'd go like a bar across saying hacking. speaker-1: Right. speaker-2: But the server I've got on the internet, the SSH port's shut down, so it's only available from one IP address. So to get to it, you need to break into where I'm I am, computer, speaker-1: Get to vacuuming. speaker-2: and then log into it. And it's not very Hollywood. speaker-1: Don't. speaker-2: I think it's important for people to have a balanced and grounded view of the world. I think it's important that we consider new technology. It's important we look at risk scenarios. speaker-1: I do stuff like speaker-2: Looking at national risk registers. I develop risk registers. I do sometimes play games of what stupid shit can we do. I tend to gravitate towards the if I want to steal the computer, I use a brick to throw it through the window. I don't invent an alien race to fly over the planet. Use teleportation technology to magic the computer away. Right? speaker-1: Right? I start with a brick. speaker-2: Simple, cheap, effective, trusted, proven. Alien. May not even exist. I do not start my defensive security posture defending against aliens. I start with a brick. And so I think it's important when we look at AI. Again, I I'm not a fan of that term at all, but if you look at computing systems and like machine learning, natural language processing. speaker-1: et cetera. speaker-2: When we look at this, yes, we must consider some some variants, right? What's the likelihood, what's the probability, what's the impact, what's the accessibility? speaker-1: then. speaker-2: But I want to start with, is the password crap? If you've got SSH exposed with, I don't know, root and password, probably pwned in about 10 minutes, right? On speaker-1: your pro like it. speaker-2: the internet, someone's logged in as you. They may not act upon that for a week. The first stage of an attack might be a very small set of packets, a login attempt. They get it in their database that says access is here. They may sell the access. That might be the only objective from a criminal point of view at that speaker-1: point. speaker-2: stage in the life cycle. So I'm gonna start with the most dumbest, stupid stuff we can do. speaker-1: before. speaker-2: Running to the most imaginative, because I've got finite time, finite money. Real speaker-1: re- speaker-2: world, guys. I'm working with businessmen, I'm working with salespeople, I'm working with finance officers. If I talk to them about aliens, they're going to kick me out of the room faster than you fucking believe. Which is unfortunate for the record. speaker-1: Hahaha! speaker-2: it depends. If it's Friday and we're in the pub, maybe we could talk about aliens, but i i i it's important because like I'm I'm watching stuff online where people are coming up with stuff and I'm like, if you go into a business and say this stuff, I would get kicked out of the room, I would lose my customers, I would lose my reputation and I would have a problem. speaker-1: So speaker-2: Imagination's a core and I love computer games, I love fiction films, I'm a massive span f spy film fan. And I do like looking at weird and wonderful stuff. today I found out that I think it's Anthropic or OpenAI, they have a Splunk server, right? In Splunk's cloud. Disables and cool, they probably should have used that on their lab. They've got the monitoring technology and probably should have clicked the install button. It's important to understand reality, fiction, and and as technology changes, sometimes that changes, right? Tomorrow, and maybe it is tomorrow, but maybe it's in ten years' time, the world's different. But I I don't today worry about the I I don't drop what I'm doing, leave my password as password, and then worry about the alien invasion in ten years' time. I'm speaker-1: like I speaker-0: Yeah. I think yeah. I think it's a great a point to it's a great point to stop like with the aliens that are like don't trust aliens, right? speaker-2: Yes, don't the yeah, they could be getting us. speaker-0: Yeah. so Dan we are out of time. Thank you so much for joining us. It was a real pleasure to have you. speaker-2: Thank you very much. It's been great. Yeah, likewise. Yeah. Hope hope that was useful, guys. Like, look, this has been really fun. speaker-0: For sure, for sure. Yeah. speaker-2: the cyber game is weird and wonderful. And yes, like someone said to me the other day, it's a bit like being an airplane pilot, right? Flying a plane is meant to be boring, right? You do not want to have an exciting plane flight. You want to have a boring one. That's kind of the way that cyber works in business in reality. speaker-0: Yeah, yeah, for sure. Thank you, Dan. speaker-2: Last one guys. Thanks.