Your Encryption Could Be Broken in 11 Hours: The Quantum Threat You're Not Ready For

**[00:00:00]** Hello everyone. My name is Armando Perez-Carreno. Welcome to the Web Talk Show. Today with us is Travis Malone. Welcome Travis. How are Hey, how are you? Good, good, good. Ah, I'm glad you're here. We're chatting off camera a moment ago uh talking about quantum computing and AI. We want to continue the conversation with the camera because I think uh a lot of you are going to take away interesting golden nuggets from Travis's experience. Uh a little brief intro here, Travis. What's your expertise? what what do you do on a day-to-day basis? Uh um so so listeners have an Sure. Sure. So, I've uh I've been in cyber security for the past 6 years and uh prior to that I was in public cloud infrastructure as a service, platform as a service, all that good stuff. And my day-to-day currently is sort of taking security and and business leaders are kind of concerned about and translating that into how they can strategically go tackle those things using different security products. And primarily what I'm responsible for doing is helping them understand that the sort of traditional approach to security where you buy a point product to solve this and you buy another product and another tool to solve this and you try and make them all work together is just kind of not going to work anymore. And to try and help them understand the sort of platform approach um to to addressing what's coming, right? So I'm I have to constantly be on the, you know, forefront of what's happening. And that that is an interesting place to be because a lot of times, you know, in in the especially in the Fortune 100s, which are the customers that I work with every day, they're really just concerned about what happens this quarter, right? Or or maybe 6 months from now. um sometimes they're interested in the three to five year conversation, but a lot of times it's if the board just asked this question and we don't have a good answer for it, help us catch up. Right? So, um so **[00:02:00]** it's it's it's kind of trying to balance those those those two things and and helping CISOs and CIOS understand what options they have to secure their enterprise. Nice. For the listeners, what is a CISO? So, a CISO typically reports into the CIO. So the CIO being your chief information officer and the CISO is the chief information security officer. So the organizationally the CISO is responsible for making sure that any information that is inside of the organization is protected and it doesn't leave. And that's I'm way oversimplifying it but most CISOs mandates sort of fall into the um you know the sock or the security operations center of of an organization. A lot of times it also includes the the network infrastructure services teams. So kind of taking those two sides of the business and making sure they're working together to protect the organization overall. That's typically their charter saying about how things are are moving super fast, right? So are they having a harder time these days with a lot of people saying, "Ooh, Chachi PT looks fun. Gemini looks good. This clot thing is working great." Is is that becoming a headache foros? Oh my gosh. Yes. The the short answer is yes. The long answer is they are sort of up against the uh challenge of making sure that the business can use AI to innovate and not standing in the way of that innovation but also making sure that AI is not you know that that expanded attack surface is not further compromising their organization and that is just come complicated more and more every day right so like initially when generative AI came onto the scene, we we sort of stopped talking about machine learning and moved into generative **[00:04:00]** AI and there was this sort of conversation about what is that attack surface and it turns out that just like anything else you can compromise a generative AI application, right? And then funny enough, some of the customers that I would talk to were like even less concerned about necessarily just using LLMs or building chat bots, right? And they were actually way more interested in exploring Agentic AI. And so those two things almost happened one right after the other pretty pretty sequentially. And so that's where that's where things got really weird because with your your standard like generative AI LLM chatbot type of an obviously there are things like prompt injection or or like data leakage or um a lot of times what we talk about with customers is like data and model poisoning. So you've got attackers who have figured out how to kind of tamper with the training or or the rag data stores that these mo that these applications are built on and and get those models to return either like toxic content or content that has like links embedded in it that you know could could be used for fishing or malware delivery or whatever the case may be. So those things were sort of like we were we were just discovering what those were and then all of a sudden agents popped into the conversation and the question became well what new risk does that posed and so then it became a kind of a a a challenge of of helping them understand like we're not necessarily worried about access to an application anymore. What you're doing with agents now is you're giving them you're you're giving them you know access to other applications to your network. you're giving them autonomy to make decisions. So, we're sort of not even thinking about like session tokens or the perimeter anymore. It's like it's like what you need to secure. Now, we're talking about like identity as the perimeter, right? And that that becomes **[00:06:00]** a really difficult concept for people to wrap their minds around, especially when they go, "Well, well, you know, tell me about MCP and and what you think about MCP." And when I turn around and tell them that I I I think you know MCP is still a little under under researched I don't think it's a very mature protocol to be honest with you. Then they kind of get all huffy and puffy like well we've deployed you know we've built all kinds of different applications with MCP and you know what do you mean by that? So then we have to talk about things like goal hijacking, tool misuse, recursive loops, cascading hallucinations, right? All of those things sort of work together to make it harder for you to to secure anything. Um, and and I've got like all kinds of examples, but overall those those are sort of the discussions we're having about AI with CISOs right I don't think most people realize all this is happening behind the scenes. They just use the tools. They hear everyone is using them and they're fantastic tools. We use them every day. I'm sure you use them every day as well. But when it comes to trying to secure something, it's this thin line between even good people, bad people kind of scenario, right? Like if you leave everything without any rules, then most people would be okay. But there will be some people that will step out of the lines and do something improper. Same here. Like you give this tool, it's amazing. If everyone was well-intentioned, they're great because they just make it so easy to do everything. But it all since it's so easy now they can be tremendously misused at a scale that is just Yes. And when you and I first kind of met and and were exchanging messages, one of the you said something that kind of I've it's kind of stuck with me, which is AI is such a great tool, but it's so misunderstood. And I think we're like we're not even we're not even at **[00:08:00]** the bottom of the very beginning of the bell curve for what AI is going to enable as a net positive for humanity. And so yeah, you're 100% correct. I mean, it's it's one of those things that can be really really great, but I but I think unfortunately and however misguided this might be, most businesses have begun to adopt it because other businesses were adopting it. And there's this sort of misconception that well if we build agents and we build them properly and give them the right guardrails we can recognize all sorts of efficiencies whether that is a reduction in headcount or you know we can speed up time to resolution for different incidents or whatever it is. Mhm. And the, you know, the unfortunate reality is that that can be true, but if you are not, if you're not thinking about how this is going to impact your organization when other technologies get sort of thrown into the mix, then you're going to completely lose sight of of what AI should be doing for you, and you're going to just run head first into shoving it into every product that you ever build without any real sort of benefit to an end user. And and that's where like you and I were kind of talking about earlier, this sort of like I don't like to use the word singularity, so I'll use like uh the word convergence of AI and quantum computing. That's what when when those two things happen, I think everything's going to get it's going to really really Where are we with quantum computing right now? Because I remember when when I first heard of it many years ago, I was interested. I said, "Oh, we can't be that far." And then I looked into it a little more and it just looked so far off because of the the scale of it. So they were making these machines that sort of did it but they were very abstract and they cost a lot and they didn't really do much of like this example like LLMs, right? LMS first started side note and people were like oh they're amazing. They were amazing but I mean not LLM starting but **[00:10:00]** CHPT like general usage but then you would ask it a math problem and it would get wrong right and like same thing now they evolved and there's reason for that and we've talked about it on the show before but the feel like at that point in time going back in time when I learned about quantum computing it was like yeah this huge thing that's the most powerful thing but it can not do like very basic stuff right and then and so so I think there's a parallel there and something very similar might happen or it probably already happened. You can you can enlighten us there that as the pace of progress in it grows, it's suddenly the capabilities just expand, not even exponentially, it's it's like it's it's probably even a different curve that we can even envision because we're talking about quantum space. Yes. Yeah. And I think that's where a lot of people get thrown off, right? is like they sort of assume that quantum computing is going to just like help us, you know, solve all the world's problems at once. And that's that's not true, right? It's also not true that quantum computing is the same as traditional touring computing. like it's it's it's it's almost governed by the principles of physics and the the constructs of quantum computing are so so different than traditional computing that people I don't think can really even grasp like how these machines are going to impact our daily lives. I also find it kind of interesting when I talk to anyone and I ask them like whether or not they understand what a cubit is, they they don't they don't, right? they they have kind of an idea because people who are in our profession, they might understand what a bit or a bite is and so they sort of take those two things and put them together, but but they're often wrong about it. Number one. And then number two, when I kind of explain to them what how quantum computing works and where we kind of are in terms of commercially viable quantum they're always their they're their their eyes like open because they just even like some of the the most tenured CISOs at the biggest companies **[00:12:00]** in the world, they haven't quite wrapped their brains around this. I'll give you an example. I don't know if you're aware that you can go right now to Amazon or AWS and you can purchase quantum computing Yeah, I saw it but I was very confused like how? So you can do the same at Azure, right? Um at AWS the product's called Bracket. At Azure I think it's called uh just Azure quantum. And what you're essentially doing is you're you're you're purchasing these you're purchasing these logical cubits and you can call them through an SDK basically. Um and the SDK will do things like spin up kind of a lambda function and a and a um sometimes it'll spin up like a direct connect straight to the actual facility where the quantum availability is. And then what it'll do is it'll draw on the coherence of the cubit and then take that and apply the uh the result it's able to garner from leveraging that coherence to traditional computing to then deliver a response. So it's not true quantum computing. It's almost like what we're calling the quantum advantage, right? It's basically saying like we're mixing quantum and traditional computing together and being able to leverage that to go do interesting things. So I'll give you some examples of how it's being used today because you're laughing but like this is this is real. Wait, wait, wait, wait. But I'm the reason I'm laughing is I when I was in college I went to the Hamburg high performance computing conference the supercomputing conference and I was there with people from CERN from the hatron collider and and stuff like that and they were talking the people from CERN the guy was like so we're we get stuff we get he didn't say stuff we get data from the from the collider right but there's so much data that we have to like there's not enough compute. There's not enough storage to process everything. So, we have to filter out a lot of stuff. And so, they were looking for partners and things to like **[00:14:00]** be able to process a lot of that data just gets discarded, right? It's just like when when the LM has to come pack, right? It's like we just you have to discard a bunch. So, when you're saying that, I'm thinking like, yeah, so the the the API call goes gets the data. How much of that does it actually get back that is available to then bring downstream and how much of it is lost or is it like just this huge like how does all that work? So there's there's that that answer is not a straightforward answer. By the way, when we talk about quantum computing in general, there's there's sort of two concepts to think about when you think about cubits, right? So there's there's logical cubits and there's physical cubits. And then in order for a cubit to be considered logical, it can't be what is called a noisy cubit, right? And so when companies like AWS are selling logical cubits to people, what they're saying is this is a cubit that we believe is going to be delivered as as a non-noisy cubit and you're able to leverage that to do whatever you're trying to do, right? And so to answer your question, how much of the data that you sort of have the quantum compute run against comes back as legitimate or or dropped? The answer is 100% of it comes back as legitimate, right? So that's where companies like EV car makers who are who are trying to crack the code on how to make like a better battery for their for their car, right? They're leveraging bracket biomed science companies, right? Who are trying to come up with all sorts of different scenarios for the drug they're testing, right? They're they're leveraging that technology right now because your point, they've got the empirical data. They need to then go run regressions against that data to determine what the outcomes would be. And just given the nature of quantum computing, we are enabling that data to be you're probably familiar with like what what used to be called elastic map reduce in AWS, right? Where you sort of take it **[00:16:00]** take the data, you break it out into different chunks, you sort of correlate the ones that make the most sense, and then you deliver what is probably um the outcome that the user is looking for to make it faster. Exactly. That's more or less the concept behind what quantum computing does except that quantum computing will deliver you the actual answer. So um that's kind of you know a roundabout way saying that the data come back 100% Okay. So let's let's unpack this a little bit. What is a cubit in terms that any person listening could sort of maybe begin to grasp? Sure. Um the way I've explained it to my father who is who who um you know was was in sort of building and maintenance facility you know roles his whole life. He knows how to build anything with his hands but when it comes to computing he's he's sort of lost. Um you know a traditional computer runs on bits and a bit is like a light switch right so the light switch is either on or off. A cubit is almost like a a coin that is spinning, right? So the coin is heads tails on it, but if it's spinning, is it heads or tails? Right? Well, it's neither heads nor tails. And it's also heads and tails kind of at the same time. So that's that's what a cubit is. Um, you know, a cubit is is the uh sort of representation of a bit being in both one and zero simultaneously. So when we talked about a cubit right we also have to talk about superposition right so super superposition is exactly that concept of those cubits existing in multiple states at once. So if you're trying to figure something out, you've got all these different cubits and those cubits **[00:18:00]** that align to kind of come to the conclusion you're trying to come to, those cubits are when they are aligned, that's they're in what is called superp position. And then there's the concept of entanglement, which is like if I had a coin and I gave you a coin and you decided to fly to Paris and those coins were sort of connected by magic, let's call it. Mhm. If I were to flip my coin from heads to tails, your coin would flip to heads, And cubits in general will exist in states of superposition and states of entanglement. And sort of the the the way that the the quantum computers come to the conclusion that you're asking them to drive to is through interference, right? So quantum computing literally runs on on waveforms and the the waves for let's call it a wrong answer sort of automatically cancel each other out and then the waves that are you know that that the probability exists that that the right answer is within that waveform those waveforms get amplified and that is how quantum computers kind of come to the conclusions. So, um, I'll I'll pause there. That that's those those are my best attempts at explaining the kind of concepts of how quantum computers work. How do you get the machine to do something? So, what is it? What is the quantum computer's input such that it can process something that you had in bits now in the quantum space? like what is that transformation? How does how does that I mean I don't know if we're getting too technical but like in an abstract way how does that transformation happen such that the machine can understand like the actual instructions that you're giving it and what you're expecting. Yeah, great great question. So **[00:20:00]** when we talk about bracket, we'll just use that since that's the most viable form of of using quantum computing right now. um at at at at its most raw form, what you're doing is you're you're feeding feeding data into like photons and electrons. Okay. Okay. So, it it's it's kind of weird to wrap your head around, but there's there's a number of different modalities that quantum computers take. Um, and depending on the modality that that quantum computer uses kind of determines whether or not the the actual cubit and the interference that the the cubit sort of takes on is going to be fast or slow. But if I were going to like write an app that was to take advantage of the current like bracket or Azure quantum or something like that, it's like you kind of run hybrid jobs almost. So the first step is like you would you would use an EC2 instance in AWS to kind of prepare the data and the question you want to give the quantum computer. Then you would send that through the SDK or through the API that that that Bracket gives you and then Bracket would use the QPUs is what they call them to kind of solve that impossibility and then would feed it back to EC2 and then EC2 being a classical computer would then take the result and then kind of optimize the result it got for the next question it poses. Right? So, it's like I'm going to give you one math problem and you're going to solve it and then you're going to tell me what the solution is, but then I have another problem based on the solution you've given me and I'm going to feed it back to you, right? So, yeah, I mean that's that's kind of how it's done currently. So, Brackets's got an SDK um and it's it's a Python SDK and that just kind of based on the problem you're giving it, it will determine what type of modality bracket should use in order to answer the question. Wow. Yeah, that's amazing. What? So is it does it support imagine because of the problems that you're describing does it support these huge data sets **[00:22:00]** or is there a memory limitation on it? How does that part work? Yeah. So typically it's like with any other utility based like cloud product, right? You're going to get build on the amount of cubits that you want to use to run your your calculation and then there's like a there's a data ingest and and egress component to it as well. But so you can you can choose the amount of cubits you want to use to solve the problem and you'll pay per cubit per hour or whatever it is and then based on the amount of data that you're sending it you'll get you'll get metered on that as well and charge for that. So it's really up to you right there's not necessarily a limitation except for the amount of people who are in Q to use the slots available on those modalities going back to mainframes for now right because that for now that explain the fundamental problem of of quantum computing right now is like I often have to help customers understand that we are not far away at all from commercially viable computers becoming cryptographically relevant such that they will be able to break the traditional encryption that we know and use today for everything in a matter of hours, right? We're like 2 years away from that. The limitation right now is the availability of the amount of cubits required to run that computation because as you know traditional encryption is basically just factorizing really large prime numbers. You know we our our traditional computers it would take you know hundreds of thousands sometimes millions of years to do that. I think the the most recent study I saw and and this is kind of if you've heard of the um Advanced Quantum Technical Institute, yeah, technical institute, AQT for short, they actually put out a a a research paper this month that introduced a brand new algorithm called the JVG algorithm. And the JVG algorithm is basically a way for them to say like if if if I'm remembering the study correctly, **[00:24:00]** it's basically an algorithm that's kind of restructuring the math that quantum computers use so that the processor on the quantum computer only handles like a smaller or like a leaner subtask, right? So it it uses classical computing to do pre-processing and post-processing and that allows the quantum computer to reduce its reliance on the amount of cubits it needs by Right. So they're saying instead of let's I think it was like you know 20 years is what a quantum computer you know originally maybe five or six years ago might have been able to um kind of break like an RSA key. They're talking about now like they they they theorize they could do it in in under 11 hours. And so that's if you have 11 hours on the system. if you have system. Exactly. And if Yeah. If you want to pay for that, which But make no mistake about it like the um like nation states um you know lots of like you know governmentbacked research um there are also like scientific lab like universities and stuff. they are 100% experimenting with this stuff and trying to figure out ways to reduce the reliance on the physical amount of cubits that you need and the amount of time it takes. Uh so it's the the actual algorithm itself optimizing that algorithm to speed those things up exponentially. And when that happens, not only is our traditional form of sort of encryption, cryptography, not only is that going to break, but like we were kind of talking about earlier, we're going to see quantum computing be leveraged to speed up artificial intelligence. And that is where things get get really hairy. Jeez, the way the computer evolved was so interesting, right? when in the first calculators, computers started and and then when we got to mainframes and they had the same issue, **[00:26:00]** right? access. Only the top universities or top research institutions had access and so but at the same time yeah like you were saying specific maybe countries nation states etc. certain organizations might build their own steals their own to try to get access to I don't know process more data to be able to do things. Eventually, we got to personal computers, which by the way, here in San Antonio, Data Point Corporation was very influential in the creation of that first processor for Intel. And and if any of the listeners are here in town or or are visiting San Antonio, you should definitely check out the San Antonio Museum of Science and Technology because they have a historical collection. not the main new building where the gaming arena is, but there's there's another building that has a lot of that history of how the first computers started, how those first personal computers started happening actually here in town, which is a lot of people don't don't know Yeah. To get enough credit, I don't No, no, no. It all gets lost. It's like, oh yeah, everything happened in California. It's not it's not true. Wow. Like uh a lot of people didn't realize the camera on the back of your phone. That was invented by a San Antonian. yes. Good friend uh may rest in peace, David Monroe. He he was the owner of the collection that I'm talking about in the San Antonio. Um Museum of Science and Technology. It's all his collection. He's even got an Enigma machine there. He had he really pushed he had a lot of patents. He really pushed the envelope in what could be done for computing and for cameras and for imaging and for a lot of things and like you said a lot of a lot of that is not known that our town has a lot going for it in uh in the tech space **[00:28:00]** which is which is very interesting. I myself didn't know a lot until I met David and uh it's it's amazing how technology if you go and take a tour like how technology evolved to a point suddenly in a very short span like if you figure when Aya Love Lace and Alan Touring were just getting into the thing how many years passed until we actually had personal computers and then from that like how many years until we had smartphones And now we have like AI and glasses and like in rings and it's it's nuts. And so yeah, the same thing will happen with quantum computing. And and people don't realize this, but as you were explaining, like I don't know if if if our listeners are are grasping the fact that all encryption, most encryption that we use in our websites, in banking and all that. It's just based on public private key exchange and we trust that that key is hard to break like you were explaining. So if this machine can suddenly break it in hours, then the whole thing comes tumbling down. So you were saying before we we got on the camera that like there's now this need of rotating keys more often and looking at the whole security scope differently. Is that is that accurate? My gosh. Yes. And I I'll I'll kind of tell you from the perspective of the governing bodies who who give business and and industry and security in general the guidance of number one what the government is telling organizations they need to do and they have to do and they must do as well as just like the general sort of like you know be aware of this sort of thing. Um, and I I'll I'll start by saying that one of the most recent attacks I've seen in the wild with AI was mind-boggling to me because um I I I had never even really thought **[00:30:00]** about this, but you know, in a traditional cyber attack, somebody will obviously get access to a network. Usually that's by like compromising compromising your credential, right? That's like the number one way that happens. Mhm. And then they will sort of find a way to install a way back in. So if they ever need to leave, they can always come back. So persistence, right? Once they've got the persistence, then they'll do what they call live off the land, which is they'll find ways to move laterally, understand the network topology, get access to further, you know, downstream devices and machines and things like that and go do whatever they're going to do, whether that's steal information or whether that's, you know, destroy things like that doesn't matter. The the fact is is that once they have their foothold, it's really hard to to find out how to get them out. And you used to have to know how to write code to do stuff like that. But one of the the the attacks that I saw last month in the wild, the adversary was able to gain they they they just fished somebody. they were able to get access to the network and then once they were on a a workstation all they that their persistence back was just an API call to a large language model and they said I have access to this machine now what commands do I need to run in order to get lateral movement and the LLM was like here's the commands you need to run but the attacker took it a step further and they actually dropped a script that was quite obviously written with AI um just based on like the commenting and some of the some of the constructs that the the the was used in the script. The script basically went it was it was it discuss it was it was a way to talk back to an agent and the agent was then able to download itself onto the the workstation and then the agent just did the lateral movement for the adversary. That's ridiculous. like you you don't even know how to have to know how to you don't even have to know technology anymore if if you could just get access to a network **[00:32:00]** and start running those those sorts of things like it it gets crazy now where I think a lot of people are are missing the convergence is when quantum computing is used to power adversaries like that and those sorts of like agents that get installed into humans will not be able to keep up There's just no way, right? So, there's this concept that a lot of professionals are talking about right now called postnural quantum threats. Right now, what we what we typically see with like malware is what we call polymorphic malware. It's malware that gets onto a machine and then it it sort of will recode itself based on some of the things it sees in the environment to evade things like endpoint detection and response tools and stuff like that. take that and then give it, you know, agentic capabilities and quantum speed. It no longer is just going to rewrite itself. It's going to completely transform what it is to the machine. And at that point, if you don't have any sort of defense, if you don't have a defense strategy that removes humans completely from from the loop, you're you're toast. And so that's that's where, like I mentioned, those government mandates. That's where those are like headed right now, right? So, a lot of what we're being told by um entities like NIST and CISA, in fact, there was a national cyber security strategy and executive order that was issued by the Trump administration telling federal agencies to prepare for this now. And what they're saying is go build a zero trust model that is completely autonomous and completely self-healing, right? And they they give you like different levels of how to do that, right? So they say, you know, number one, you need to be able to detect um and and have the visibility and the rapid alerting that you need inherently like built into whatever that strategy is. That's level one. Level two is automatic remediation, right? You you need humans on the loop, **[00:34:00]** but you want as many responses as possible to be automated. And if if if the if if the system or the agents that are kind of running this strategy sees something that they've never seen before, then they need to obviously like be able to reach a human so the human can make the decision on how to proceed. But level three is where this executive order says that you need to go. And that is self-healing, autonomous optimization, humans out of the loop. And the crazy thing is is I mentioned that we're we're two or three years away from that quantum commercially viable threat to to to kind of, you know, hit Most organizations haven't even gotten out of level one yet, right? They most organizations don't even have the visibility that they need. They don't understand what they're trying to protect and they really don't have a strategy for how they want to protect it. Because when we talk about things like zero trust, I'm sure you've probably heard of of that framework. But the Diesa gave us what zero trust should be and what it should look like, but they didn't tell us how to get there. And so organizations had to figure it out themselves. Um, and so lots of organizations are just now finishing the work of trying to figure out what that zero trust matrix should look like and then how they're going to implement it. And if you if if you can think about big behemoth corporations moving at the speed they typically move, which is at a snail's pace, like I don't know that they're going to get there by the time that this all happens three or four years from now. So yeah, it's it's a it's a very exciting time. Maybe for me right now. Yeah. And then yeah, you need to watch Tron and all the classics to see like or do I don't know if you are you a fan I am not. I haven't. I know people are going to hate me for this, but no, I have not. I have not read Dune. I don't know. I don't know. I think I started reading it and we didn't we didn't meet I2I. Like I'm a big fan of all the Asimov stuff. I've read most of his like foundation **[00:36:00]** stuff, all the Foundation stuff and and uh related. And I like science fiction like Ready Player One. But dude, I don't know. I just it just wasn't the time. You know how books you read them? Sometimes it's just not the right moment to read it and just so I'm going to try again. Yeah. But uh tell me what what well I was just about you know there's a there's this kind of like a car I want to call it's an archetype I'll say okay of people in the Dune universe called menntats and menntats exist because the human race basically outlawed computers like hundreds of years prior to doing happening but because it was cat what happened was catastrophic when when humans got really good with computers and artificial intelligence and things like that. And so humans said, "Okay, that's it. No more AI, no more computers." And so humans then had to then become the computers again. And the these men this Menat archetypes are they kind of follow the the leaders and the rulers around and the rulers go, I have a question about this specific issue. And the the Menntats know everything, right? And that's kind of how it used to be back in I'm sure medieval times, but Right. It's it's it's funny to think about things like that just how quickly this is happening and and and yeah and some of them got got a lot right. Obviously you've got the scary side of how and machines making their own decision especially at critical levels. There's this movie that I watched a few years back that's not I don't I don't think it's like an a movie or whatever but I don't know why I watched it but it was I I thought the premise was good. It's with this what's his name? I don't remember the actor's name right now but uh it's called I or something like that. Have you watched we called Eagle Eye? It's about this AI that has it's supposed to be very smart and it's attached to the military or whatever and it uh determines **[00:38:00]** because it knows everything and it sees everything. Spoiler alert, if anyone wants to watch watch this movie, skip ahead. It sort of sees that humans are good, but then it starts seeing all the wars and all the footage and everything and all the discussions and what happens, etc. And then it hears it overhears a conversation in the Oval Office or whatever or in the war room and they're talking about like this is this guy and then it that we want to take out and then the computer gives them like oh but it's just a 60% chance that it's the guy but there's going to be collateral damage blah blah I suggest a no go right and then the commanderin-chief or someone says like no no no just just just take take them out or whatever and so it starts realizing and then comes to the conclusion that the the the way to fix everything is to get rid of the chain of command that is like poisoned non morale or whatever. And so it does this whole thing and there's many stories like like this one, but it's it it just runs into like you're saying it's super powerful and super fast. So it goes it it generates voice to tell people things and like overrides things and and just puts this whole thing into plan and into action such that everything aligns for this attack that nobody knows about and then they pins it on someone else like he sends weapons to someone's door like so they arrest that person and it's it's a fun but but it's like if you if you step back and like analyze it like yeah Even if you've been playing with these agents recently, they sometimes make mistakes based on fallacies, right? Or or a how humans work. And yes, we make a lot of mistakes ourselves and sometimes very bad decisions, but if you're telling it to be completely truthful or completely whatever, it's going to do that. And that could back against you. it cascades because what what a lot of organizations are doing right now is **[00:40:00]** they're daisy chaining these agents together. So you've got one agent that writes the code, you've got the other agent that lints the code, then you've got the other agent that runs the regression tests against the code, and then you've got the other agent that pushes it into the uh uh UAT. And you know that like there's there and that each agent's using maybe a different model or it's it's trained on different data and like if one of those models gets poisoned for instance what happens then when you know it gets sort of caught in this loop of wrong information or writing bad code or like injecting malicious you know uh payloads into that code and then it passes that information on to another agent. Now that agent's model and memory starts to get poisoned, but then it gets stuck because it can't verify the code. It doesn't recognize the code, right? And so now everything downstream is stuck. Um, and so it's like, yes, like not only are these things black boxes, but if we give full autonomy to them, you know, we could we could end up shooting ourselves in the foot and finding ourselves in a real bind. You know, it's it's again finding that balance between helping folks understand what the risks are, but also helping them like look forward enough to to be able to enable innovation for the business, but also to protect from what's coming. That's like the hardest part of a CISO's job. And the hardest part of my job is communicating it all. Because then what what most people will ask me is like how do we if we're going to build this autonomous self-healing like closed loop security stack where do we start and the honest answer is is like there there is there is no like runtime for this stuff you know obviously with like agents or LLM like you know people are using Python as kind of the lingua franca of you know of of how to deploy those things like u I think like lang chain lang graph crew AI all that stuff is is python based but then you've got like actual policy enforcement and most providers are giving people terraform to **[00:42:00]** do that or there's open policy agent which is kind of the I think the closest thing to an industry standard for like machine readable policy but then then you get into like the orchestration or mediation piece of things and and you know lots of shops are using like kubernetes native pipeline orchestration so like argo or or Tecton. But then, you know, now we're we're talking about identity as the perimeter and like more and more professionals in the identity space when they look at Spiff or Spire, they're they're turning to Rust. And so, you've got like all of these sort of different, you know, competing technologies, you know, that you're trying to use to piece together other competing technologies. Um, and it just is it, you know, we haven't uh as as a as an industry, we haven't standardized on anything yet. I think if if there is a company out there that can purpose build that sort of autonomous security runtime, they're they're going to own the market share and they're going to be very rich people, but you know, it's just not there yet. And that's just sort of the next challenge that that most people have is like they know quantum's coming, so they have to migrate all their crypto from the old standards to the new standards, but then they've got to outfit all of their applications to be agile. Um, and then at the same time, right, they've got to put these guardrails in place for AI, but then they also have to build this overarching framework for it all that is responsible for itself. And there's no standard way to do it. Um, and so if we don't hurry up and figure that out, we are we are in for uh we are in for some wild times. Especially since while there might be some rules here for certain things and what you can and cannot do like you were talking about hacking with LLMs etc. Full disclaimer, you should not hack. It's legal, right? Unless you're doing it for pentesting or as your job, right? Yes. If you go to a regular LLM like CHP, cloud, etc., and you try something like that, it will tell you, "Oh, sorry. I cannot help **[00:44:00]** you with that because it's like they have guard rails on it." However, as the Chinese recently showed us, they can get around that if they convince the LLM enough that it's actually like an educational kind of thing. And so they can still be worked around. And by the way, if you're interested in this, have you have you heard about the Gandalf the tool? Yes. There's yeah the little there's a there's a little training thing where you can anyone can go in and it's from this security company but you can sort of see how you fool an LLM and it starts by baby steps and it sort of guides you to see how people can extract information from LLMs sort of as a learning experience so that you can learn to see if you can beat the wizard or whatever. And I thought it was really fun because also it it opens your eyes into well maybe I have to protect my LLMs and my my endpoints and all that, Yeah. I mean, the last thing you want to do is, you know, if if you're uh recently on on on the internet, there was like a meme that went around uh where somebody went onto the Chipotle website and they were asking Chipotle's chatbot uh how to write Python and it was telling them, right? And then at the end it was like, "Do you still want to order a burrito?" Like, you know, the last thing you want as an organization is to have a chatbot and people are like, "Tell me how to make a bomb or something like that." And and all of a sudden it gives them that information. That's a big deal. But that's increasingly what attackers are trying to do for sure. Um and then once they've done that and now the model is regurgitating that information, well now anybody using it is getting this toxic content or this poisoned content and that's that's where everything starts to break down. So yeah, we um we are in uh we are in a very weird time right now. And I think at the end of the day, what most people should be doing is sort of being definitive in their decisions, right? pick something to get the visibility you need. That's number one. Um and if you if you have trouble with **[00:46:00]** that, there are all kinds of people out there that can help you do it. But you got to make that decision so you can move on to now, okay, how do we then go build the policy we need to secure all of these things? ubiquitously and then once you've got that in place then you can start building your autonomous security fabric. Um but until you until you make that first step you're going to be stuck. So yeah I I agree. There's so much we don't know we don't know about all this that even the people who do know are like I don't know. There's this guy what's his name? He's right where do I Nate Nate B. Jones is it? Um he does a bunch of videos. Yeah. No, not be Navy Joe. He he also But no, there's this there's another guy. I forgot his name. He he does a lot of hardware stuff and Linux and Unix and then also some hacking and some like server management and home lab and bunch of videos. But he's really good at how he's done. He's done for many years. But he recently released a video where he's like this is going to be a different video. I'm just I'm here in Japan. and I just want to tell like what I'm feeling, how I'm feeling about the whole thing, blah blah blah. And he's like, I'm feeling the same thing. I'm getting overwhelmed. Like there's so many things happening so fast that we're all trying to play catchup, but at the same time, we're missing out on other things that might be more important and stuff like that. Like speaking of like openclaw and everything that's that's coming out and you just want to build stuff because you can and you're going to be left behind, right? But then he says like I'm gonna I'm gonna record more videos now of me actually just sharing what I'm learning because in all the all this AI stuff. Yeah, I know some stuff. But he's like I'm going to be humble and just like I don't know a bunch of stuff because it's so moving so fast. We don't like nobody knows all this stuff. So I'm just going to tell it as it is and not hype it as much because that's how I'm feeling now. And I thought that was that was really good cuz I think many of us in the industry are like, "Yeah, you could be an **[00:48:00]** expert in security and software and engineering and and whatever the field might be and get better every day. But when it comes to all these things that are just every few hours something new comes out, there's not one expert out there. It's just not possible and there's so much happening that it's it's just unknown. So yeah, I think it's it's very fun. I have a lot of fun with these tools all the time. And I think we can all just be mindful of how we're using them and what we could put in danger if we like expose certain tools or connected to certain tools. Um, so that's that's just what I think. But what uh before we go, what what would you recommend to anyone who's just dabbling like starting using these tools maybe for a business? Um, in terms of your expertise, what would you recommend to them so that they can not slow down their progress, but rather embrace all this that's that's out there, but at the same time do it with a certain care to be safe. Yeah, there's a number of things. Primarily what you want to make sure you're doing is finding a way to give your users and I'm I'm speaking purely from the perspective of like an organiza like even a small to medium business right find a way to to to give those users access to these tools prevent them from like protect them from themselves right um a lot of what I've helped customers do up to this point is put policy in place that will for instance allow developers to access cloud code will allow them to upload information into cloud code but if that information has intellectual property or PII or anything like that block it right so apply some DLP controls there um and then also say look you can't download anything from cloud right so like **[00:50:00]** maybe you can download a a Python file or maybe you can download an XML file but you can't download a zip file and you can't download a DLL file you can't download anything like that right? So, so put those guardrails in place and then determine like which one of those SAS applications for instance like claude and and chatgbt and those are all delivered through like a SAS vehicle but there's also things like Grammarly and you know Eureka and these things that you can install an extension into your browser and helps you figure out which one of those applications you can tolerate and that you're okay with and then anything else like just just block it like you know what I mean? So that's kind of the first one. The second one is when you're building agents, right, you you or even just enterprise AI app like LLM apps, you always want to start with the code you're writing. An example that I'll give is there was there was a recent attack that was just identified where attackers are basically spinning up Squarespace websites that look like the clawed website and they're using SEO campaigns to boost those websites to the top of search results. So users who are looking for like the VS Code plugin or something for Claude like they go to these websites and they download the the what they think is the plugin and it's a compromised plugin, right? You have to start there, right? You have to make sure that everything you're building is being built with tools that are official that have been sanctioned by your organization and that um you know are are are sort of actually lending itself to making sure that whatever code you're pushing Yeah. it's safe code. Um, and then you know from there anything else that you implement, make sure there's a purpose behind it, right? Like if I don't want my senior engineers reviewing code all the time, hey, that's cool. That's totally fine. But, you know, when you offload that to to AI, don't offload the sort of like pull request approval process. Mhm. You know, make sure a human actually has to see **[00:52:00]** what the agent did and see what the agent caught and if there's anything that doesn't look right, go verify it. Um, you know, just make sure that you you still keep the human in the loop currently because to your point like these these systems are still black boxes and we don't know what we don't know. We don't know what's going to happen with them. So, you don't want to rip the human all the way out, but make sure you're purposely including the human and offloading any of the rope work that you that you can do. And then I would say probably the last thing is, you know, make sure that you're upgrading your searchs, make sure you're hardening your VPN tunnels. Make sure you're doing all the things to protect yourself from from quantum threats because like right now what's happening is these attackers are exfiltrating this information that is encrypted and then they're going to use quantum computers to decrypt it all later. So any any information you lose now will unfortunately pay dividends negative dividends in the future. So you know just just make sure that you're you're taking steps now to protect yourself from those threats. And I think if if you do those things, you're doing everything you can right now. Um, and as we learn more, obviously there will be more things that we can work on. But, you know, don't don't try and go boil the ocean. Start there and uh and work your way work your way towards the uh, you know, uh, the the golden gates. Amazing. That that was super good advice, Travis. Thank you so much for coming on the show. And uh it'll be cool if we could talk again about more of this cuz there's so much to talk about. I think uh people will really enjoy this particular episode. Thanks everyone. And if you haven't liked and subscribed to the show, please do so with Spotify, YouTube, and Apple Podcasts. And thanks again Travis for coming along and we'll see you in the next one. Thanks. Thanks for having me.