AI Native Cybersecurity
Welcome to AI Native Cybersecurity, hosted by Shahar Peled and Anna Sarnek of Terra Security.
AI is changing how software gets built, how attackers work, and how security teams have to respond.
In this podcast, Shahar and Anna talk with CISOs, AI leaders, and security experts about the decisions they are making right now: how they use AI, what is hype and what is real, and what they think comes next.No vendor pitches.
Just honest conversations with the people shaping the industry.
Welcome to AI Native Cybersecurity, powered by Terra Security.
AI Native Cybersecurity
Former Google CISO Gerhard Eschelbeck: Why Fully Autonomous Security Is Still Years Away
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In episode 4 of AI Native Security, Shahar Peled interviews Gerhard Eschelbeck, a cybersecurity veteran with nearly 30 years of industry experience and former CISO of Google, to explore the rapid evolution of AI-driven cybersecurity and what it means for the future of security teams. He also shares how he helped create the Common Vulnerability Scoring System (CVSS), a framework that gave security professionals a common language to evaluate and communicate vulnerability severity.
Tune in to hear insights from one of cybersecurity's most experienced leaders on AI-powered defense, the future of vulnerability scoring, human-guided autonomy, and the innovations that will shape the next generation of cybersecurity.
TIMESTAMPS
[00:00:16] Introducing Gerhard Eschelbeck: Nearly 30 Years in Cybersecurity.
[00:01:32] How Generative AI Is Changing the Cybersecurity Landscape.
[00:03:36] Why Vulnerability Chaining Is Changing Risk Assessment.
[00:04:29] Google's Security Strategy and Its Position in the AI Era.
[00:06:51] How AI Could Transform Security Tools for CISOs.
[00:08:14] The Origin Story of CVSS: Creating a Common Language for Vulnerabilities.
[00:10:16] Why Traditional Vulnerability Scoring Needs an AI-Era Update.
[00:11:05] The Future of Offensive Security and Automated Penetration Testing.
[00:12:55] Connecting Source Code, Production Environments, and Attack Graphs.
[00:15:15] Human-Guided Autonomy vs. Fully Autonomous Cybersecurity.
[00:17:13] The Future of Self-Healing Networks and Human Decision-Making.
[00:18:13] What CISOs Need to Know About Their Organization's Attack Surface.
[00:20:26] Mythos and How Advanced AI Models Are Changing Cyber Defense.
[00:21:59] AI Agent Security, Sandbox Escapes, and the Need for Guardrails.
[00:24:27] Why Cybersecurity Startups Could Shape the Industry's Next Generation.
[00:26:03] The Next Wave of Cybersecurity Innovation.
QUOTES
- "AI is the tool to allow us to leverage penetration testing in an automated way." – Gerhard Eschelbeck
- "We are better off today with optionality where we take advantage of the AI-based scale together with the human decision-making." – Gerhard Eschelbeck
- "Agents are everywhere. How does an organization now manage and supervise those agents in their environment?" – Gerhard Eschelbeck
SOCIAL MEDIA
Shahar Peled
Instagram: https://www.instagram.com/shaharpeled_/
LinkedIn: https://www.linkedin.com/in/speled/
Gerhard Eschelbeck
LinkedIn: https://www.linkedin.com/in/gerhardeschelbeck/
WEBSITE
Terra Security: https://www.terra.security/
This is AI Native Cybersecurity. Field notes from security teams defending against and operating like AI adversaries. Because the best defense is a good offense.
Shahar PeledGerhard Eschelbeck, welcome to the AI Native Cybersecurity Podcast. It's a pleasure to have you. Thank you for joining me.
Gerhard EschelbeckThanks for having me this morning. Excited to be here and uh certainly looking forward for our discussion this morning.
Shahar PeledSame here. Maybe let's start with a quick intro. Would you introduce yourself to the audience?
Gerhard EschelbeckOf course. Gerhard Eschelbeck, I'm a longtimer in the cyberspace, uh almost 30 years by now. Started in my early days at McAfee, which is uh certainly a very pioneering amount of work there. Uh, then was part of the early team at uh at Qualis and had a journey over the many years, was uh the CISO at Google for uh for a number of years. And then most recently, uh the last few years been really working with the startup ecosystem, with the community, with the early stage founders, and that's why I'm super excited to be here today.
Shahar PeledPleasure to have you. Obviously, one of the most experienced and you know impressive careers in cybersecurity. Uh, very lucky to have you on our side and very happy to have the conversation, especially these days that everything is changing. Maybe let's start with the main question that can start us off. What has changed? Gen AI obviously was adopted, started to be about two and a half, three years ago. What is the main change you've seen? Not around marketing, the actual change in cybersecurity.
Gerhard EschelbeckYeah, no, I I think certainly the the creation of this uh new world that we're living in today with AI being um a main technology revolution, so to say, has created great opportunities for companies, for uh individuals, for governments uh to use data and make it a meaningful uh access to the data. But at the same time, it has created a tremendous amount of new security challenges. Uh every new wave of innovation brings new challenges on the technology side and on the security side. And AI is no different. Uh I think we are in the middle of this revolution, and right now we are seeing some of the early challenges of cybersecurity. So, particularly what has changed is the speed, the scale, the amount of vulnerabilities and risks that we are seeing is unheard of. And I think it's probably an order of magnitude uh that it changed from even looking back a year ago. Uh so I think the volume is one of the challenges. But also uh AI has not only created a larger amount of volume of issues that we need to tackle as CISOs, but it also has created a new set of issues, uh prompt injections, uh various uh specific issues related to AI front and center to our to our data.
Shahar PeledAI is an attack surface on its own.
Gerhard EschelbeckAbsolutely, absolutely. Yes, yes.
Shahar PeledUm but you you said speed and scale, and I totally agree. I think some perception in the market keeps saying that we'll see attacks that we've never seen before in mega, mega, mega zero days, and we've seen a little bit, but most of the attacks that did happen autonomous or human-guided autonomy, as we like to say at Terra, or not autonomous, these are just the same vulnerabilities we always knew at a different scale and pace, which is a challenge on its own. But correct me if you think otherwise, I have not seen an immense change in the type of vulnerabilities that we're facing.
Gerhard EschelbeckI think what has changed is the chaining of vulnerabilities. It's no longer just a single entity of a vulnerability that's being exploited, but it's very often a chain of individual vulnerabilities that may not be seen as uh all of them individual, high-profile vulnerabilities, but by combining all of them together, you really create a scenario where you have suddenly a new set of vulnerabilities that wasn't discovered before, wasn't found before, which is really related to the fact that the mindset has shifted from purely the severity of a single vulnerability to the exploit uh attack graph that really is relevant now for looking at the vulnerability set of vulnerabilities.
Shahar PeledMaybe that's gonna change soon and we'll see some new things we have never seen, but I don't know, yet to be determined.
Gerhard EschelbeckTBD.
Shahar PeledSo until a few years ago, you were the CISO of a small company called Google, the search engine.
Gerhard EschelbeckYes, indeed. Yes, indeed.
Shahar PeledOne of the largest and most influential companies in the world. Also one of the largest and influential companies in cybersecurity, now especially after the the Wii's acquisition. One of the most, you know, the most interesting parts about Google that I at least see is that they're the only company who holds the entire value chain of AI from the data centers to the inference to the chips to the models to the application layer. You know Google security better than 99.99% of people in the world. Um let's start with the fact that how do you think Google is positioned in the cybersecurity world, related and unrelated to AI? Do you think they're gonna emphasize more on cybersecurity, not just for themselves? That's something we can talk about after, but they're gonna go into the markets, the cybersecurity market stronger than they've been before. Waze has been a huge move, but there haven't been so many moves before that that they've done in cyber in the past few years. Do you think cyber is a big, uh a big chunk of what Google is gonna expand into the next few years?
Gerhard EschelbeckWell, clearly Google is uh is a huge target. And uh even the years I was the CISO at Google, we uh certainly were um all the time under attack. We were a target uh all the time. But the reality is Google has built this incredible internal proprietary infrastructure, but not only the infrastructure, but also an incredible team. Uh the security organization is the one of the most powerful organizations I've ever seen. And uh thanks to you. I had the opportunity to be part of it and I'm very excited to be uh to be part of it, or was part of the team. I think what's really unique and uh and interesting in in Google's philosophy is this ability to look forward, uh to predict what's the next move gonna be. And we worked on some really uh interesting projects at the time from uh Zero Trust, uh Passwordless. Uh some of those initiatives came into the open market just many years later, but really we were pioneering some of those early work, and I think that's what makes Google very unique in a way from a security perspective and from a security architecture perspective as well.
Shahar PeledI I agree. That's it's really fascinating. And and having them touch so many of the areas in the full value cycle in one company is I think they're positioned very, very well.
Gerhard EschelbeckIncredible, yeah.
Shahar PeledAnd specifically, if you can double-click on Google security, which you know very well, do you think something is changing, has changed or will change soon because of this new era of AI that Google is in and the full value chain they have? Because you know, physical security needs to be part of it, right? And then the GPUs, and then they inference the application layer. Each one of them is such a complex ecosystem on its own to protect. What do you think is changing?
Gerhard EschelbeckYeah, I think the models that uh that we see of from from any of the larger organizations will be uh gearing and starting uh towards taking advantage of uh the scale of compute uh and using it to help security. I think it's gonna be tooling that's gonna be available to security organizations, to CISOs, to companies. So I think that certainly is an opportunity that we should not underestimate. Even I don't think Google will become a security company necessarily, but the tools that and the tooling that they will be providing will help many of us, many of the CISOs in the world to uh build better security for organizations.
Shahar PeledOkay, that's a good point. I didn't think about it this way. Um let's move over to the next one. Many people don't know that you're one of the creators of CVSS, common vulnerability scoring system, that everyone in cybersecurity knows. It was so needed when you built it. Like we needed a way for everyone to score severities of vulnerabilities.
Gerhard EschelbeckYep, yeah.
Shahar PeledIt was very challenging. So before I talk about what CVSS means today, and I want to talk about that, talk to me through like how you built it, what led you to building it, um, what's a what's possible behind the scenes. I guess everyone in my generation in cyber would love to know the story behind one of the most known terms and then let's say policies or whatever in cybersecurity.
Gerhard EschelbeckYeah. Well, it started as a grassroots effort, really. Um, and we were seeing the challenge coming our way when different vendors and different partners uh shared vulnerabilities, disclosed vulnerabilities, but there was no common way, no common framework of sharing the severity of a vulnerability. So we were somebody was talking about high, low, medium, somebody was talking severity one, two, three. So you couldn't compare it as a C so what does it mean actually? How do you actually compare all these vulnerabilities? And which ones are the ones that you really need to patch today, and which ones you can wait for the next week, maybe. And so that really created the need and the opportunities. Getting together and with the three of us, we uh sat down for uh for some time and discussed and came up with the initial standard definition of CVSS, and it was a it was just a great way of suddenly having a framework in place that we all could have the same language. It's like almost a language that we use to to share uh vulnerability, severity, and it it really, I think it changed the world in many ways.
Shahar PeledUm it did. And to double-click on that, by the way, I would love to have been in that room to think about like to create something that literally everyone in the cybersecurity world has been using for so many years. It's to me, this is one of the most impressive and special things you've done from the very, very impressive resume. Obviously, CVSS now has been in many conversations and might be out of date, although there are new versions of C VSS coming out all the time. Do you think CVSS is still the right severity mechanism, severity scoring mechanism for today's day and age? And if not, what do you think should be?
Gerhard EschelbeckI think the uh the value of uh having a common language is still very appropriate. That's still needed there. But I think what has changed is um we are not just talking about severity anymore. Uh we are really talking about uh uh looking at attack graphs uh and not just at an at a single individual vulnerability. So the chaining of vulnerability that concept is not existing there in uh in C VSS necessarily. So there is a couple of concepts that I guess are missing. Uh, and so I think it's an opportunity to uh either create a revision of CVSS or create an entirely new system that's built in the AI world today. And I think that's the opportunity that we have in front of us here.
Shahar PeledWould you want to be a part of it?
Gerhard EschelbeckLove to be.
Shahar PeledThat could be a fun project. I'd love to do it. I think I think it's time. What do you think the next generation of CVSS with the same name, but for the AI or it's a totally different system. I think it doesn't matter, but I think cybersecurity needs an updated version of scoring severities of vulnerabilities. Yes, indeed. Yes indeed. Let's talk about offensive security for a minute. Offensive security, pen test, red team, ethical hacking has been an adjacent part of cybersecurity for many years. It was a very small spend versus other areas, although not cheap. But it was a project-oriented, service-oriented value proposition, point in time, highly, highly scoped to a small part of the attack surface. No organization was truly able to test a majority of the attack surface, doesn't matter if it's the web attack surface, cloud, infra, AI, data, um, and especially not more than once a year, once a quarter. This has changed. It's changed in a way that we believe that more and more and more cybersecurity professionals, Visual and CISOs are moving away from defensive to offensive, from reactive to proactive because vulnerabilities on their own are just not manageable. And there is one question in cybersecurity we need to ask, which is had adversaries attacked me today, what can they achieve? Which is why we believe it's time to change and move towards proactive mechanisms that actually do this mimicking of adversaries. Um, but especially now that the attack surfaces change on a daily basis, sometimes hourly basis, whether it's through code changes, whether it's through new CVEs, new AI models and capabilities, misconfigurations, whatever. Um do you believe in that, in that change? What do you think is going to happen now that offensive security is actually a category with technology?
Gerhard EschelbeckYeah, I think the biggest um the biggest opportunity that is in front of us here is the uh the changes that uh any change that you make in your code base today, uh as small or as big as it may be, has some immediate and direct impact to your production environment. So that the whole chain is necessary to look at it today. You cannot just look at the production environment only or at the source code only. You have to really look at the whole chain to understand kind of attack graphs and attack patterns. And the ability to do that, I think that's really what uh AI made possible today. I mean, AI is the tool to allow us uh to leverage uh penetration testing in an automated way and do it uh regularly, frequently, automatically, mostly autonomously, partially autonomously. But I think the reality is we have some real opportunity in front of us here to change the way how penetration testing has been done in the past and how we're gonna do uh using AI in an automated way, leveraging especially the fact that we have now uh the automation capability that we didn't have before.
Shahar PeledLet me double-click on the automation part. Everyone's talking about autonomous, semi-autonomous, fully autonomous. Eventually, there are some areas in life that AI is not yet ready to overtake autonomously. Maybe because the tech is not there, maybe the user or the buyer will not let it. A good example I like to give is a way more autonomous. It's autonomous to some extent until an edge case happens, and yes, they're trying to decrease the edge case as much as possible. But what happens if there is, God forbid, a car crash or there's a police stopping? Someone is gonna take the wheel behind the scenes. Probably not someone in the car, but someone at the back office talking about heart surgeries, God forbid. The technology is not yet there, but I also don't think the perception is there. I would not want this to be fully autonomous because the cost of mistake is catastrophic. I think the same in cybersecurity and more specifically in offensive security, because you're doing offensive actions that the cost of a mistake could be catastrophic. Do you think the world is going towards fully autonomous everything in cybersecurity? Do you think there is a human-guided autonomy on the way, which then you talk about scale? How do you envision this?
Gerhard EschelbeckWell, look, the the Waymo car didn't come overnight either. Uh, it took more than a decade actually to develop it to even where it is today in an autonomous way. Of course. Uh I think the reality is we also need the same kind of evolution in the in the automated pen testing, in the automated testing. So security, cybersecurity, fully autonomous, uh is not a reality today. I would love to get there. I would love to get there in the next decade. Uh, but I think today where we are is is really kind of human-guided autonomy. That's where kind of autonomous vehicles were like five, ten years ago, and that's where I think we are in the cybersecurity space today. And I think uh that uh scale and that ability to automate combined with the human decision making, I think is the most powerful combination. So I think there's a lot of hype in our industry today about fully autonomous and this and that. But I think the reality is we are better off today with uh optionality where we take advantage of the AI-based scale together with the human decision making. I think that combination is very powerful. And over the years, it will clearly move towards a more automated system where the human is less and less uh involved in the decision making as well.
Shahar PeledI couldn't agree more, and I love the equipment. Let's take you said five to ten years, and Waymo took ten years to build. Yes, the technology now is much better than what it was ten years ago, so it should be faster and incremental. But again, it's still not fully autonomous. So I think if I if I can push back a little bit and ask this question, do you think it's ever going to be fully autonomous? Like no human guided autonomy. By the way, I don't have the answer. It's it's a it's a conversation I love to have because what it means to be fully autonomous, like zero human interaction, no someone behind the scenes, no someone clicking yes or no in edge cases, full agents making decisions on the most sensitive actions and systems in the world where a cost of one mistake is catastrophic, and then obviously accountability becomes the next question. So do you think because Waymo is not 100% autonomous yet? I don't know any.
Gerhard EschelbeckIt's the corner cases, really. I think those are the cases where we continuously need the human decision making, the the complex decision making, uh which really requires the human involved there. And also safety is also a big, a big topic as well. Uh so I think we will be kind of moving towards a model that's almost fully autonomous. Are we gonna get ever to a self-healing network? Uh a fully autonomous self-healing network, would love to get there, but it's probably a vision goal that uh will take a lot of a lot of time to get us there.
Shahar PeledI agree. I agree. Moving us forward, um, I talked about the question that I love to ask that I think is the holy grail of cybersecurity. If I'm attacked today, what can adversaries achieve? As a CSO now, but CISO in the past many, many years. What picture can you, let's say today, if you're asked to present to a board, what do you show them in that slide? What can happen if adversaries attack today?
Gerhard EschelbeckYeah, I think the as a CISO, the biggest challenge we are facing is the having a complete picture of our environment there. Every organization that we are running security for is rapidly evolving, rapidly changing, continuously changing. And how do we actually have a complete picture of what I have? What is my environment looking like? And I did a study uh a few years ago where I looked at why do companies get hacked, why do companies get attacked? Um, and the successful attacks that we all know about that we are reading about, that I did take a look and studied kind of the reasons, and there's two primary reasons that I found. One is uh user error, um missing multi-factor authentication, for example, so user gave away the passwords. That's an easy target, uh, easy to fix as well, but it is one of the targets. And the second one is uh not knowing what you have in your environment and a system that got stale and got hacked, uh, for example, uh, is one of the other uh main root causes why companies got attacked and targeted. So I think it's uh there's a real opportunity as a as a CISO to have a um a completeness, a complete picture uh of your environment. Uh and until we have that, it's gonna be a challenge. It's gonna continue to be a challenge there.
Shahar PeledSo, what I'm hearing is there's no single dashboard that anyone can show to answer this question.
Gerhard EschelbeckNot really, not really today. We'd love to build it as well, but that's the next startup opportunity.
Shahar PeledWe're on it, we're on it. Um okay. Before we get to the closing part of things, um, Mythos, elephant in room. And in the same same question, I'm gonna talk about Hugging Face as well. That happened a few weeks ago. We're filming at Black Hat. Um take these two different mega hyped things, right? Mythos as a model, mythos-like models, um, hugging phase as an incident with the which happened with OpenAI. Um I can split the question into two. Do you see mythos-like models as a real significant change in the AI ecosystem with cybersecurity, or do you see it mostly as a lot of noise that hasn't really made too much change on the ground?
Gerhard EschelbeckI think it had tremendous impact already, even the last few months. Let's let's face it, uh, we have seen a large wave of uh security patches that uh have been released by organizations, largely driven by some of those language models that have been released. And I think Mythos is one example, but there is many others as one, and I think that's probably part of what we have today. What I think those tools do, they bring balance into this inequality between the attacker and the defender. They really kind of uh neutralize the kind of advantage that the attackers had in the past. Now the defenders have the same set of tools and the ability to really uh look at uh uh cybersecurity in a holistic way in their environment.
Shahar PeledOkay, I agree. I I think yes, it's a stronger model. Yes, he made a lot of changes and he found very, very interesting things and impressive. I think the hype is a bit overhyped. Let's say it is.
Gerhard EschelbeckOf course, of course, there's always media attention, is always important, and I think that's probably what it is there. Yes, yes. And I think people also starting to wake up towards the cost uh factor of it. Uh it's not cheap, uh, they're super expensive. Uh slow. Slow, expensive. Uh and I think that's also thing things that people will be waking up. But like with any new things, there is always a tremendous amount of wave uh of uh media and attention. And I think the reality is it will be balancing out over time. But I I do believe that the benefit of um having the attacker and the defender on the same level is is is a is a huge a huge advantage.
Shahar PeledI agree. And if as The same thing about the Hugging Face incident with OpenAI. For those who don't know, the agents get a task to solve some tasks within Hugging Face. And they creatively find the way out that actually caused a massive incident and went onto a real environment and not a sandbox environment. The task was within a sandbox. And then in order to solve the sandbox, they were creative and made themselves out of the sandbox to the real environment. I think it was production. They solved it. Just not in the way that people thought they will. And I don't know if there are guard drills who could have blocked it, or maybe different tasks or prompts. I don't know. But the reality is it autonomously made its way through something that could have been catastrophic. What do you what's your take on that?
Gerhard EschelbeckI think there is this is a very unfortunate uh incident uh that we have seen here. Uh and I think there will probably be more of those. This is not going to be the last one for sure. They have been, that we don't even know about exactly. And and I think uh it is it is part of the kind of evolution of technology. I think from those, the most important part out of those incidents is the learning, and how do we put better guardrails in place? How do we actually make sure that uh uh agents don't escape? Um so I think the one of the real challenges that I see is also the the management and the oversight of the rapid proliferation of agents. Agents are everywhere. How does an organization now manage and supervise those agents in their environment? It's one of the bigger challenges as well.
Shahar PeledI agree. I think it gets even more challenging when you use agents from that are based on different models from different providers with different guardrails, then you build different tools and skills that could be non-deterministic as well. And some of these tools are also open source, and some of these skills can it get it gets very, very tricky, very fast, yes. It's just another full ecosystem of supply chain just within AI, and it's just starting. To finish off, um, I always like to ask my guests uh prediction. 12 months from now, and I'm gonna go back to the archive to the archive in a year, so let's see what happens. Um, what is one thing you believe would happen with cybersecurity and AI? And your peers today will disagree with you.
Gerhard EschelbeckYeah, I mean I I think the biggest um the biggest uh opportunity we have is is following some of the early innovators in the space, looking at what is what is happening on the ground right now in the startup ecosystem. And I'm I'm a big believer in the startup scene. I'm uh um following many, many folks that are starting companies in the early stages there. Uh you are one of those folks that have built a very successful company in the penetration testing uh space, you know, autonomous penetration testing space. And I love what you have done for your industry. And I think there's many others that uh are to watch for today and look for today, what they are doing for the industry, for the cybersecurity space, so to say, in a in a way uh where they build the next generation of technology. It's the innovators to watch for. And is it gonna be the fully autonomous network? Is it gonna be the self-healing network? Is it gonna be passwordless? I don't have the answers for all of those questions, but I know that innovation comes largely out of the early uh startup ecosystem, and I think that's what we should support and and uh and build upon.
Shahar PeledOkay, so if I have to double click on that or take out the specific point that I heard, you think that because of all the changes that happened, yes, they're cyber giants, but you think that the next wave, not necessarily of innovation, because innovation always comes from comes from startups, but the next wave of large companies might not necessarily the existing companies we see today, but maybe the startups.
Gerhard EschelbeckThat's exactly right, yeah. Yeah.
Shahar PeledInteresting. That's that's a hot take.
Gerhard EschelbeckYep.
Shahar PeledI think the cyber giants will stay cyber giants. Maybe the cyber mediums are not positioned for the future well enough, and there'll be an overtake from the startups that they can become cyber giants.
Gerhard EschelbeckI think it also depends a little bit on some of the external factors. Like three years ago, nobody predicted this rapid growth of AI. So I think AI by itself will create and has created a whole set of new companies, a whole set of new cybersecurity companies. And I think uh that's why it's very hard to predict what we're gonna see in five years from now. I think the next year is pretty clear. It's the early stages that are continued to build out and be successful, and I hope you're gonna be the next large or one of the next largest uh as well.
Shahar PeledSo too. Uh Gerhard Eschelbeck, thank you so much for joining me on the AI Native Cybersecurity Podcast. It's been a pleasure, super interesting, a lot of hot takes, and uh I appreciate you having being here.
Gerhard EschelbeckSame here, thanks a lot, and uh looking forward continuously working with you.
Intro/OutroThat's it for today's episode of AI Native Cybersecurity. Your attack service probably changed while you were listening. So take something from this conversation and go test it. We'll see you next time.