A (3:24)

Move to the second layer. I just want to make sure I'm understanding this is a research competition. Was it that there were six AI applications that were developed and of those six, they were five that were so vulnerable that a bad actor could use those vulnerabilities to remote in and run this software any way they wanted to?

B (3:44)

Correct. There are six technologies that are used to build AI technologies like Nvidia, like PostgreSQL, these are Redis. These are common technologies that are used to build the application. Out of the six, four actually had critical vulnerabilities at the highest stake. And by the way, AI has the most vulnerabilities now that are being disclosed. Wiz Research itself competed in this pwn to own contest. It won the first place. But we need to understand that this is an area of active research. It's a new technology, hence it has a lot to do with the maturity to it in order to get to the level of, let's say, trust in the software.

A (4:25)

So it wasn't just an output of an AI enabled technology that was vulnerable, it was the actual building tools that are used to build AI itself or vulnerable.

B (4:38)

Exactly.

A (4:39)

All these companies, all these engineers who are relying on artificial intelligence tools because they're so new, they may not know it, whereas what you're claiming that they may not know it, but bad actors could be basically hacking into the code that they are writing with these tools and then not just controlling the tools themselves, but the outputs as well.

B (5:01)

Exactly. And this is what we see. It's scary. It's a new technology stick. It's being built, it's now maturing over time. We are learning about it, we're securing it, we're improving it. But we also need to remember that as with any technology, this is new software and it's now being out there tested by pen tester, but also by threat actors. So that's one. It's a new technology stick. It's important to understand, and that's the most important thing to understand about AI. It's software like any other software.

A (5:31)

Five minutes in you've already told me about something that I didn't even anticipate coming in would be a problem. Which is it's not just the code, it's the foundational tools that are used to output it.

B (5:43)

Correct. You're bringing new software.

A (5:44)

Okay. And is the second thing that you're about to go on to the unreliability of AI produced code itself?

B (5:50)

Not even that. First we're going to look at the infrastructure. Right. Because in the end you're running these new tools, but they're running on top of infrastructure like any other application. You have identities, workloads, you're using the basic components. You are storing your training data sets in buckets that can be now publicly exposed. You are using identities that can be overly permissive. You are using VMS or containers that can be also compromised or misconfigured.

A (6:18)

So just to translate to a non or less technical audience, all these. The infrastructure that you're talking about is people are building AI programs and then they are relying on the cloud.

B (6:31)

Correct.

A (6:31)

All the tools that are used to support basically anything that runs online. Software that runs online. And there could be vulnerabilities there as well.

B (6:39)

Exactly. Vulnerabilities, misconfiguration.

A (6:42)

And in fact, you have a program and it's running off of a cloud system, but you're exposing a huge part of the way that your program runs without knowing it.

B (6:55)

Exactly. So if we look at incidents around, let's say, AI applications, we had an incident where a very large, I would say software provider exposed a bucket through with all of the training data sets in there, a lot of sensitive data, and they didn't intend to expose it. This is a very basic security issue that we have experienced in the past decade now applied to AI. So again, when we think about AI, it's not all new, it's new software, it's existing infrastructure. And when we think about securing it, the basics apply. And it's very important to remember the fundamentals apply. The fundamentals of patching vulnerabilities, securing configurations, managing identities and so on. So this is like the second layer that we think about the infrastructure.

A (7:44)

Okay, and can I ask one more question about that? So basically what I imagine a lot of organizations are doing right now is when they want to build personalized AI, personalized AI tool tools for their company, they're saying we want to build something specific to us that does what we do that can, you know, maybe replace our employees or augment our employees. Let's just download everything from the organization and will throw it into the bot. And they've been promised by some API provider that the bot will not use that data for training or it won't spew it out elsewhere. But the thing that they miss is as they download all this important secure information from their company, they might store it somewhere.

B (8:19)

Exactly.

A (8:20)

Because where else are you going to have the treasure trove or this like motherload of data other than when you're starting to train a bot specified for your company?

B (8:28)

Exactly. So if I'm a threat actor, why wouldn't I use a good old techniques I'm using to expose buckets, databases, misconfigurations in order to exfiltrate this precious data? Why would I go through complicated AI prompts and so on if I can go directly at the infrastructure? So that's a second layer that we have to be mindful of. And these are all best practices that we apply and we follow, but they apply the same way on our AI applications. And actually when we look at AI related incidents, the majority of them is within this layer of using the infrastructure in an insecure manner that allows threat actors to just do what they are used to do in cloud in the past.

A (9:13)

Okay, and now let's circle back to the question I asked in the beginning because you know, we've been building up to it, which is how vulnerable is AI written code.

B (9:20)

So AI written code is interesting because one, yes, it may be more vulnerable because we haven't instructed it to be secured. As we use AI to build applications, we need to also same way that we instruct it in what we want the application to do, we need to instruct it in the way we want it to be secured. Right? Apply list privileges, remove data if not in use. So there are all of these best practices in security that we apply today. If we have a human developing it and we review it, but with AI, we need to now specify the same manner. And as an example we have released, the research team have released, it's a rule set that you can feed into AI, generate generator, like code generator through AI. And this rule set will guide AI to build a secure code more than if you didn't. And this is just one aspect. But the more interesting thing around securing code that is generated by AI is what happens if you actually need to fix it. And who is the owner of this code now? There is a very interesting question that arises because if I'm the developer of the code, and now there is a vulnerability of someone reported a security issue. I wrote the code, I know it by heart, and I have someone reviewed it. But let's say I vibe code my entire application. And it's funny because on the first blog post introducing vibe code, the person who posted it said, you can forget that code even exists, but no, you cannot. The code is vulnerable. And now if something happens, you need to get back to it and fix it, but you need to know the code to do it. So there is a broader questions on responsibility around the code that is generated through AI and who is going to go back and fix it if something happens. By the way, it's not only security, also reliability, availability, scale. How do we assure that we have the proper capability to maintain the software that we have shipped across? Security, availability, reliability in the long term.

A (11:36)

Okay, so I think that what you're saying is basically AI can build pretty secure code if you introduce the instructions or the protocols to build secure code into the prompt.

B (11:48)

Correct.

A (11:49)

That seems fairly. That seems good. But then the other side of it is what we're seeing now is engineers are starting to be removed from the code that they are writing because oftentimes they are shepherding that code process. They're vibing it, they're vibing it and the AI is doing the rest. Are we already seeing cybersecurity problems within companies who have had developers that have just vibe coded or AI coded applications?

B (12:19)

Yeah, actually there are very known examples of some people that have posted, they vibe code an application and then it got hacked hours after or days after, and now they don't know how to recover because they didn't have the skill. So the way I'm looking at it is that vibing code is a great way to accelerate, but it doesn't remove you from the responsibility of actually knowing your code, being able to address issues within the code and guide AI farther into the maintenance process as we go and use the application and mature the application. So this is, I think, a maturity thing that we need to go and to do. And another aspect is when we think about how would an agent architecture work, because you are talking only on one role, which is the developer. But if we are fast forwarding agentic concepts into the future, so why wouldn't you have a security reviewer that is also an agent? So you commit code, you are developing your code, and next you should have a security review to your code, but also performed by an agent. That is minded for security with all the security best practices guideline that we have provided it. And someone should maybe look at the architecture and someone should look at data privacy and you can think about. And again, it doesn't exist today yet. But as we fast forward, AI doesn't mean that it's only doing the coding. It can do other things in the development lifecycle that we can rely on.

A (13:55)

But this is scaring me even more because I thought when you said, okay, people are going to vibe code, by the way, vibe coding is coding just by prompt. So you say, build me this application and it builds an application as opposed to just doing the code yourself, I thought, okay, so now we'll have defined roles. We'll have. If vive coding becomes a thing, people will vibe code applications and the role of the developer will be to audit, monitor, look for vulnerabilities there and address them. So that will be their expertise. It's not actually building the things, it is securing them and maybe improving them in some ways that the bots can't handle. But what you're saying now is we're looking towards a future where there are going to be agents like another AI bot whose role is to do that thing, to secure the.

B (14:42)

Exactly. Why not?

A (14:43)

So then, isn't that like a triple risk then, or a double risk? Because you're now going to have AI bot whose core competency is to build and AI bot whose core competency is to secure. So if something goes wrong now, we're definitely not going to have a human with the skills to be able to diagnose and address.

B (15:01)

So it actually doesn't conflict with each other. We can accelerate or automate more, but it doesn't actually remove us from the need to have, let's say, a human in the loop that we will be able to then account when something happens in the details. So there are two separate questions. That's why when we started, I said, yeah, vibe coding can be improved. It can be bad, but it can also be improved in various ways. But also it doesn't resolve the actual need to say who is the owner of the code. What is the operating model? When something happens, and let's say AI cannot fix an issue. So who does? And who owns this? Over time, and this is maybe the most I would say when I look at the challenges, technology will continue and accelerate and help us to build applications. But still we have to figure out what is the operating model, that when something happens, we know who to turn to and we know that we will be able Think about it. You are an enterprise, you are developing an application. Forget about security, let's say basic reliability. The application goes down for whatever reason. You cannot just wait and rely on, maybe AI will figure it out, maybe it doesn't. Right.

A (16:17)

We need to seems like a bad strategy.

B (16:20)

So we need to have accountability in the end. Okay, so how do we maintain this? And if I'm a developer and I shipped an application and it goes, doesn't matter if I built it myself or through AI, but it goes down tomorrow and I'm not able to fix it and get it up again, I'm not doing my job right. So this is a big question. How do we continue to own the application while leveraging increasing amount of automation?

A (16:45)

Okay, so we've talked about the liabilities and the tools you use to build AI liabilities in the infrastructure used to store key information for AI, and now liabilities in AI code and vive coding in particular. But there's another side to this, which is I love the fact that people are able to build things just by a prompt now or maybe a little bit more sophisticated. They prompt, they code a little bit, they build off each other. And now something that, that might have taken a team of 10 can be done with one person. But we've called them threat actors. I call them bad actors, basically bad guys looking to do damage or looking to hack into computer programs. They have all these tools at their disposal as well. So are we already seeing them putting those tools to use in an attempt to hack into software? And has that increased the sophistication and the level of hacking that we're seeing already?

B (17:38)

So there are several ways in which threat actors can use AI. Okay. And the first thing is just hacking into an AI application. Right? And if you think about how do I hack into AI application and why it's appealing, there is what's called like the trifecta of, let's say, risk factors in AI. One, it's exposed. Second, it has access to private data. Third, you have untrusted content like the query, the, you know, the chat queries that you send to it, the prompt that is also exposed to the threat actors. So one, you have the layer of directly aiming at the AI application and trying to extract via the prompt, sensitive data. So that's one. The second thing that you can do, you can automate and iterate more on what is already known. And I think this is one thing that we look at AI, it's really good in automating repetitive tests. So instead of trying one type of attack or one type of vulnerability I can automate and iterate through an AI like built, like purpose built application to try and test many more options. Right? And then you have the third layer that am I able to discover new type of threats using AI? Am I able to do vulnerability research to find new vulnerabilities? Because I've trained AI to do something very specific in that case. So these are the three levels that we can look at. And I think that the interesting thing is, and let's try to take a little one by one, when I'm targeting directly the AI application, it's like any attack surface we need now to understand how to secure AI applications. The prompt, the data within the model against threat actors and that's almost like application security applied to AI. Whole new stack, whole new type of applications, still a lot to learn how to secure it. And yes, threat actors can do it today. I think it's very available for them today. Any, we should think about it that any prompt that we expose is going to be tested also by threat actors. So that's one second layer, the automation. This touches a very, I would say the deep issue with security, which is the attacker to defender asymmetry. And if we think, you know, historically, just for those who are not familiar, the asymmetry means that a defender has to defend on all fronts, all of the time, while the attacker has to find only one thing that actually works in order to get in. So the asymmetry is insane. It means that the more state we have, the more we need to secure as defenders, the harder and harder it becomes because we need to secure everything out there. But the threat actors can find one vulnerability and they will still get in. So historically we've become better and better in improving the ability to secure the foundations, remove the risk, proactively detect and respond. This is where security has been throughout history, okay, in improving our ability to cope with this asymmetry and now with AI. The interesting thing that the threat actors can automate a lot more, but from a defense perspective, it doesn't give me the same order of magnitude of, let's say, improvement that the threat actor can gain. So in essence there is an aggravation, a significant aggregation in the asymmetry that we are going to face. Because I can test the threat actors, I can try more as a defender, it doesn't help me to detect more at the same order of magnitude. So this is a challenge that we are going to see more and more, more automation on a threat actor sign. And the reason is that from a detector detection point of view, from the defender point of view, I cannot withstand a high false positive rate. From the threat actor's perspective. I don't care about the false positive, I just need one from the detection. If I have a high false positive rate, I'm done. I cannot find the hay in the stack because let's say even if I have 0.1%, 0.001%, you take it as much lower as you want, just multiply it by the attempts of threats that you get, you're going to be bombarded with noise. And noise is the enemy of security.

A (22:09)

They can like DDoS, your infrastructure, pretty much with attacks.

B (22:12)

DDoS, your security team. Actually the security team.

A (22:15)

That's what it is. It's like a physical, like we're just going to. DDoS is basically you send a bunch of traffic to your site and take it down. So go ahead.

B (22:22)

Yeah. On run security, I can tell you a nice story, not technology oriented. One time I walked into a building and I saw a new employee training for the guards of the building and they were just next to the fire, like fire alarm. And then the one who was giving the tour was saying, if the alarm is sound, if you hear the alarm, this is a fire alarm. This is where you go, you turn it off and then the alarm stops. And then a lady asked from the new employees, and what if it's a real fire? And he told her, listen, it's not a real fire, it's a false alarm. Why do I love this story? Because this is exactly the challenge the security teams have to face with false positives. If we are not going to be really accurate in detecting the high fidelity alerts and really making the extra mile to make sure that if a security team sees something, it's actually a thing they need to take in very fast, we are becoming unable to cope with the real threat.

A (23:26)

So this is the risk is that if you automate, if you're a threat actor or someone trying to hack into a system, you can automate it. You can overwhelm a security team because every false positive they need to put attention to and then eventually you get past their defenses.

B (23:39)

Exactly. And this is why there is a good way to tackle this. It's actually reducing the noise by investing more in the fundamentals and less on the detection. Okay. So if we are always trying to detect, it's going to be hard. We're going to be bombarded with alerts. If we're proactively reducing the risk and the chance of being attacked in the first place, namely, patching vulnerabilities, fixing misconfigurations. We are in a much better position to turn down the noise. So this is where security has invested a lot of time and now it's going to aggravate if AI is being applied by the threat actors.

A (24:16)

Who are the threat actors?

B (24:19)

Everyone. AI is very accessible as a technology. We need to understand that. It's accessible, it's simplified, it can be used at scale. It's not very expensive. You don't need to be a superpower to use it. You don't need a lot of funds to use it. To be honest, it's accessible.

A (24:39)

So is it like organized crime, governments.

B (24:42)

Everyone, state nations, organized crime, teenagers in basements. Teenagers in basements, yeah, it's easy. It's accessible. Why wouldn't you do it?

A (24:49)

Well, you could go to jail if you get.

B (24:50)

No, I'm saying as a threat actor, why wouldn't you try to use AI?

A (24:54)

Same answer. Why wouldn't use AI? If you're, if you've, if you've already decided to do something, if you are already a threat actor, of course you're.

B (25:00)

Going to use AI, Right?

A (25:01)

So that's why I'm saying right up.

B (25:02)

As a threat actor, it's a new tool in your stake, and now you can leverage it to accelerate to automate. I think there is a thing about cybercrime. It's. It's a business, right? And we don't, I don't know if we always. It's not, it's not just bad guys. These are like businesses. If you think about ransomware, it's a business. It has rules to it, right? You know, when you pay the ransom, you get the data, right? It's a rule. You're not tricked into it because it's a business and they need to maintain their reputation. So as a business, same way that any other business is looking into AI and thinking, how can we use it to accelerate to automate cybercrime? State nations, they have the same logic.

A (25:44)

Today in the year 2025. These bad guys, teenagers in basement, whatever you want to call them, businesses, they've had access to pretty sophisticated generative AI for a couple of years now. So talk a little bit about the curve that you've seen since the introduction of ChatGPT. Are we seeing about the same threats like we've seen previously, or has it.

B (26:04)

Escalated exponentially currently from what we see and we are seeing, and we are releasing like the, the state of AI and we're monitoring for threats there is a site threats with IO that we monitor all of the cloud AI related incidents and we do a breakdown of basically what was the root cause, the techniques. To be honest, we're seeing today more of the same. It's not that we are seeing a significant shift in the threat landscape. It's really turning back to the infrastructure. The things that we know that are.

A (26:36)

Working so now, the same vulnerabilities, but the magnitude of attacks, have they gone up?

B (26:40)

Always. They're always trying to do more. But you need to remember that also on the blue side, on the defender side, we are really getting good in protecting our infrastructure. And I think that we're seeing a trend and it's always easy to talk about the increasing rate of attacks and the red side. And I think it will always be that way. They will always try more. It doesn't mean that it will be more successful. Because on the blue side, we have major transformations in how we look at security that is helping us to improve the foundation, improve the processes, improve the ability to proactively reduce the risk, the ability to detect and respond. And I think that we are seeing like, since we are still at the foundations that we know how to secure, I think that we are seeing that we are not at the phase that you are saying where we are seeing crazy stuff happening.

A (27:35)

I'm not saying it, I'm just asking about it.

B (27:37)

No, not yet. Yeah, we're not seeing the crazy stuff.

A (27:39)

Why do you think that is? Because I think that with every conversation around generative AI, there's always this, oh, you gotta be looking at cybersecurity part of the conversation. Because all these tools are in the hands of the bad guys now for all these reasons, but yet we're, you know, in, in reality, the actual ability of them to get through is not higher. Is your answer just because companies are doing a better job securing their infrastructure? Because if so, then it's really not a big problem.

B (28:10)

I think that we are in the process, when I think about the last decade, right, this process of automation, it's not new to us, right? Automation has taken place before AI, long before AI, and will take place after AI, long before. We are always in a journey to continue and automate what threat actors a decade ago have done on the keyboard manually. And when we look at cloud attacks, for instance, the ability to automate, what happens. As soon as I walk into an account as a threat actor, what do I do? Well, I can automate quite a bit. And we have seen this level of automation, we have seen this level of automation, for instance, with ransomware. We have seen automation happening over the course of the last decade. And. But in response, within security, we have basically developed the capabilities to respond to this automation. AI is another layer that allows us to automate more. Right now, okay, I'm not talking because we're not seeing the crazy new threats yet. We are right now at the phase where we are seeing accelerated automation of the known threats, known risks. And this is a journey security has been into in the past decade, and it's only one step up.

A (29:25)

So it's interesting because I guess what you're talking about sort of follows, I think, a lot of the progression of generative AI to date, which is that it's a very promising technology. But companies trying to put. Because if bad guys are companies, they're businesses. Companies trying to put it into action have seen mixed results. And so actually what I'm getting from you is same thing is happening with the bad actors. But then the question is, if these models get much more intelligent, does that open us up to bigger risks?

B (29:54)

I believe there are areas that, yes, if they become increasingly better. Let's take vulnerability research as an example. Vulnerability research is one area. Let's explain what's a vulnerability? Okay. Vulnerability by definition is the ability to move from one trust level to another trust level in a way that is not permitted. For instance, if I can run remote code, then I'm moving from the outside to the inside. And this is the worst that can happen because I'm literally running code from remotely external in your internal environment. Right. So that's a vulnerability. A vulnerability can be like unauthenticated access. So authentication bypass. I'm logging in and I have this trick that I'm giving false password and I'm still able to log in. That's authentication bypass. I was able to walk into a higher trust without the permission to do so. That's a vulnerability. The ability to research and find vulnerabilities. This is the bottleneck of the security space because vulnerabilities are what allow threat actors to move from low trust to higher trust environments. The ability to automate research by AI of vulnerabilities can open up maybe a race where you can find many vulnerabilities and unable to patch them at the same pace. There are solutions today that are already leveraging AI to detect vulnerabilities. These are companies, security companies, that are using AI to analyze your code and find vulnerabilities in your code that up until today, you needed a very deep research to do it. And it's automated. It's really nice. From a security perspective, threat actors may try to weaponize it to find faster and faster vulnerabilities and automate the process to put it into action. So that they can always use what we call in the industry zero day vulnerabilities. Zero day vulnerabilities means that it's a vulnerability that besides the threat actor that found it, no one else has seen it. So it's under the radar. That's the meaning of a zero day vulnerability. So if we look into what could be a risk. Yes, if we're able to create a perpetu mobile that finds vulnerabilities, operationalize them and puts it into action. It's scary, but I think that today we're again looking at the detection mechanism and so on. Keeping the fight.

A (32:37)

Okay, so we're deep into the conversation, but good time to really introduce what Wiz is. I mentioned in the beginning that Wiz has sold to Google for 32 billion. What your company does is basically just that it looks at everything, you correct me if I'm wrong, but looks at everything that an organization has in the cloud and tries to find those vulnerabilities proactively.

B (33:01)

Correct. Wiz connects very quickly to any cloud environment and assesses all of the risks to the cloud. We call them attack paths. So risks that we talked about, the noise, we remove the noise. We focus on really what are the critical attack paths that threat actors can use in order to gain access to an environment. And we surface them to security and development teams so they can proactively reduce the risks before the bad guys find them. Okay, so that's what Wiz does. Interestingly, it doesn't happen only in cloud. We can do the same thing on prem. On prem private cloud, but also during the coding phase, which is even more interesting. This means that as you are developing, I can actually predict what would become critical attack path and guardrail and let's say help developers make the right choices while developing before it gets to the cloud.

A (33:56)

Right.

B (33:57)

So this is like even preventive security. So this is. And the last pillar is like detective, this is where we don't want to be, but we have to have the controls is to monitor the cloud environment and if we are seeing a suspicious activity, detect and respond to it. So this is a bit on what Wiz does.

A (34:13)

Do you think increasing threats from generative AI played into why Google wanted to acquire the company?

B (34:18)

Security again, we're in this battle for decades. I think we are improving dramatically as an industry. I think that Generative AI increases one, the use of software. So cloud security has become even more important today because we understand that AI will be used at scale by any business. By the way, historically we're looking at cloud. There were, I call it three migration waves into cloud. The first one were the cloud natives, Spotify.

A (34:50)

Okay, so we're recording in Spotify studios.

B (34:52)

So cloud native started in cloud, born to cloud, then came Covid. Covid actually brought the second wave where businesses realized that staying on prem means they may be disconnected from their business and moving to cloud is the way to go. And they then came the second wave, right, the during COVID the major, the banks, the financial pharma, they all moved to cloud because they realized, okay, that's not the playground, that's a strategy. And now we're watching at the third wave, which is AI. So AI augments the use of cloud, it augments the use of technologies transformation into cloud and AI and it all piles up in the same direction. So just going back to the point, yes, AI increases, the use of software, increases the importance of cloud, AI security, what we call the fundamentals, how do we build secure application? So absolutely, it plays into how do we think about strategically what's important for an organization, cloud and AI security. So that's one second thing. I do think that AI requires us to innovate faster than ever before. And this is an interesting piece. We talked about the new stack, we talk about new threats, we can talk about the pace of adoption of technologies, which is also mind blowing. We are at an era that we are innovating more than ever before. In order to keep up, you need solutions that can help security to enable the business to move forward and not secure a solution that will block the business from using technologies. I think that when you look at Wiz, Wiz is really a technology that allows organizations to adopt technologies in a way that enable the business rather than blocking the business. And this will become a critical attribute for any company out there. Otherwise, AI is moving and businesses will move with AI and any other business should keep up with AI because this is going to define businesses in the upcoming years.

A (36:54)

Yeah, I mean, I think that I'll just add that obviously Google is trying to be the one where startups, anyone who's trying to build, I mean, they're seeing tremendous growth in cloud, I think like 30% quarter over quarter in the most recent quarter. But that's been the case for a while as AI has built up. And if there's a chance that these vulnerabilities actually start to show themselves and become more difficult, then I would imagine that Wiz is pretty well suited to help them out there.

B (37:23)

I think there is an interesting concept for those who are not familiar with the history of cybersecurity. There was a very interesting initiative back in even I think back in the 2000, okay, 2000, by Microsoft. And I want to just. It was called Trustworthy Computing. And Trustworthy Computing was around this concept that if humans want to be able to trust their compute, they won't use it. And that started a whole initiative within Microsoft to secure Windows, to secure all of the software that they ship. And this changed the security industry forever because this created the concept that security has to be built in in order for people to trust it. And when we think about AI, when we think about cloud, security has to be baked into these processes so we can trust anything that we leverage in terms of cloud and AI moving forward. Security is a cornerstone for our ability to use technologies at scale. And yeah, this is, I think we have to solve it right? We have to make sure we can trust cloud and AI.

A (38:35)

Clearly Google thinks so. I mean, I think this was the biggest acquisition in their history.

B (38:38)

Correct.

A (38:39)

So it tells you everything you need to know. All right, we got to take a break, but after we come back, I want to talk about what might be like the big black swan events, whether bad actors can hack into humanoid robots when we see them. And also you had a very interesting report on Deep SEQ when it came out. So we'll touch on that as well. All right, back right after this. Shape the future of enterprise AI with Agency AGNT CY now, an open source Linux foundation project, Agency is leading the way in establishing trusted identity and access management for the Internet of Agents, a collaboration layer that ensures AI agents can securely discover, connect and work across any framework. With Agency, your organization gains open standardized tools and seamless integration including robust identity management to be able to identify, authenticate and interact across any platform, empowering you to deploy multi agent systems with confidence. Join industry leaders like Cisco, Dell Technologies, Google Cloud, Oracle, red hat and 75 plus supporting companies to set the standard for secure scalable AI infrastructure. Is your enterprise ready for the Future of Magentic AI? Visit agency.org to explore use cases now that's agn tc y dot org and we're back here on big Technology podcast with Inon Costica, the co founder of Wiz, also its VP of product. Let's talk about DeepSeek right away because there was an interesting release that Wiz put out when Deepseek was sort of all the rage and it got a lot of headlines. It was here. This is from your blog. Wiz Research Uncovers Exposed Deep SEQ Database Leaking Sensitive Information, Including Chat history. Now, I took issue with this because I had just met you at the Amazon the Re Invent conference and you had talked to me about how what Wiz does is it proactively. We just talked about before the break, it proactively looks through your code and finds things that might be exposed. And I saw this blog post and I was like, well, that's what Wiz does. And you could probably write similar headlines about many different companies. And I think what happened was the news cycle ran with it and people who were afraid of Deep Seek pointed to it and said, ah, look, deepseek is trying to leak your information to the Chinese Communist Party, which is actually not what it was. So what do you think about my read on the situation there?

B (41:15)

Okay, so let's go back on the details on the Deep Seq. What happened? Deep SEQ was introduced and gained a lot of traction and interest from businesses, media, everything. And from our research team as well, looking into Deep SEQ and the research team found again, we talked about it in the fundamentals. It's not a very advanced AI centric capability. It's an exposed database. In the end, if we simplify it, that includes a lot of sensitive data from Deep seq. They closed the exposure after we disclosed it and then we published a blog post about the fact that that Deep SEQ had an open database. If we simplify it. And I think that Deep SEQ has a few Deep SEQ that week. There are a few interesting things that worth noting. One, Deep SEQ adoption was something that we need to pause and understand how fast technologies can propagate within our state. Within a week, almost 10% of organizations were using Deep Seq. That it sync okay, within a week. I think that this is just one aspect of when we think about from a security perspective, what do we need to do as we are facing this rapid adoption of new technologies? In the end, Deep SEQ raises interesting questions around. I'm going to use it to train on sensitive data. I'm going to what are the questions that we should be asking when we bring a new technology into play? By the way, it's not just Deep seq, it's any technology that we are bringing in. And I think Deep SEQ was interesting because of the rapid adoption on one hand. Second, the questions around, okay, where did it come from? It came out of nowhere. Right. Nobody anticipated it. And all of a sudden everybody started downloading the models, using it. And third, we found what is a critical vulnerability that we felt like we need to let people know about. And I think that these areas, this is really important for teams to ask the question on how do we enable safe use of technologies that doesn't prohibit the business from using them, but does instill the security measures we need in place. This is really around what we've learned from Deep seq. Technology is adopted faster than ever. We need to know what our teams are doing with the technology. And third, we need to do the scrutiny around what is the security posture of the technologies we use. Again, Deep Seq, like any other AI tools, as we talked about earlier, it's a new space, new software. It's being developed as we go. We need to scrutinize the technologies as being adopted as we have done in the past decade. It's nothing new.

A (44:26)

But my point is I feel like you guys find similar vulnerabilities like the one that you found with deepseek every day. And this release let the media run with this. There was already fear of deep seeking and the media kind of ran with it and said, see, the whole point of Deep SEQ was like a data exfiltration exercise.

B (44:47)

This is what the media takes of it. Right?

A (44:49)

Yeah. But I want to hear from you.

B (44:50)

What was really, I think that in general, we focused on the research on the entire AI stack. Over the past year and a half, we launched products to secure AI aispm, which is AI Security Posture Management. We have released the AI State of the Union, like what's being used, not used. And I think we have also researched many, many AI technolog and that we found to be vulnerable. You can cover in our blog, as.

A (45:15)

You said earlier, five of six of these foundational technologies.

B (45:18)

Exactly.

A (45:18)

So were the vulnerabilities in Deep SEQ any worse than what you typically find?

B (45:24)

It was a very, what I would say, typical exposure that exposed a lot of sensitive data. But yet, if we look historically, there were similar incidents, for instance, with Microsoft releasing a token that had access to sensitive data in a bucket used for training AI. And I think there are vulnerabilities more complex, but vulnerabilities in Nvidia's toolkits that they release for developing AI. I think what media does with it is one thing. Right. But the research shows, like throughout the past years and the research we have shared around AI, that AI really the messages that we started with AI is a new stack of tools There are software tools. They have vulnerabilities. We need to put the effort to secure it. By the way, the tools we use and the infrastructure we use, it doesn't mean that we can obviate the need to do infrastructure security the same way that we have done. And that's basically the message I think that we should all take, including from the Deep SEQ incident. This is a very common misconfiguration that happens.

A (46:39)

Yeah.

B (46:40)

Exposing a database.

A (46:41)

Having spoken with you, I was thinking that as well, and I was trying to be like, come on. But I think that it's good that you guys talked about it, but I feel like I'm looking back at the story of the context the media missed, is this stuff happens all the time to everybody.

B (46:54)

Yeah, everybody.

A (46:56)

All right, let's talk about some wild stuff now. Some more pie in the sky things. So all these AI labs are trying to develop superintelligence. They've gone from, like, AGI to superintelligence over the course of a couple of months. One of the things that you would imagine you would find with superintelligence is the ability to break into anything. I was speaking with a researcher recently who gave this interesting thought experiment that I kind of want to run by you. I do want to run by you, which is that if an AI lab built an AI system that was an expert at detecting vulnerabilities could get past any system today, no matter how good. No offense to anyone working in AI security, but the idea of superintelligence is you can surpass what exists today. Would they then have. And imagine they saw other labs getting close to building the same thing, which would eventually destroy security on the Internet. They then have a responsibility to go in and delete the competing AIs because of what could happen if this technology proliferates.

B (48:06)

You are looking at it from the threat perspective. I always try to think about it from the defender perspective. Are we able to generate faster technologies that will defend against this sort of stuff?

A (48:20)

But doesn't the idea of superintelligence then worry you a little bit? Because, a, we're not quite sure. I mean, if it happens and it's still a big if, we're not quite sure if it's going to be something that's developed by the good guys or the bad guys. And even if it's developed by the quote, unquote, good guys, there could be hidden bad stuff going in there, going on in there.

B (48:37)

It's true to any technology. Quantum computing is another example. Like, there are many new technologies and there will always be new technologies that will be introduced. And I think that as defenders and in general as the good guys, we need to think about how do we create the proper guardrails to secure and build foundations that help us to continue and adopt technologies in a way that is controlled manner and secured. Now of course you can think about scenarios that are far exceeds the worst case scenarios that we have seen up until now with security. But I think that historically that wasn't the case. And historically the same things were said about cloud, they were said about data lakes, they were said about containers, they were said about many different innovations that were introduced. And I think that we were good enough in an industry. I'll give you the positive perspective. Most of the technologies that were introduced in the past were introduced without thinking about security at all. And if we think about the sad story of why we ended up in, I would say a huge deficit of security in the industry, it's because security always came later on. It came bolt on. We invented autonomous cars and then we figured out okay, so they can now be controlled and being crashed into places without control. So we need to secure it. The concepts behind I think today AI security are actually well thought today rather than much later. And it shows a bit on the maturity of not only the security teams. By the way, it shows a lot on the maturity of the industry that we are having the discussion on what's safe AI, how to secure AI architecture. And unlike before talking about new technologies and their security was always something that was perceived as inhibitor to innovation. Let's get it right first and then let's secure it. Let's not get the folks scared about it so they won't block it. But today it's actually a much more, I would say integrated discussion. Of course we're going to use AI, but of course we are going also to put the guardrails, the security in place, the standards, we're going to ask the questions. So as we do this, we also develop the muscle of let's secure AI at the get go. And I think it actually shows a lot on the maturity. I'm positive actually about this because I think we are at a position that we've never been in it before. It's relatively new that we talk about technology and the security of the technology at the same time. It's great.

A (51:30)

I mean, I know you're in security and so I guess I'd be optimistic if I was co founding a security company. But I think you're underrating the risks.

B (51:40)

I just think that as an industry there are always risks. The risks are always scary and they far exceed what ordinary people can think about when they just see a technology. But I do think that we have strong foundations within the security industry and the development industry to build the proper guardrails. And I have trust that we have established enough track record on how to secure applications environments that we can continue and innovate. And as I'm going to talk not only about wiz, like the entire industry, if I look today at the pace of innovation around AI security, look at the number of startups that are already focusing on providing solutions to to AI security. AI firewalls, AI security posture management tools, AI security gateways, right? There is a whole array of security industry solutions that are basically creating AI aware controls that are now already in place. Already in place. It's not like futuristic, it's existing solutions. Yes, we need to mature, yes, we need to research. Yes, we need to do a lot more, but we are building the capabilities towards there.

A (52:54)

What about quantum? You mentioned quantum. Everyone talks about how quantum will break encryption and break security. You have a big smile already.

B (53:02)

Hearing the word quantum is interesting. I think it's also. It's an exercise, right? What if exercise? I think we're seeing a lot more awareness to post quantum cryptography. Right? The ability to basically use cryptography that will withstand quantum. You know, if quantum computing happens, what.

A (53:23)

Do you think is going to happen?

B (53:24)

I don't know, I'm not in this case where speculations. I do think that, you know, it's good practice to improve. Again, it's the same thread actually. It's interesting because you will hear me giving the same speech about quantum computing like security as AI, there are foundations we need to be really good in running initiatives that will secure our foundations. In AI it can be whatever we talked about, but in quantum computing it should be about using the cryptographic tools that are suitable for this era. And I think the one thing that has changed around quantum computing is a thought of maybe it is closer than we think. And then the threat actors are stealing data now, so they will be able to decrypt it later once they figure it out.

A (54:14)

That's interesting.

B (54:15)

So you know, still now decrypt later? Yes, it's, it's an experiment, it's, you know, it's a thought. Right. We need to also give it attention. And I think there are more and more companies, especially regulated companies, government, that want to make sure that they are securing their data today against crypto, like post quantum cryptographic risks.

A (54:38)

So all right, I want to end with this. I want to talk a little bit about the threat to hacking autonomous vehicles and humanoid robots if they eventually show up. Now, I have to ask this in a way that we don't get the same. I'm confident we'll be able to get ahead of it response, because I actually want to know, like, you know, what's likely to happen if this actually goes down. I heard you speaking recently about there's been attacks on hospitals where their system, their systems have broken down and they've been unable to send dispatches out and give care to people. If this happens hypothetically, is it a magnitude greater risk if, let's say humanoid robots, which might be in people's houses or embedded in society, or autonomous cars, which we know are like all over the roads now, if there's a vulnerability, there is that like a level up from just the typical software hack.

B (55:35)

You know, there are things that happened in the history of cyber that were very surprising. For instance, what would happen if I break into your, I don't know, oven it has an ip, I'm breaking into your remote cameras. What will happen? Just I'm getting. Installing something there. So historically there was a DDoS attack that actually utilized all of these IoTs in order to target one specific, let's say target. And they brought it down because we're talking about millions and millions of devices that are flooding a certain area. But who thought that this will be the use of these iot? Right. It's a very creative use by the threat actors. Right. And I think that the way we think about the risks. Threat actors are very creative in how they utilize these sorts of stuff. And it's not always in, you know, the direct way. Okay. We're going to hack autonomous cars to run into people and robots to go into people's house and do stuff. But I think that this can cause significant outages, for instance. Right. We can use it. It can cause significant dysfunctional, let's say, societies. Right. If we rely on it too much or we don't have the proper ways to secure against it. So I do think that as things evolve, we do need to always think about how this can be used not only in the direct way, but also in part of other campaigns that can take place in cyber. And with this creativity is amazing when we innovate, it's frightening when it's used by the threat actors. We're seeing it, you know, not only in cyber, we're seeing it in everywhere. Right. And I think we should be always I would say it careful about the way we apply technologies to our day to day and the guardrails we put into play. Because the last thing we want to do is to build trust on something that breaks at some point and we don't have the guardrails to protect against it.

A (57:52)

Yeah, it's kind of crazy because throughout this conversation we've talked about how we're relying more on AI to code and now we're going to rely more on technology to work autonomously in our society, whether that's agents taking action our behalf on our behalf with computer use or whether that's cars or humanoid robots. And so I think when people like look at the trend of why cybersecurity is important, the more cyber we put in our life, so to speak, the more security we're going to need.

B (58:22)

Correct? The more estate we cover, the more security we need. And one concept, one concept that I do want to say something that was a cornerstone for, at least from my experience with customers in cloud security. It's a concept of democratizing security. Because if we think about just security teams doing security for the whole world, it's not going to scale. But if we think about our responsibility as employees, humans, developers, security teams, IT teams, and how can we contribute to this resilience of our systems, the cyber resilience of the technologies that we use, this is going to end up in, I think, a much more scalable approach to security. And one of the most interesting trends in cybersecurity is actually that democratization aspect. Another very simple example is phishing emails, right? If we're just a security team, one security training team tries to filter all emails. Something will get in and in the end will work.

A (59:24)

But yeah, we haven't even spoken about people spoofing other people's voices to try to get in.

B (59:30)

That's happening now, that's happening, that's happening. We've seen live cases and protect against that. This is another, as I said, all security products across the industry should become AI aware. And if I'm thinking about anti phishing tools, they need to be able to cope with emails that are created via AI generated capabilities that are very, very targeted and seem very, very credible. So email security tools should be able to find them phishing through fake voice messages. Like basically these are the deep fake stuff. So these are the things that all of the industry has to mature across all of them. There are solutions that are now in play. Security is not security today. And this is like me being optimist. I think security today is geared towards finding new threats, responding to them in ways that we can operationalize at scale. But we have to be aware that with new technologies come new threats.

A (60:30)

Right. There's a human side to it also, I imagine where when somebody is telling you they need money or something like that, like, you gotta, like, you can't just believe them anymore. It's weird. Even if it sounds like a parent or sibling, child, there has to be like. You have to be like, I'm just trying to be sure that this isn't an AI generated voice. Here are some things I'm gonna do to triple it.

B (60:54)

Correct. And that's our responsibility. In the end, everybody, as humans, employees, developers, in how do we make sure that we are part of this resilient system?

A (61:03)

Right. So I'm trying to think about how I'm gonna. How I feel after today's conversation. I mean, on one hand, you sort of helped broaden the spectrum of threats that I ever thought about, whether it's the ability to hack the actual tools themselves. On the other hand, it seems like the way that generative AI is working today, it hasn't led to this massive uptick in security vulnerabilities or even attacks, even though we're always seeing them escalate. So I'll choose to leave with your perspective. Optimistic. And we'll have to keep talking.

B (61:37)

Amazing. Thank you very much for having me.

A (61:39)

Thank you, Enon. Thank you, everybody, for listening. We'll be back on Friday to break down the week's news. Thanks again, and we'll see you next time on Big Technology Podcast.