00:23:40 · Yinon Costica
So AI unlocked that as well. And the third thing, we have the model advantage, which is if we are using the best models. And let's take like very sim... Let's say I have a model that performs one percent zero per... Like, percentile better than the models that are available to attackers today. Okay? This means that I can unlock millions of websites globally that the attackers still cannot. Right? So these small differences, they matter. So the ability to leverage the best model before the attackers is key. So when we think about the benefits of, let's say, a Cyber Gemini to scan for Good, scan customers, scan all of these environments and have the first... The best model to scan them, this is key. And I will say on top of this, I think that the, again, diversity of thought here matters. One model will find things that another model cannot find, and the ability to actually benefit, going back to the multi cloud, multi model concept, the benefit of scanning using multiple models, multiple perspectives will actually allow us to find more stuff than an attacker with one model. And maybe a reference to this is with Atlas. We released with Atlas. It's our vulnerability research. It ranked first on the cyber gin rankings because it doesn't use a single model. It uses multiple models, each one for the best task it can perform. And then we basically unlock immense ability to find vulnerabilities that, you know, single model solutions couldn't find. So so I think there is a pattern there when you take these three aspects, the context, the threat intel, and the model advantage. I think defenders have a good way to create an advantage in this world.
00:25:32 · Sebastian Johnson, Yinon Costica
[00:25:32] S0: And is there a way that defenders should be thinking about defending themselves from attackers versus this sort of like range of autonomous model attacks that we've seen? You know, the open AI Hugging Face incident, Anthropics had a few incidents. It, you know, that seems like it's it's almost like a very nebulous or opaque thing to defend against because it's it's just the model, right? You're not there's not an attacker behind there with an intention. You're trying to kind of guess whatever a model may or may not want to do. Like, is there a different way to think about those two types of attacks? [00:26:04] S1: So it's it's exactly the thing that we have to assume we we are being attacked using AI. And the best way to predict what they're gonna do is just to try and attack ourselves with AI. Okay? Because we cannot predict. So let the AI do it. Same same way that we are surprised again and again by what the red agent finds. It's it's exactly what we should do across our attack surface. And the and the Hugging Face OpenAI incident is really that example that persistent AI will operate in a very... Like, going back to how security operates, we always think about likelihood and... But when you have abundance of intelligence, maybe you don't need the likelihood factor. You're just looking for everything. Right? So the ability to really try AI against ourselves first to find these low risk things that we didn't think about prioritizing earlier, but AI can take advantage of, I think these are the things that we should be all doing right now.
00:26:04 · Yinon Costica, Sebastian Johnson
[00:26:04] S1: So it's it's exactly the thing that we have to assume we we are being attacked using AI. And the best way to predict what they're gonna do is just to try and attack ourselves with AI. Okay? Because we cannot predict. So let the AI do it. Same same way that we are surprised again and again by what the red agent finds. It's it's exactly what we should do across our attack surface. And the and the Hugging Face OpenAI incident is really that example that persistent AI will operate in a very... Like, going back to how security operates, we always think about likelihood and... But when you have abundance of intelligence, maybe you don't need the likelihood factor. You're just looking for everything. Right? So the ability to really try AI against ourselves first to find these low risk things that we didn't think about prioritizing earlier, but AI can take advantage of, I think these are the things that we should be all doing right now. [00:27:09] S0: Yeah. Yeah. I mean, it's it's just such... We're hearing more and more about these incidents cropping up. [00:27:16] S1: Yeah. [00:27:17] S0: Are you... Like, some of these things seem to be happening without anybody knowing. Like, how should organizations be, like, prepared for attacks that they may not even, like, know about?
00:27:29 · Yinon Costica
[00:27:29] S1: I think it's... And what we see in the reality, it raised a significant awareness for what what we call AI threat readiness. With with Google, we launched AI's threat defense. The concept of understanding that these threats will evolve and evolve very fast. And actually, it it says... So what what do we need to do? So we we actually have a a four four step program. And the first one is use the red agent. Okay? First and foremost, most important, to scan the attack surface because that's where they're gonna hit first. Second, we have to improve our ability to remediate and correct, like fix our applications at scale. Because if we just find the stuff and we don't fix them, it's not gonna help anybody. So the ability to accelerate and scale out the remediation processes, segmentation capabilities, architecturing the environment in a way that is more resilient. I think this is a second area that we have to invest. That's where the green agent operates to automate. And then the third pillar is actually about code scanning where Wiz Atlas operates. So the ability to scan the code at the get go before we even launch it and try to find all of these holes and vulnerabilities before we even deploy. And then the last piece is AI threat detection and response, which is the blue agent in ways it's the ability to continuously monitor the environment using AI, not not humans. And then AI should find the attack, but also should respond live to it. And that small point of responding live means that we need to think about defense as much more of a dynamic space than static. And by by dynamic, I mean that if if we see an attack, I want the blue agent not only to tell me that, yes, it's a malicious attack. I want it to actually create additional detections to guardrail environments. I want it maybe to create some remediative actions or response sections to cut the blast radius as much as possible. And we need to think about it as a very dynamic response to a threat that doesn't require humans. So we need to push the limits on what we allow automation to do through agents so we can actually be effective in responding to threats live. So all of these these four pillars, these are the fundamentals of how we think about security against AI threats.
00:30:13 · Sebastian Johnson, Yinon Costica
[00:30:13] S0: Is the future of cybersecurity a sort of always on system? It's like you're constantly having things that are attacking, scanning, fixing, just run permanently and autonomously. [00:30:24] S1: Absolutely. That's... It's it's like we we are almost there today in that again. It's not like futuristic. Like, this is what this what... Exactly what needs to happen because we want the red agent to compute... Continuously test, the coding agents to continuously test. We want to fix automatically. We want to respond automatically, not only to attacks, but also to red agent findings. And we want all of this to operate in a self improvement loop that will allow us to continuously say, okay. I you know, I will use a metaphor here. Like, the agent wakes up in the morning. He doesn't really wake wake up. But when it starts, it needs to ask, okay. So what are the tasks I'm going to be active on? And in the end, we're gonna find ourselves asking as a security organizations, how much capacity do I want to allocate to these agents to run on on my behalf? Right? Do I I want to have an equivalent of, you know, one continuous agent, two, 10, 20. Okay? And then it will be cost... In in the end, it will be cost versus value question. How much can I accomplish with an agent? But it will be an always on. So, yes, absolutely. I think that's the direction security is heading towards when we're not far from there. Like, I, you know, I actually think 20... Let's say, '27 early, we should be already seeing loops like that.
00:31:50 · Sebastian Johnson
[00:31:50] S0: We've got time for one more question. I wanted to ask, you know, when you look for rules the next, like, six, twelve, eighteen months, you're talking about this sort of always on coming on. What is the one thing that you're most excited about working on?