It's great to be here. Thank you, everyone. I am very excited to talk about a topic today that I think everyone is thinking about and worried about recently. And I will try to share some of the things that we are seeing by talking to customers and organizations that are working with the new Mythos model. You just heard a great presentation about how to secure AI agents. I'm going to talk about pretty much the opposite. So how do we defend against AI? AI models that are going to find the issues in our identity security.
And I want to start by explaining what I think are the core issues that those agents are now exposing. Maybe before that, just for those who don't know, Mythos is a new model by Antropic that has shown a great capability in finding security issues and exploiting them. And we've been working with customers that have access to Mythos in order to test against it. So this is just background for those who don't know, although I think everyone these days are following what's happening with Mythos. So a little bit about the background here.
I think that identity is a pretty easy target for these models because it has so many old vulnerabilities that we thought we can just forget about for many years. Identity was built over decades. It was built in silos, meaning that we had legacy identity, like Active Directory, LDAP. And then we never really fixed the security issues there, but we already added some other identity environments like cloud identities and SaaS applications. And then we added non-human identities and now agentic identities. And all of these different identities are managed in silos.
So even if one of them has good security, you'll find another that doesn't. And this is exactly what these AI models are very good at. They're good at finding the holes in your defense. For many years, this was manageable because we thought even if an attacker will come into our network and will try to find some of these holes, it will take them time. And in this time, we will detect them, we will respond to them, we will figure out what to do. But this is changing now because AI agents and AI adversaries are really exploiting these holes in our defense.
And they are making identity much more important than ever. I hope that everybody in this room knows it. Everyone in this industry today has a very urgent role. I think that many people argue about where is AI security really sitting. I think that at the heart of it, it's really an identity problem. Because the risk of AI is really defined by what it can access. But all of this is now being challenged by these new models. And specifically, MITOS is really breaking a lot of the assumptions that we had for years about security.
I'll tell you in a minute how we are actually seeing it in real environments today. But what we are seeing is that a lot of concepts that worked well for humans are now starting to really break. For example, static access rules that we built for many years defining what each user can access. I think some can argue that it never really worked even for human users. People did their access reviews and just approved everything and kept all the access. But there was at least some attempt to manage it. With these AI models, it just breaks completely. Because it's way too static.
The model is moving so fast that you can't adjust to the speed. The second is privileged access management. Which is an important solution, I have to say. But it's very narrow. It's only protecting a few of the accounts. So what we are seeing is that organizations spent years trying to scale privileged access management on board another account and another account. And it's just not enough. And now these models are exposing the fact that most of our privileged users are really outside of the PAM. There's so many of them. And the last paradigm that this is challenging is detection and response.
We assumed for years that when an attacker will come in, it will take them a few weeks, sometimes even months, to run the full attack until they really take over our network. And in this time, we will get alerts and somebody will investigate them and somebody will go and do something about it. And now there's just no time. What we are seeing in these attacks is that the model can take over even a large network in a matter of minutes. Sometimes. Up to a few hours if it really works hard. And that just doesn't leave enough time for detection and response.
So over the last weeks, we've started working with organizations that have MITOS and started testing against it in red team, blue team scenarios. We are sharing some of the learnings here because I think that the fact that MITOS is today limited to very few organizations that have access to it. I understand why it's happening because of the risk, but it's also not allowing the rest of the world to prepare. And I think that we have to prepare. So I'll show you what we are seeing. The few things that I think make these attacks very different. Everybody talks about vulnerabilities, right?
Zero day vulnerabilities in code. It finds a lot of vulnerabilities. I have to say that from the organizations that we are working with, most of the vulnerabilities are not exploitable. So some of them are. But it just finds a lot of vulnerabilities and there's a lot of manual work that is actually required to understand which of them is real, which of them is actually risking me. What I think is the bigger issue from the scenarios that we are seeing is the speed.
So even if MITOS is just using simple attack techniques, stealing credentials, moving laterally, the same as your pen testers are doing, the speed is what makes this really hard to defend against. The whole attack is done in a few minutes and you have no time to react. So if your defenses, if your prevention, if your controls that actually prevent these attacks in real time are not solid enough, it will find a hole and you have no time to adjust. It's very adaptive. So every time you add defense, it will find another way around. It's amazingly creative.
So it just fails and learns and finds another way. With human attackers, it wasn't like that. They would hit a wall and then they would have to rethink and there was some psychological impact of, okay, we failed here, are we going to try again and maybe we'll try in another week.
Here, there's no delay. And the last thing is it doesn't really make a lot of mistakes that you can detect. So in the scenarios that we have seen, none of the detection tools was even able to generate an alert because it understands what to do in order to avoid the mistakes that create anomalies. So all of this, I think, is just a step. MITOS is not the end of the AI evolution. It's just another step in the direction. I think it is an opportunity to use it as a wake up call, even internally within our companies, to talk to leadership about this is continuing to evolve and we have to prepare.
But it's going to continue to improve. As we think about how to change security, and especially identity security, in order to defend, there are a few concepts that we now believe in and think that they are really critical in order to adjust our security controls to this new reality. One is if static access rules are no longer enough, we have to shift into something that is more dynamic. And I think that there is no choice but to let AI actually be part of that. Let AI make some access decisions in real time. It's scary, I have to say, for everyone.
But it's the only way to make decisions fast enough. So people now talk about making decisions more dynamically, using all kinds of signals and context. But the truth is we, humans, are not able to handle the complexity of so many signals, so much context. Even if we had all that context, we cannot make decisions fast enough using it. We cannot build the rules that exactly define what to do in every scenario based on every signal. And I think AI will have to be part of the decision in real time.
So somebody accesses something, you will have some static rules, of course, for your critical assets, maybe regulated assets. You will have static deterministic rules. But for everything else, AI is going to make those decisions in real time using real-time context. The second thing is privileged access management. I think we have to figure out a way to scale it to every privileged identity. Apply concepts like just-in-time. Go beyond the old paradigms of putting everything in a vault and make sure that every privilege is protected.
Using multi-factor authentication, using just-in-time access, zero-standing privileges. But most importantly, get it to scale. Not stop at a hundred most critical privileged users, but really get to all the privileged accounts.
You know, the tier 1s, the tier 2s, and so on. And the last thing is, if detection is too slow, we have to move into run-time enforcement. Meaning that if we see lateral movement, if we see something bad happening, we have to block it. Block the access itself, not the next thing that it does. This is what we got used to with humans. Where we would see an attack, and we will not stop it, but we will analyze it, and then we will stop the next thing. There's not enough time. You have no way of predicting what is the next move. And you don't have enough time to figure it out.
So we have to move into in-line enforcement. Meaning when access happens, if we think this is bad access, we should block that access specifically. Not analyze it and then do something about it later. All of this led us to think about how to change the whole concept of identity security. And develop this idea, which I think many are probably thinking about now, of autonomous identity security. Because as I said, I think that we are getting to a point where humans just cannot handle the scale and the speed of these attacks.
We can also talk about that from the context of how do we protect AI agents themselves. More related to the previous talk, which is also calling for the same solution. We have no way of handling that scale and speed. And the idea is that if we combine runtime enforcement, meaning we have the technical ability, the architecture to control every access and make a real-time decision about it. And we combine that with an AI engine that can make those decisions. And we combine that with context that you need in order to make these decisions correctly, which is hard.
You get to what I would call autonomous identity security at runtime. It's a very hard thing to build, but I feel the more we look at these threats, that we pretty much have no choice. I cannot think about any other model that will actually survive this evolution of AI threats.
It's, I think, something that we have to build pretty quickly. And us, and I think others, are probably moving to solve it pretty quickly. The good news is because software development is becoming much faster now, you can actually build pretty ambitious things pretty fast. Let's talk about the context piece. How do I know what decisions to make? So somebody is trying to access something. How can I make that decision in real time? There is a lot of context that you can use. Things that we are not using today, by the way.
Things that I wish we could use even 10 years ago, but we just couldn't because there wasn't AI on our side. But now there is AI on our side. Let's use it to our advantage too, not just to the attacker's advantage. So what we can do with AI, as the defenders, is take all this context that no authorization rule was ever going to be able to handle, and feed it into LLM and let the LLM make real-time decisions using all that context. And I can tell you from testing that this is actually really good.
I don't think that any human comes close to the accuracy of these decisions, even with today's models, it's really impressive. And it's just because the amount of data that you have is amazing. I'm not talking only about identity data. I'm talking about data from the HR systems, from business context. I'm talking about the context of the prompt themselves, if we're talking about AI agents. And by the way, I'm not only talking about AI agents. I'm talking about protecting humans also, and regular service accounts and stuff like that. We can take all that context and use it to make a decision.
But without runtime enforcement, that decision is just a recommendation after the fact. That's not going to help us, right, because the attacks are too fast. So we have to combine this with runtime enforcement. So I will show you what I think is the framework for applying all of this together. I'll do it pretty quickly, given the time. So if you have users, machines, agents, accessing resources, they do it with all kinds of identity and infrastructure solutions. Some of them are using Active Directory, some of them are using Entra, Okta, Ping.
Some of them are using accounts that they have in SaaS applications locally. Some of them are just managed within some AI platforms. We need to take all of these platforms, all of these places where identities are managed. And you're going to have a few. I know a lot of vendors say you should put all of them with me. You can try. But you're probably going to have a few. Because some of your employees will create agents and manage identities in different places.
And what we need is to get all of them, whenever they're being accessed, whenever an access request is coming in, to forward it to a central brain that can make that decision. So that identity security platform that sits behind all the infrastructure, the legacy infrastructure, the cloud infrastructure, the agentic platforms, and can act as the unified decision engine. And that engine can use AI in real time. By the way, a good question is, can it be fast enough? And the answer is, there are models today that are able to handle this very complex work in under a second. And give an answer.
Which is, in my opinion, that's sufficient for most scenarios. Once the AI engine provides a decision or recommendation, it can create all kinds of actions. It can step up the authentication. It can ask a human. It can offer permissions. It can send an alert. It can do all kinds of things. And of course, this has to work with your entire security and identity stack. No one can solve this alone. This brain has to be able to talk to your MFA provider, and to your SIEM, and to your XDR, and to your HR system, and all these different tools.
But again, as I said before, the good news is, integrations are becoming much easier now with AI. Once that decision is made, and the decision is not just a yes or a no. It can also be, ask a human, because we are not sure. And learn from this. Once a decision is made and applied, you simply send back the answer to the identity provider. To the agentic platform. To the directory. And let them continue the flow. So this is a way to insert an AI-based autonomous runtime decision engine in the backend of every identity solution.
Now all of it, of course, has to come together with a full identity security platform. And again, there are many ways to achieve that. I'm only showing our approach to it. Which is really something that covers all the different types of identities. All the different types of environments or identity silos. And provides all the different security controls that are part of what I think identity security is. And by the way, I know that the term identity security today is very unclear. So I would say at least my opinion.
I think that there's a difference between identity management, which is the layer at the bottom here. And identity security. And the reason that there is a big difference is that in order to do identity security, you have to do it across all the silos of identity management. Because if a company has a few different identity management solutions, on-prem, cloud, human, agents. It doesn't help you if each of them have some local security controls. You need some layer of security on top. Just like when you have endpoint security, it has to protect all the endpoints.
And when you have cloud security, it has to protect all the clouds. So it's the same for identity security. That's how I think about identity security. We only have like half a minute. So this is the summary. Some key takeaways from this. First of all, I think that identity has a critical role in stopping these AI threats that are actually going to be a huge problem for the world in the coming months. I think that we understand it, but a lot of the world doesn't yet. And I think that it will become very clear very soon.
The problem that these AI threats bring are first and foremost the scale and the speed. The vulnerabilities that it finds, of course, are dangerous, but we will get to it. We will use the same models to fix them. That will be solved. The scale and the speed really calls for a different approach. When we see those attacks with that scale and that speed, they really challenge the current security paradigms.
And I believe that the way forward is through this new concept, this new paradigm that combines context, very rich context, the AI decisioning in real time, and the inline enforcement everywhere. All those pieces, I believe, are critical. We are out of time. This is a short slide about us. If anyone wants to talk to us, we are here. And you can see here our booth number, also my contact information. Thank you very, very much for this. I really appreciate all of you coming here.
We've got the coffee break coming up next, so people can stick around if they want to hear some answers to the questions. We've got a couple of questions here. We've heard a lot about the non-deterministic behavior of AI agents. From your point of view, which problems or challenges does this bring up when using AI as autonomous runtime decision point? Any good or bad experiences so far?
Of course, the main risk is that AI will make decisions that are unpredictable. I think that there are ways to control that. For example, you can get it to a point where whenever it's slightly unsure, it will ask a human. So human in the loop, right? But you can't do human in the loop every time, because then why do we have agents if we need to approve everything they do? So you need to gradually minimize it where the AI learns from every feedback until it becomes better. And luckily, we don't have to do it on a particular organization.
We can do it across a thousand organizations until it gets really good. But I think the real answer is that we have no choice. I don't think that we'll be able to handle the scale with human capabilities. So I think that getting AI to be able to do it well is pretty much our only choice at this point. But the next question is about scalability. It says one second response time for each API call seems super slow. So can this model really scale? Token verifications are usually asked to be tens of hundreds of millions. It's true.
There are scenarios where this will not work and you will have to create more static rules because AI will just be too slow. I think these are the minority of scenarios. But in these scenarios, yes. And I don't think we'll ever get to a point where AI makes every decision. Some decisions will be made deterministically. The ones we have to be very fast. But also the ones who are very critical for compliance or for, you know, very clear certainty that are the most critical. But all the rest of them, I think we're going to face a choice where it's either AI making decisions or nothing.
Because nobody is going to be able to build policies fast enough. We have agents created and five minutes later they're already accessing something. Nobody will be able to build policies for all of them. And I think that for that long tail, I think AI is a good answer. Great. Thanks very much. Thank you.
