Video: Agentic Enterprise 101: Understanding AI-Driven Risks | Duration: 1836s | Summary: Agentic Enterprise 101: Understanding AI-Driven Risks | Chapters: Agentic Enterprise Introduction (21.765s), AI Security Priorities (124.265s), Four Key Risk Areas (230.505s), Agent Identity Types (425.06s), OpenAI Insider Threat (612.615s), Agent Risk Characteristics (704.0649999999999s), AI Agent Risks (831.995s), Essential Agent Controls (951.8249999999999s), Securing Agentic Systems (1072.905s), Open Source Integration (1217.275s), AI-Powered SOC Operations (1437.365s), Agent Integration Solutions (1595.755s), Conclusion and Next Steps (1692.155s)
Transcript for "Agentic Enterprise 101: Understanding AI-Driven Risks":
Hey. Thanks for joining today. Today, we're gonna talk a little bit about how the agentic enterprise has evolved. We'll give a little bit of a background around what's happening in the industry today. I'm Scott Clinton, vice president of AI security and market strategy here at Exabeam. In addition to what's happening in the market, we'll talk about how, what the key challenges are for an agentic enterprise, how do we define it, what people are focused on. And then lastly, we'll talk a little bit about what Exabeam's doing to help organizations cope with this new onslaught of agentic risks and agentic security, and how we can help you get started. So let me go ahead and get started here. As we know, today, AI and Agentic is everywhere. We have been seeing a broad adoption penetration of AI and agents across most enterprise organizations. When I'm often asked, what is an agentic enterprise? The answer is most enterprises today, whether they know it or not. Organizations are seeing both vendors start to roll out agent and age... Agentic solutions. We are seeing individuals, citizen developers, pick up agents to... Using things like Copilot to create their own agents. And, also, we're seeing, an adoption, of course, of, the Frontier models and applications across the state. This is happening at an unprecedented pace, and it's no longer a single initiative. So it's not just one group that's picking it up. It's literally every part of an organization that is investing in AI and agents to help them improve processes, to be able to accelerate their work, offload activities. But all of this is coming at a cost and a set of risks. As we look at the priorities for 2026, while AI and agents have been a priority in the past, it really jumped up just in this last year to be the number two concern by CISO on a global basis. This is the highest ever position for AI in this, only steps just right behind cyber incidents. These have been the top priorities, and we've seen this reflected in what organizations have been deploying. While this has been rolling out in organizations, one thing that's been falling behind has been governance. So with this adoption running quickly, every organization is really struggling to figure out how do we just simply get visibility around our state, how can we get control, and what things can we do. What we know for sure is the adoption is growing much faster than programs are able to keep up. Part of the challenge is that when we start looking at this broad adoption, the security teams are struggling to keep up. So it's not just that the broad adoption's happening, opening us up to more risks, both machine generated risks now that, both the good guys and the bad guys have AI, where SOC chain... Teams are super challenged with their ability to respond to this. We are already seeing a lot of new exploits hopping up up in the market that are AI driven exploits, things that are being coordinated by AI systems. Not only to that, you all... You have all this deployment of AI throughout your organization, trying to figure out and keep up with the signals while still keeping up with the same signals we've had to deal with for the last twenty years for just regular AI security is just tremendous and a huge challenge for organizations. With that, when we start looking at an agentic enterprise, we really are seeing four key risk areas as organizations rapidly adopt these. These are, first of all, shadow AI. The first goal most organizations have is just getting a handle on what their estate is, what agents are being developed, what AI people are bringing from home. It's very easy to do that. Also, what type of coding agents? We've seen a broad expansion of the use of things like codecs and, and Cloud Code. These types of agents are being dropped on, desktops. People are able to to install them. Understanding what's there is one of the biggest challenges. What we also know from history is that things like deciding to just block access really doesn't work. We all kinda lived through that in the cloud era. Now we're kinda facing those same challenges. This is also the same kinda challenge we faced in every new technology adoption area. Whether it be mobile or whether it be cloud, we're now facing these same challenges. So our goal has to be on how do we make it easier for organizations to deploy AI safely and securely as part of their organizations because it's happening. It's happening with their vendors. Another key area risk is vendors who are already embedding AI and agents in their solutions. In many cases, vendors are telling you they're putting those in their solutions. In other cases, they're rolling them in to provide more capabilities. Now there's a huge value of agents, and we'll talk about how Exabeam is leveraging those to make your security operations more efficient. But it's also, being rolled out in ways that you're not really quite sure how you're gonna secure. So just as we're creating those opportunities, there's also these challenges around how are these agents being secured, how am I getting visibility, and how is that really impacting my overall estate? The fourth third one is citizen built AI. AI has become so easy for individuals to create their own agents that we have people who have never developed applications before, who are not using best practices, who are simply leveraging the resources and the tools to be able to improve their work from individual work processes. This is creating a huge proliferation of agents, whether they're being developed in Copilot or other, AI systems. Individuals are hugely proliferating this use of of AI agents. Then lastly is custom agents. So agents that organizations are intentionally developing to build and to improve process automation, to help drive new innovations in information gathering and analytics, be able to better target customers, for new products, and even, as we've seen, really using them for development. So we've seen a decreasing use of AI agents. We mentioned Cloud Code before for development. So these four risk areas are the fundamental areas that most enterprise organizations are dealing with today. That's why when we talk about having a a Jintiq enterprise, you kinda have one already, whether you want one or not. So with this whole shift to agents has been this other important paradigm. This shift from what we've had traditionally is of human insiders. Today, agents are acting as human insiders. You've probably heard, people talking about this in the industry. You may be experiencing some of these challenges yourself. So with agents and AI, we are now looking at four new kinds of identities. We've had human identities in the past. We now have human and agent assisted and ident... Identities, and I'll talk about that in a second. We have agents, who act on their own, and then we also have the rise of these digital imposters. We've seen the deep fakes. These are big challenges organizations have to make sure the insiders they're dealing with are human. Or if they're agents, how do we manage that securely? So today, you have your human or your human agents, your human insiders who we built programs for for years. We know how to build those programs. What we have now is this new classification of human and agent assisted insiders. These are humans who can... Are leveraging agents to do tasks, as we talked about before, things such as citizen developers who are using agents to enhance their tasks. Sometimes, these x... These agents or human assisted agents' identities are operating, with accidental actions. Agents go beyond their scope and do things the human didn't intend. Other times, agents are acting at the direction of humans. So one of the big challenges is how identities flow from human to agent without any without any guidance, without any guardrails. A lot of times when an agent would... Creating an agent their identity credentials, acting on behalf of them, having access to all the systems they already have access to, which creates a huge challenge. The other kind are agents that are stand alone agents. These are automation agents. These are built to help us provide improvements in business processes. These agents are acting on their own or acting in collaboration with other agents. So this represents, basically, a digital employee in many cases who is now acting in in exchange for another, human employee. But by the same token, we're struggling with how do we manage identities, how do we track these individuals as part of that. Then last is the digital imposter, deep fakes, people who are doing voice, voice, impersonations. This really changes the landscape behind how do we start to think about both the security environment for AI systems and applications as well as this new aspect around agents acting as humans or this new role of insider threat. Case in point, so we've all heard about Open AI and Hugging Face Robot. If you haven't heard about it, you know, it's certainly been out there in the news. It really was, in many ways, sort of an insider an insider, exploit. It was an agent who was acting as a privileged insider. They had access, and they had they had the ability to, basically leverage that privileged access to capitalize on governance gaps. That agent had way too many... Much permissions to be able to leverage and access to credentials that could use... They could use to escalate and, as of a result, kind of delivered a... Was able to escape from their boundaries. There were detection blind spots. There weren't behavioral baselines in place. There was no matching of what, that behave... Monitoring or what that agent was doing. As a result, it, again, broke out of its boundary. And then finally, the trust boundary failure where the agent came into, the... A partner's autonomous system, not an employee, but was able to access that third party system as an anonymous employee and impact the Hugging Face platform. So in many cases, while there are a lot of different aspects to the OpenAI Hugging Face, exploit, and there's a lot of questions and and answers about how that worked, it is very much a insider threat case study and how an AI can act as a privileged insider. So what makes these agents different than regular humans? Look, we talked about a little bit, it's trusted credentials. A lot of these agents are acting with trusted credentials. They're given access to API keys. They have OAuth tokens they can use to be able to act on behalf an individual. What makes them more challenging is they're automating. They're working autonomously. So they're not having, personal insight. They're just executing the task that they've been given. So to do that, they're accessing a variety of different APIs or taking access to resources or maybe sending emails. They're doing what they were asked to do, though not necessarily in the way they were expected to do it. So as a result, this whole challenge around this anonymous autonomous execution helps to, you know, really expand the difference between where we were potentially with humans, where we understood sort of the behaviors to now this autonomous execution where these are operating on our behalf, and in many cases, not necessarily the way we intended. The last one is around distributed activity. These are now not just in an office room. They are not just in a data center. These are activities that span a variety of different locations. They're trail... They're, they're leveraging, prompts, memory tools control, API logs. This activity for these agents to work in distributed and even coordinated fashion, is a big challenge. Well... And what's really important is that these agents themselves were, were designed to act this way. So it's not like these are security issues. In fact, these are expected behavior. And so what we have to do is start to look at how do we manage that expected behavior, understand what's not intended as part of, these applications and the app... And the usage of these agents across the enterprise. So when we look at this, we really kind of, arrive at one key concept that there is really kind of a shared risk surface that exists when you start looking at these different categories. This shared risk surface covers a number of key areas. So unauthorized tool use, whether they're having access to tools that they should or should not use, the access to data exposure to sensitive data, and then, obviously, you know, different types of privileged use, so excessive privilege or expanded privilege. In fact, the OWASP, GenAI security project and OWASP top 10 list for agentic applications lays out a clear road map, for what these top risks are and are, aligned to how AI is being exploited, everything from, goal and agent goal hijacks to rogue agents, to providing unexpected code execution, to helping to address the issue as we get more complex in our in our architecture or application architectures to represent things like cascading failures where one, system fail... One agent fails and default... It rapidly, you know, impacts others. So with all of that, really, the fundamentals here is where you do start. And with this move toward, what we used to know as deterministic applications, we knew what was the... What the inputs were and the outputs were to these new architectures, which are centered around AI agents. We now struggle to understand... Which are nondeterministic in nature. You don't always know what they're going to say consistently. They may step outside of the bounds of what they normally operate. Creates a huge challenge for organizations. What's really important here is we need to rely on behavior, monitoring behavior and the foundations that is... That has in order to be able to provide a baseline to understand what agents should and should not be able to do, because we no longer have those deterministic controls we could otherwise apply. So what that means when we start looking at our controls is we need to look at really kind of six key controls that we need to focus on. One around nonhuman identity life cycle governance. That's not just a, an agent that's also other... The nonhuman identities, obviously, focused on agents. We need to make sure we're we're applying the right level of agency or providing, making sure that we're applying the least level of agency for the agents, so make sure that we are controlling what they can and cannot access. We also need to look at how do we make sure that we have agent telemetry, pulling back into our system. This is one of the bigger gaps we have in, the architectures that are out there in the industry today as people are developing new applications using these agents, understanding what these agents are doing, how much... How many how many tokens they're using, because that has a direct impact on costs. All of those activities, we need to have visibility into them. And today, there's not a lot of consistent ways to do that. We understand... We need to understand how memory integrity works across these different agents, and, ultimately, we need to have this foundation of behavioral monitoring, which extends across not just the agent, but the full agent workflows to incorporate integration with existing systems as well so you have a foundation. This, of course, does not, eliminate the need to make sure you're executing on all of your security fundamentals. In fact, as we talked about, AI and AI agents simply help accelerate a lot of those fundamentals. If you don't have your security fundamentals in place, it's a big... You're just exacerbating the challenge. But these are the fundamental controls you have to have in place. Again, focusing on making sure that behavioral monitoring is extended to both agent to agent workflows. So we're gonna talk in just a couple of minutes around how Exabeam is helping organizations address these issues. So to begin with, the insider threat challenge, Exabeam has been focused on insider threat for some time, focusing primarily on the human aspect of it. Now with agents, we've done a lot to extend what our capabilities are there as well. We can provide visibility around both identity and credentials. We can continue to monitor behavioral change. We can start to build, you know, rapid change activity and see that through watch lists and tie that all together throughout your overall program visibility lens, from entity behavior to sequencing analytics to really being able to have a comprehensive investigation timeline based on behavior from the first, from the first action taken all the way through to the exploit when it occurred. The other step we're taking is to help you secure your agentic estate. As we talked about, there's a lot of components that need to be brought together. So first, we need to be able to verify whether or not the agent is doing what it's supposed to do. Using agent verification analytics, agent behavior verification, we're able to... And tools like Praxin, an open source tool that's been provided to help developers understand how their agents are developing or are building, are being built, and what, actions they are approved to take and when they go outside of those bounds. So number one, we need to be able to verify what agents are doing, and we're helping organizations do that, through projects like Praxin. But, also, what's important is having that observability. We talked about earlier that fact that we need to have insight in telemetry. Having that in, part of, a a a solution to secure your eugenic enterprise is central. Third, we need to make sure we have the ability to analyze behavior, make sure we understand the DNA deviations that are happening, and create baselines that allow us to respond to points at which agents or humans or humans and agents interact together and start to move outside of normal baselines. And, ultimately, it needs to start to be folded back into your program so you can improve your outcome... Your outcomes continuously, So as part of a continuous improvement program with your platform... With your programs. And what's fundamental, as I mentioned before, is we started providing tools for open source. So one of the things for Agentic AI today is that open source is found in... Foundational. We know this from the models that have been developed. What they've scraped from the Internet has been open source tools, open source libraries. So what they're building in your agents are they're leveraging open source tools and libraries. This is important because this means we have to start the security process where applications and where agents are being developed, And we have to do it and be part of doing that through open source. That's where things are being developed. Those are the libraries people are consuming. So as a security practitioner, we have to make sure that we're focused on not just, being or understanding that these are coming from an open source perspective, but we haven't been active participant in that. And for that, we've created an open source community to help us connect, the practitioners and the solutions together to be able to address these broader issues. When we look at the complete stack in bringing telemetry, we take that open source capabilities, and we bring that to that first step in helping to secure custom agents. So Observa is an open source project. You can find that in our open source community. It is a library that helps you incorporate that into your custom agents. It can then, send telemetry back using an open source schema that is... That can target ourselves, can target OTEL, or any number of other solutions that are out there. The goal is really to make sure that we have the ability to bring any custom agent telemetry back into the platform where you can start to work on it and operate on it from a security program perspective. The next part is we also have brought in to market this idea of Edge Exabeam Agent Sensor. As we talked about earlier, there are a lot of CLIs or endpoints being dropped on variety of desktops or endpoints. These include clog code, these include, codecs and other CLIs where people are starting to to do regular work. These can very easily be installed without security oversight. What our agent sensor does is it makes it easy for us to both discover and then map the telemetry back. We use the Observer's open schema for doing that so that you can bring that back into the Xbeam platform. We leverage our existing connector libraries to have access to the broad set of foundation libraries, of foundation models, and the libraries that they have so we can bring telemetry into those as well as existing applications where they're integrating with. Almost every enterprise application or systems of record today is creating an MCP server for acting with and integrating with agentic systems. This... Our collector library makes it possible for us to already bring that telemetry together and correlate that, as well. We can bring that into our agent behavior analytics, which is an extension of our UEBA capabilities, which we've been obviously using with human identity and behavioral analytics and even nonhuman identities for some time. We've extended that with behavioral analytics, so that brings it into your platform as well as helping you incorporate that into your overall program to make sure that we're driving the right side of outcomes through our outcomes navigator. So we're providing this complete stack to help you be able to bring that telemetry in and get the observability that you can act on and incorporate into your programs. The other piece is not just a state... Not securing your estate, is also how we can use AI to... In agents to help accelerate SOC operations. As I mentioned earlier, the big challenge is not only are these threats coming from just securing agents within your enterprise, the, the bad guys have AI as well. And we're seeing an increased rate of orchestrated AI events, and AI exploits that are challenging organizations. So just as we're continuing to look at an increase in areas of coverage and telemetry coming from these agents that are being developed and these AI systems that are being used, the, additionally, AI exploits are driving more alerts, more challenges, more noise into the SOC. The only way to really address that is to start to do things like leverage AI to help. And so what we've done is we've continued to build on our original set of agents that the, around Exabeam that have made it possible for us to not only be part of the, the, providing operations acceleration for each different role on the SOC. We've now started to make it possible for you to also build and customize your own agents. We're seeing many organizations starting with a core set of agents that we offer to kinda... Like we offer to help accelerate, AI investigations from start to finish. But, also, we recognize that not every SOC is built the same. You're gonna want to be able to customize your operations. So as part of that, we're enabling you to bring your own agents or to build your own SOC. So through our MCP server and our extended content that we're providing with that, that helps you be more effective in really tuning and building the SOC that you want. Also, we're providing integration into both your existing systems, and we already have integrations into different systems within the SOC. So we can make sure we can operationalize your, program across across your program. And lastly, we bring all of this to our analytics back end with our combination of both rules and behavior analytics that help you build on a... Build a foundation that you can continue to grow and secure your agentic estate while also helping your SOC team shorten investigation cycles, be more efficient, at, delivering and identifying threats and being able to defend the new state of AI solutions and applications that are coming into your enterprise every day. We've done that in two ways. As I mentioned, we've got XBeeNova, which is our built in, agents that help you to scale, for analytics and for for analysts. The other thing is to bring your own agents through our Exabeam server, which gives you full read write capabilities to be able to really incorporate, the agents in your agentic solution. We've also started to provide skill kits, so skill packs that make it even more efficient for you to be able to incorporate Exabeam in terms as part of your existing agents or to operate it directly from the CLIs of choice. We see a movement toward more headless operations in a lot of cases. So being able to access Exvene's full security context through MCP servers, through your codex or your CLI... Or sorry, codex or, cloud code CLIs are great ways to kind of incorporate it into your daily workflow in a headless fashion. We also see this extending over to groups outside of organizations, the security SOC, where organizations may incorporate or work with you. So organizations like HR, legal, and others who don't wanna go to yet another interface, the opportunity to interact through natural language to get access to, data and to provide and kick off workflows, is a real area we see of growth in the, in an agentic enterprise. So with that, you know, what's really happened? You know, AI and agents have really kind of changed the rules. And Exabeam is really here to focus on helping solve, three areas. Really, we wanna help organizations address this new insider threat class that's been developed with... We talked about before between agents, between humans, human assisted... Agent assisted humans, agents, and also, the, the the types of agents that are really kind of driving your exploits in your organization. We're helping you to protect against this growing attack surface of AI. So not only are your, these new agents that are being developed, these adoption of AI across the organization, each of these represents a new attack surface. By bringing in the telemetry, the visibility, and the access, for that and underlying that with behavioral analytics gives us the ability to, help to secure your entire agentic, your entire agentic estate. And lastly, you know, we are leveraging agents to help really tame that soft noise, really come... Make sure that we have, the opportunity to improve the overall, capacity that you have for your teams, be more efficient in investigations, be quicker to the mark to be able to do that. Again, all of this layers on top of helping you support your daily security operations that are critical to any organization. As I said before, agents and agent security doesn't eliminate the twenty years we've been spending trying to make sure we've got the right processes in place to secure our organizations. What it does is it does increase or accelerate the risks that we're dealing with, And so it exacerbates gaps that we have, and so we're there to kinda help you fill those gaps, both through the products we offer and the services we offer to make... Get you successful in helping build a secure agent take platform. And with that, you can go to our website to learn more. We just released a number of new guides on how to secure the agentic enterprise, and we look forward to working with you, kind of going forward.