Video: Securing Agentic Intent: What Your Agents Are Built For vs. What They Actually Do | Duration: 2844s | Summary: Securing Agentic Intent: What Your Agents Are Built For vs. What They Actually Do | Chapters: Welcome and Introductions (4.19s), Why Intent Matters (221.745s), Understanding Intent (406.07s), Goal vs Task Adherence (462.245s), Agent Goal Hacking (855.68s), Agent Classification Framework (1005.3s), Agent Types and Prioritization (1255.57s), Risk Tolerance Policies (1572.36s), MCP Governance Strategies (1869.17s), Balancing Security Control (2213.565s), Protection Methods Discussion (2348.8s), Least Agency Framework (2402.89s), Closing Thoughts (2572.36s), Future of Agentic Security (2642.055s), Closing Remarks (2782.015s)
Transcript for "Securing Agentic Intent: What Your Agents Are Built For vs. What They Actually Do":
Hello, everyone, and welcome. My name is Christy Hart Smith, and I am global director of analyst relations. And I am thrilled to be moderating today's conversation with our special guest from Forrester, Jeff Pollard. He's VP, Principal Analyst. Welcome, Jeff. Thank you for having me. Excited to be here. Great. And, Yuval Sarel, who is a Senior Product Manager from Cyera working on Agent Guardian. Welcome, Yuval. Hey, Christy. Thanks so much for being here and a big welcome to all of you who are on the line. I'm seeing lots of amazing drop ins from folks in the chat. I'm going to start off with our icebreaker today. Our session is going to revolve a lot around intent. And so I'm gonna start with an interesting icebreaker. And for those of you who are listening along, feel free to, drop your own answer into the chat because I would like to hear yours as well. My icebreaker question is, what is something that you do with intent? Something that's intentional. I think it's an interesting question. Jeff, you all? Yeah. I mean, I'll I'll I'll go. I mean, it's sort of a a a lame answer and maybe shows too much of, too much of how work might dominate my life, but certainly from a research perspective. Right? I love what I do, and so, you know, all the research I conduct is is is with intentionality or intent, if you will, for sure. Cool. Yeah. And that and that's... That shows, which we will see in a minute. Thank you. Yeah. My my answer is gonna be, cocktails. I I love making cocktails to... For guests I host at at home, and, I like finding out what it is they like or what it is they don't know they wanna try and and make them a cool cocktail. So I love putting thought into that one. That's a much better answer than me. I've had, like, a bunch of different answers that have flown through my head after I came up with this question approximately two and a half minutes ago. And I'm gonna say that, like, the answer that I'm going to give mostly because I'm getting on the road for both the Forrester Security and Risk Forum next week, as well as Cyera own DataSec AI conference, is that one of the things I do with intent is I blend teas. Something I do at home that's just a part of my afternoon... In the morning, I'll put, like, a pinch of this and a dash of that. And then, in a little tea strainer, and then I have the tea in the afternoon as a pick me up. And I started making these teas when I travel so that I can take the practice on the road, and it's become so, I've become so known for it that people will reach out and ask me for a particular tea blend specifically before before I leave. So I guess that's something that I do within 10. Sort of overlapping in a way with a cocktail theme. So. I guess. I have to find a way to get into. mixology of of some sort. Yeah. Exactly. Yeah. Exactly. Yeah. You're gonna need to bring a beverage too, Jeff. I guess so. I guess so. Yeah. Well, I thought that this was an interesting icebreaker because we think a lot about intent from a human perspective, behavior, behavior analytics. Right? Bringing AI and agents into the picture opens up an entirely new window around intent and what it means. So, Yuval, I think one of the first questions that you brought was why is intent important now? And, you know, I think you want to have a little bit of a conversation with Jeff around, you know, why he turned this into a whole research topic that's worth everybody's time. Yeah. Like, I think, that... That's exactly true. And, like, I think I I think... I I don't know. Many people here in the... You can comment in the chat, but, like, many people joining here today might, already be familiar with why intent is such a hot topic in security and security for AI today. But maybe some of them don't, like, don't fully get it. And I think, like, it's a good first question for Jeff. Like, you did the whole research about it. So what's your take? Why is it important now? Yeah. I think the primary reason is because this is the first time that we've ever really dealt with software that thinks. Not assigning it a personality, not saying that it's alive or has consciousness or or anything like that. Right? I'll leave that to the to to the Frontier Labs folks who may or may not believe that's the case. That's a sort of a different issue. I don't think we wanna go too far down that path, for this session. But we we are operating now with a software that actually thinks. Right? We call them agents because they have agency. They have the ability to make decisions. And when you have the ability to make a decision, to reason, and then to execute actions as a result of that, it becomes of, I think, frankly, paramount importance that you understand the intent of why those actions were taken. So what were the inputs? What were the factors? What was considered? And, ultimately, what were the choices that were made that elected for a given set of actions to happen once the software was given a goal? And that goal could from... Could come from or in the form of a prompt from a user, which is the way that we often think about AI working. But when you start to explore the more sophisticated use cases of AI, the more sophisticated deployments of AI, you'll find that sort of that relationship of one to one prompting of person to prompt, it stops existing. There might be a system prompt. There might be a prompt that we've programmed into an agent. But in true multi agent systems, especially when you get into loops and graphs and routines and things like that, it's not a user and a prompt anymore. It... It's instead a set of inputs, and those inputs could be processes. They could be data. They could be tools. Those can also be outputs as well. And so we have to understand why the software has decided to do what it did, what else it considered, and how much that deviated from what we possibly wanted in order to get a sort of valid understanding to secure this for our environment. So I think, ultimately, that's why I think intent is so important now. When software was a series of calls and we. pretty much knew what it would do, but even then, we didn't always know what it would do, but we at least had a pretty solid idea that was an easier world to live in now that we have this other aspect where it's not just, you know, a system call or a function or an API or a library or a DLL that tells it what to do. We have to understand everything that went into it to know exactly why what happened happened. Yeah. I, like, I I agree with all all the statements you made there. And, like, I think I think one way that that kinda maybe helps make sense of it, to me to me at least, is that, you know, we used to try and and gather intent by UEBA or anomaly detection or, like, mechanisms like that that try to understand patterns and kinda deduce that, okay, risky... Oh, like, sorry. Anomalous is, malicious. Right? Or, like, Yep. that's, like, what's the assumption. And now we're in a much more sophisticated world where, a, we can understand intent because we have a thinking software, which is agents. And we must, understand intent because everything you just said... Because that's, like, the level we'll operate... We let... We're letting the users operate in today. Yep. Yeah. You. know, you were, Jeff, you were talking about, in how how agents go about doing the business, getting instructions from a human user, but then also beginning to delegate work out to others. And you used a phrase with me, a few months ago, about the difference between task adherence and goal adherence. And I... That's the clearest I've heard anyone explain that. I know we're gonna get into, like, your formal definition of intent and that kind of stuff. But as you're as you're thinking about... You know, and Yuval, you also just brought up, like, the behavior analysis side of things. Right? And I did some time as a business analyst and I would write very elaborate business requirements for systems. Right? And and I think of those these days when I'm writing a prompt for an agent to follow, but then the agent goes, you know, then the agent goes off and does its other thing. So, like, there's, like, different layers of where intent is coming from. Is it jumping the gun to ask for a definition of these two areas, or do you wanna start with more of a broader definition of just... Of what intent means to you? No. I think that's a that's a great place to start. So when when we started thinking about how to define intent, we did the research, we talked to a lot of different organizations, a lot of different practitioners. Yeah. It really kind of came down to looking at the goal adherence, the task adherence, the set of constraints that were imposed upon the technology, and then the actions that the technology takes. So the calls, the systems, the data that it interacts with, and that really kinda forms the foundation of the definition of intent and then obviously securing that intent and applying some constraints and guardrails to it is really important. And when you think about goal and task adherence, the the way that I I sort of often simplify that is, as a user, there's something that I want to accomplish. Now again, that might come in the form of a prompt, It may not. But I want to accomplish some sort of outcome, and maybe I type that directly, and it could be something like summarize my email. Or, you know, it could be build a brand new application. And, of course, because we wanna be thorough and it's gonna... And we want it to be perfect, we just add make no mistakes, and we know the AI will do it without a problem. When you think about that, though, like summarize email, for example, that's that's a really complex request where we're asking it to interact with an inbox, which is another piece of technology that we probably integrated via MCP and APIs. We're asking it to then look at emails to classify, to summarize, to provide the salient points, to make determinations about that. We're effectively allowing the software to decide what should what should my priorities be. And so as it begins to do that, what are the individual tasks that are performed within that chain of actions? And if those tasks suddenly start to deviate from what makes sense in terms of summarizing my e mail, well, that becomes a pretty important thing for us to know. When the agent starts to behave in a way that we don't like, we need to understand that and those tasks can become clues whether or not they are serving the purpose of what our ultimate goal was, which could be summarized inbox. Well, if that's the request I have and suddenly the agent starts to authenticate to websites that have nothing to do with anything that was in my email, suddenly the tasks are no longer aligned with the goal. We need to be able to classify that so that we can understand why it happened, what may have led it astray, is it a security issue, is it a configuration issue, is it an ambiguous request, did the user just not specify it correctly. The other part that becomes really, really important, and this is the one that I think a lot of security teams often don't think about and it goes right to, Yuval, what you said, is so interesting. Right? In security, a lot of times, we either knew something was bad because we knew about it, so we had definitive proof whenever this happens, it's bad, or we saw an anomaly and we sort of assumed bad because we had never seen it before. And if it was rare, it was probably bad. And so we had kind of these boxes that we were able to put things in. When you start dealing with agents, though, it's it's not that neat. It gets way more messy. The example that I use, and this is a real example I've heard in in several scenarios, where you have something where maybe it interacts with a CRM system, and a user is asking an agent for account info. And it uses a correct API to get that account info. It's granular. It's restricted. It only shares the information that user should be able to access. That's awesome. When you actually dig into it though, what you discover or what you could discover is that it only used that API because that was the easiest route to go to get the information. But if that hadn't worked, it could have invoked the deprecated API. The deprecated API would literally give that user all of the information, all of the information about the account, maybe account numbers, maybe billing numbers, maybe credit card numbers, things that that that person doesn't need access to, and suddenly discover that that action the agent took when it provided exactly what the user asked for, that was only safe or okay from a security perspective because it was the easiest option. If that option had failed, then you actually would have had a security issue. Not saying an incident, not saying a breach, but you would have probably had an exposure. And that is something, again, that we we really haven't dealt with. Right? The the the scenario where a user goes and tries to use one API versus another, there's no average user that's gonna do that in any environment. That is... That, you know, that is not something that the average corporate user is gonna do. A dev might do that, but certainly not rank and file employee. But with agents, the agent can decide to do that. In fact, agents are persistent. So when one thing doesn't work, they try other things. They adapt because they want to accomplish the goal that they were given. And so there is an example where we could really be confident about our security practices only to eventually find out that the reason why we were successful was not because we had done anything necessarily to prevent that from happening. It was because, that was just the faster option it had seen more often. But it could have done the other thing. We need to know that so that way we can secure it correctly. So I know kind of a long example, a longer answer, but that to me is why that goal and task adherence is so important because we need to understand the deviations. Yeah. It's so important. And you've all... You... The... There's a customer example that I know you, we talked about in the... During the prep for this, which was a personal email inbox as well. So in your answer, like, we'd love to hear an example of that that we've caught. Yeah. There... There's, like, so many. Right? Like, both from, like, stuff that we caught from agent guardian and, like, from public incidents as well. The the first thing that come... Jumps to mind is actually not an... A customer insight is, like, a a incident. Sorry. It's like a it's like like... They found out that, sorry, that, like, OpenAI models, if they had a personality, they're not proactive enough. So they had to bake into the system prompt of codex. Like, example of when the user says, can you check? Can you do? Can you? They should do it and not just answer the question, which goes to show, like, how hard it is to, like, assess actual intent when, like, building a program. And and, yeah, like, we... We've we've caught incidents where, like, the agent extracted credentials from, like, the... It was told to con... To contact and access Jira, to organize tickets in Jira, and it had an MCP, and the MCP was unauthenticated. Like, it it couldn't actually do the job. So it kinda did, goal hacking, which goes to the examples you gave, Jeff, and it found a way to get Jira. It bypassed it. It found an API credential in another process's memory, extracted it, and use that to access Jira. So it didn't do... Like, it didn't access any data. It shouldn't have. It just used an unsanctioned method, and it found a way to do it. And usually, the user doesn't even know it... It's happening. Like, to the people in the chat listening right now, it's probably happening happening to you on a daily basis. So, like, how do you actually stop something so powerful, so relentless, that's always finding the the way to get to to get to its goal, even if it bypasses your organization's intent of how it should be used. Maybe that's even, like, a great segue to to kinda ask you, Jeff. Like, okay. So I think... I I hope, like, people listening right now, like, have a good grasp of what intent is and why are we talking about it. But how do you how do you formalize this? Like, when you're building a security program, when you're evaluating your risk in the organization, intent can manifest in many different forms. So you have, like, a very good, structural definition. So can you walk us through that? Yeah. I think the first thing that you have to do with this is start to classify what your agents are doing and how they do it. And we think about this in sort of two by two terms where we've got sort of emergent helpfulness, engineered helpfulness, and then accidental harm and purposeful harm. And the reason why it's important to have those top two, the engineered helpfulness and the emergent helpfulness, as a classification is because on the engineered helpfulness side, that's where the agent does what you wanted it to do. Work with Jira. It did it in a way that you didn't want, which might be accidental harm, but it did accomplish the objective. So you wanna understand that. On the emergent helpfulness side, what's really important is that as agents are creative or figure out new ways to do things, what we have to do as security practitioners is balance things. If we overconstrain with guardrails and policies and technology that limits what AI can do, we run the risk of eliminating some of the magic. And and here's the secret that we all don't really talk about, but it's true. When AI does something that I wasn't expecting that I like, I take credit for it, and I get praise for it. Right? That's that emergent helpfulness. When AI does something I don't like that I wasn't expecting, well, I blame the AI. That's when that's when all of a sudden it's the agent's fault or the AI hallucinated or it went rogue or whatever we wanna call it. So we've gotta be able to sort of make sure that we nurture and understand each of those scenarios. And then what we have to do is start to layer the technology controls on top of it such that we can facilitate the engineered helpfulness of what the agent's supposed to do. We can allow for aspects of emergent helpfulness that we want when appropriate, but we can also understand when something is accidentally harmful or purposely harmful when it's using an unsanctioned method to accomplish its goals, but it did that in a task. And. so then that really, like a lot of things within security, comes to applying technical controls to agents such that we can understand the agent's chain of thought. We can understand the identity of the agent and of the user and what it's operating on behalf of. We can understand the initial prompt that might have caused this behavior, that initial goal, and also mechanisms in place to actually have guardrails and to stop the agent from doing what it's doing in the event that it does go a bit sideways so we can offer the organization some degree of certainty that we can enable the use of AI because we have mechanisms in place to stop and enforce. But I think it really starts with understanding how to classify these things because that is when you start to figure out, okay. Once we've classified it, how do we think about the risk levels? Do we want to curb accidental harm? Absolutely. Do we want to stop purposeful harm? Of course. Do we wanna nurture and encourage emergent behavior within certain constraints? And do we wanna make sure engineered helpfulness is actually happening? 100%. And so the way that we use that is through something we created at Forrester called the Aegis framework, which is an agentic, specific agentic security kind of, for AI agents. And then within that, we use things like an agent on a page template that helps you understand what is the agent designed to do, who works with it, what are its goals, what are the tools and technology that it can interact with, what's the data that it can access, and then what are the outcomes it's supposed to produce, and what are the goals for those outcomes. That's that's how we think about it. But it's both a policy question, like a lot of these things are, but it's also. a process question, and technology and observability and enforcement is required for these things. We've gotta have all of those layers, and that's really complex because this ecosystem of technologies is changing every hour. You. know, So there's a question that's a... Yeah. From the chat, like, maybe we're going for the same thing, Christy. So I see some questions in the chat about deploy time, like intent around deploy time of agent. Like, what was the agent designed to do? And I think it it goes to, like, a... Maybe even a bigger question than that, and I I think it's gonna it's gonna also help answer the question in the chat. So AI agents, I think from the, like, infrastructure and and technology tooling level, kind of moved in the past six months maybe from Yep. more of, like, let's call it application agents, like agents that are serving a unique purpose in the organization. It can be broad, it can be narrow to... Let's call them personal agents. So, general purpose general purpose agents or personal agents like code that can do a lot of stuff under some policy on... In the organization. And I think intent, all of the types of intent that you mentioned, applicable to all... To both of these types, rough types. But, like, how do you... Where does that meet you, and, like, what do you think about it? Like, does that change anything in your thesis, like, this transition in usage, or is it, like, exactly the same on both categories? No. I... So I think it does change things, at least in terms of prioritization because we've gotta be pragmatic about this. The example that I that I often use here is let's, let's talk about Copilot, given sort of its presence from an enterprise deployment perspective. So if you take something like Copilot m three sixty five, would I say that you need to classify or even necessarily investigate a Copilot m three sixty five agent? No. It's not realistic. One, because frankly, a Copilot m three sixty five agent is a stored prompt. It's it's not really an agent. It's not orchestrated. It can't reach outside of the environment to send an email, to post to an API, to interact with data. Right? It's it's much more limited, and you should have other controls in place for that. Could it generate an e mail and could it possibly include things like account numbers or whatever? A 100%. You should have other controls that would stop that person from copying and pasting that into an e mail and sending it. Wink. You know, there are controls for that. There are things you can do. When it comes to something like Copilot Studio though, if this is gonna be an enterprise broadly deployed, curated, supported agent that works across multiple data sources like a CRM, like a productivity suite, like e mail, does it interact or integrate with other APIs? Then, yeah, suddenly that's something that's performing an enterprise process on behalf of users. It's expansive and you can do a lot of things with that agent. Yeah, you absolutely probably want to look deeper at an agent like that. Then you can go to something like Azure AI Foundry where you can do anything you want with that. It's highly technical to build it. You have to know what you're doing. You have to know some AI. You have to know some development. But the agents that you build with that are immensely more powerful. And so you need to think about it that way. So when you start to think about intent, I often say start with, like, one to three agents. Pick the ones that are centralized, that are curated, that are actively being used to accomplish a business purpose, something like a customer support agent, something like an accounts payable, accounts receivable agent, maybe a sales agent, something like that because it's customized enough, it's widespread enough, and it's central enough to a business process that it makes sense for you to understand that as opposed to trying to do it for every developer with Cloud Code, for example. Because sure, Cloud Code absolutely can spawn multi agent and it does and you can tell it to do that. But at the same time, that is the harness that controls those decisions. So instead of necessarily inspecting the intent of everything that a clogged code might do, what you need to inspect is what you produce with clogged code because you still need visibility into what that harness and what that orchestrated system might produce. But that's really a build problem. It's building something for you, and so you wanna look at the end result of what's built, not. necessarily what's going on through. If that code's gonna make it to production, it needs to run through your normal AppSec processes. If you're building an agent with it, of course, but it's a different sort of scenario. So I think that's part of the confusion because, really, when you start talking agents, we're talking about agents that are bundled into design software, agents that are bundled into productivity software, personal agents that I can download and install like Open Claw, Hermes, or something like that, autopilot from Microsoft, not the other autopilot from Microsoft, but the new autopilot, you know. Okay. Let's not go. too far down that road. But... And then you've got the tools that you can actually build things with and those can create other sorts of agents. I think that's where a lot of the complexity comes from. But for me, it's about the agents that you are building, curating, deploying to the rest of the org. Those are the ones that you wanna think about the intent for. You wanna track. You wanna classify and apply the controls too. That's what that's what matters. I wanna get a little bit, like, narrower on this mostly because I'm having... Like, Yuval has asked some amazing questions and follow ups in the thread on his. And he's he's asking about, like, on on goal adherence and task adherence, like, when the two disagree, say that the agent finds a faster path that the builder never planned, which to me is like that emergent helpfulness, could be, emergent helpfulness or harmfulness depending on whether that's like. a log into a production server that it's created. Like, Yep. the question that he's asking is which one does the control trust? Now, personally, the way I would handle that is to have different sets of policies. One that is looking at the outcome with the original goal of what the builder provided and another one that is looking at what is the risk tolerance for how agents operating on this type of data or agents with these types of permissions are allowed to pursue their tasks. So I I personally might look at it with two different policy sets, but that's like a level of complexity, especially if we're used to just thinking about, like, one policy that's going to solve one specific issue. We're trying to handle a variety of different kinds of issues at at at once, really. Yeah. It's a it's a really cool scenario, because it it it is very organizationally and task specific. The example I would give is that when I talk to an organization in the financial services or insurance sectors, if I am talking to them about something to do with underwriting or lending or something like that, their tolerance for how creative or how locked down an agent needs to be is drastically different than if I am talking to more of a software company or a technology company that is attempting to build out an agentic workflow for agentic software development. In that scenario, they might want those agents to learn and create more efficient patterns, more efficient workflows that weren't originally thought of. If I've got a customer support agent, I might want it to learn a more efficient way to retrieve data as long as that it doesn't cause me problems with exposure, for example. So if I'm a consumer product goods company, my tolerance for that variance is probably higher than. if I'm a credit card company, right, as an example. So I think that's a... It's a it's a really important question. First of all, I think for most organizations, they wouldn't be able to answer it in the first place because they haven't sat down and classified these things. So if you don't know goal adherence and task adherence, if you don't have the sort of runtime controls and visibility, you can't answer that. That's why we created that two by two. That's why we released the securing intent research to equip organizations with the questions, with the classification. And then from there, it is about guiding the organization through the appropriate level of risk. And I'll be entirely transparent here. It is staggering when you are in a meeting with a business leader and you start to talk about agentic accountability. Right? From from a security perspective, we've often said that we are there to help the organization understand how and what risk they are willing to accept, and we will give them answers that tailor the amount of risk acceptance they'll tolerate based on what they tell us. Right? We're coaching. We're not deciding. Agentic is wild and different because you can take a business leader that you've worked with that has a human workflow, that has mistakes, that has problems, that has errors, that has issues, that have major consequences in terms of litigation and regulatory fines, and they are accepting of that. That is their job. They understand as a leader. They have to accept it. And then you tell them we're gonna reduce the size of your team by 75%, and we're gonna add 200 agents to it. And you can watch them visibly recoil and say, woah. Woah. Woah. Woah. Woah. I'm not accepting the risk for those agents. That's that's not it. Because you've now taken a process problem, and you've made them accountable for a technology problem. And so they expect technology to be responsible for that technology problem. And so this is a really different world from a security. practitioner perspective to start having those conversations. And so the other part of this is to be able to have a framework where we can put it in terms that they understand about things like goal adherence. Our goal is to execute this process with a minimal number of errors. Here are the tasks that are acceptable. Here are the tasks that are not, and then let's apply guardrails and controls. Maybe it's at the data layer. Maybe it's from an MCP perspective. Maybe it's identity in the tools that are authorized or can be invoked. Maybe it's about the actual agent itself and how much it can do before you need to get to a human handoff or not. Some organizations might want the workflow to be nearly autonomous, and others want far more checkpoints or check ins with a person. Those are all valid, but we've gotta be able to answer some of those questions. And, you know, frankly, before our research, there really wasn't a way to do that. Yeah. Yeah. Well, What. That's like... Oh, yeah. So I I loved, like, the way... Like, all these conversations are super interesting, and I especially, like, I wanna double tap on what you said about, like, MCPs and tools and identities and, like, finding the ways that you can actually govern all of this. Like, how can you bring it into practice? And something interesting that I found and I'd like... I'd love your perspective on is that I find a lot of customers I speak with really want to lean into the MCP governance and skill governance and, like, the tooling of agents, aspect of things. And I find that interesting because it makes total sense. It's like the highway of agents. Right? This is where most traffic and most permissions are used. But there's... It's not only highway. Right? The agent can go through the small roads of data traffic. And through the wilderness and, like, a hike through the mountains, like, it will find a way to get what it needs to be done. And, like, if you focus only on what feels comfortable because you can govern it, then you kinda miss out on all the soft stuff that if you close all, like, all the borders around the main the main line, you're gonna have all these, holes, all these, like, potential mishaps or breaches. So I know we do both. Like, we we tackle both angles. Like, we're working around MCP governance, like, finding the malicious MCPs and making sure they're not in use and, like, making sure the correct MCPs are in use. And also monitoring for, like like you said, like, intent drift. Right? Like, where is the agent being, helpful, because it's not adhering to the task they have stated or to the company's policy. So, like, in your mind, like, when you're talking with business and and technology leaders, where do you see they're spending their focus on? Like, what do you think about this, like, tension between these two... Between these two, fronts? Yeah. It... It's interesting because so many of these conversations, one, it sort of depends on where the person comes from inside the organization. If we're talking to someone that might be more on the network side, then often the idea is, okay, is this a gateway solution? In other words, is this a proxy or a gateway of some sort? If we're talking to someone on the development side, then often it's a conversation about API gateway because that's where they've come from. And so they think, oh, well, let's just add MCP to an API gateway or let's add runtime to an API gateway, and that's what they understand. If they've been more on the endpoint side or in some cases a lot on the security side, they might have an EDR answer, or an endpoint protection answer to this. And the truth is we're gonna need all of those. None of those is the right answer on its own. None of those is the wrong answer on its own. We're going to need all of them. MCP is an incredibly alluring place to start to think about implementing controls. One, because MCP makes things very easy. Also, the old joke, which is still kind of true, the S in MCP stands for security. So if if that is going to be the mechanism with which we use for integrations with agentic systems, then, obviously, we need to have that as a part of our control plane that completely and totally makes sense. The the problem that I have is that there's a lot of sort of silver bulletishness around MCP, if you will, because it's this idea of, oh, well, if we're there, we've solved a lot of the problem. Well, here's the issue with that. Yuval, the example you gave is great. Extracting an API key from memory and then using that in order to interact with the same system but in a different way. Keep in mind, all I have to do is have access to API docs and then I can take that API key. I can write my own script if I have that ability as an agent and then directly interact with that system via API and not use MCP at all. I don't have to have that pathway. So MCP is a really, really important part of this. It absolutely needs to be part of our our plan and how we're gonna secure these things. It's just not the only answer, and we can't get too hung up on that. And we've also gotta be careful because there are a lot of scenarios where some teams are resistant to force everything through a proxy and introduce a single point of failure into an architecture that's been designed to be robust and resilient outside of that. So there are some legitimate sort of development and architectural concerns with trying to have this kind of single point of visibility, which also in a lot of scenarios might introduce a single point of failure. So, I do think there's a lot of where folks come from. Naturally, politics and organizational dynamics play into it. It might be easier to do something like MCP than add another endpoint agent to every single endpoint. That that makes a lot of sense. Maybe API integrations can make sense in some cases with SaaS platforms as more and more... Especially even Frontier Labs go to, you know, cloud compute and VMs in the cloud and things like that. So I think there's a lot of answers, but the real sort of ground truth to this is that this is gonna be about assembling a bunch of different solutions to give you everything that you need. And that is painful but necessary. And it's painful like any new technology is because this is a new technology, and so we're not gonna have a one. a one size fits all answer right now. Yeah. Like, from a... I, like, I love it. I I... I'm only feeling bad because I I see the time running out, but I have so much to say. I'll just say that, like, I find it, like, from a builder's perspective, right, building a product in this space, I find it, really, like, I need to balance between give more rope, so to speak, to customers whether they want to govern and where they want to apply, like, the controls and against, like, having an opinionated stance and, like, giving them the veggies along with the dessert. Right? Like, this is what. will eventually get you safe. This is what we think from, like, our understanding of the world and, like, having both of these, like, tracks within the product development life cycle internally, etcetera. So eventually, we can, like, keep up with things and and, like, have them be secured. Right? That's the goal, but super. interesting. And and I think this is this is something every security leader is dealing with right now, and it is it is a challenge because these technologies are are so different. We used to kind of have these neat categories of software. And once you add agents and Agenca capability to it, we do start to sort of reduce those categories. They don't exist in the same way anymore. Right? It's not just a CRM system. It's a CRM system that can now reason and take action and write a script if it's not controlled. Like, all of those things are true, and we've never dealt with that kind of capability being handed to a nontechnical user and being very, very usable for that nontechnical user. Right? The ability to build now to your point, it's drastically different. I was really bad at Python. I'm no better at it. But if you look at the stuff I make personally, it looks like I'm better at it because I can actually make things that work now. And before I kinda couldn't. But but, you know, we've we've really given a powerful technology to people, and it's part of our job to make sure, as security practitioners, that they can take advantage of that. I mean, that that really does need to be our goal, and our intent is to equip the organization to do what it wants to do with these technologies. Because if we don't, they're gonna go around us. We know that. Shadow IT taught us that, and Shadow AI is certainly reminding us of that. Well, I know, we are getting close on time. I've I've, been interjecting with the questions as we've gone along, but a reminder to everyone in the audience, if you do have questions, please drop them in the chat or in the q and a tab, and we'll get them answered here as we're wrapping up. But I'd like to circle back to the Jeff, point, Jeff, because you brought that up and you built this framework around different types of violations that you expect to see. And also you are giving some guidance as a part of that in terms of some of the best methods of protection. So we've talked about MCPs here just recently. We've talked about a couple of different ways to think about both the goal adherence and the task adherence. Are there other kinds of violations or other kinds of, protection methods? And we talked about data classification. Right? But, like, others that this audience could walk away with from this call. Yeah. I think some of the most important, certainly data security. And there's an interesting dynamic from a data security perspective because it might be one of the first times that understanding data, labeling data, classifying data actually has a legitimate strategic enterprise benefit. Because if you've done those things maturely and well and you have that sort of strong data governance, you're now benefiting the AI efforts of the broader enterprise. So I often joke with my colleague Heidi Shea who covers data security that, like, like, her thing went from we're all mad at DLP to this is one of the most strategic things the organization needs to have. And it and it happened. like that. Right? It it... You know? So I think that's that's one part of it. Agentic identity is a whole other problem. You know, I often say... I've covered emerging technologies for years now and securing emerging technologies. And one of the lessons that I consistently learn when it comes to securing emerging technologies is that they exacerbate problems we already had. Like, yes, they introduce some new problems. That's absolutely true. But they exacerbate the problems we already had. When IoT was an emerging technology, if you had network visibility problems, well, now you have more devices on your network. If if you had asset management problems, well, now you have more assets. IAM was certainly not a solved problem. Now we have agents. Those agents have different types of agents. They can spawn their own agents. They're temporal. They're ephemeral. And so if you had identity problems, you now have more identity problems. That's a big one. And then finally, I think risk management. We are used to this world of risk management where our environments were relatively steady. We did a once a year sort of checklist. We kind of figured out our risk. We did a business impact assessment. That doesn't work anymore. You know, in this. world where I can spin up an agent, you can spin up an agent, we can write software on demand to solve a problem, that process of once a year checklist kind of risk management doesn't work. It's it's about continuous risk management. And then finally, I think the last thing that this has really done, Forrester, of course, we originated zero trust, put it out there. Obviously, it sort of took took over, the market. I think the next phase of that is really about least agency. So zero trust was really about least privilege. When it comes to agents, it's about least agency. We have to be able to give an agent the minimum amount of authority, access, permission to do what it needs to do for the shortest amount of time possible, and that's really the concept of least agency. So least privilege was about what can it what what can it get to. Least agency is about what can it get to, but also what can it do with what permissions on whose behalf. And so that's really kind of the next expansion. So I think outside of intent, the other two real sort of pillars of of Aegis that we think about is least agency and continuous risk management. That's great. I was gonna... I was... I... I'm... Of course, I'm always interested in where you're going next. And, Jeff, you're one of the voices that I look to when I'm looking at emerging technologies. So thanks for throwing in some of the things that you look at as you are covering emerging technologies to see, like, how they do exacerbate things... Problems we've been, faced with for some time. I... We've opened up the survey. So for those of you online, if you wouldn't mind responding to the survey and letting us know what you enjoyed about this, we've got just about a minute left. Yuval, is there a parting question? Is there a parting thought you'd like to end on here? Oh, wow. That... That's a tough question. Yeah. Like, I I think, like, maybe because I have so many more questions, maybe, Jeff, I'll I'll ask you. Like, what what what should should people focus on when speaking about intent? Like, if if you had to to take one takeaway, one thing to do differently when you're assessing your next moves in securing AI, what's, like, the the the one... If you had to take one thing, like, what would it be around intent? I think it's it's gaining the runtime observability of the agent. That is that is super important. You need to understand the chain of thought. You need to understand the traces. Again, you need to understand what the agent did, but you also need to understand what the agent could have done and maybe decided not to do for some reason because that can reveal a lot of gaps and issues that you weren't aware of in your environment. So I think those are are the things that are so important because once you have those, you can start to classify, understand. And from there, you can develop your policies, your guardrails, your constraints that you wanna apply to those. So it's really about getting that data of what was the agent doing and why, and then starting to apply those controls to it because then you actually understand why what happened happened. So you're saying just like, it... It's an amazing answer, but you're saying that, like, the runtime observability can can lead to the policy itself. Right? You... If you don't know what your policy is, you can start from the visibility and, like, build from there. Yeah. Absolutely. I think that's a... That... That's that's an absolute necessity, again, because you you need to understand how this agent is reasoning, what it has access to, what it decided to do, what you don't have documented. And then from there, you can kind of work backwards to patch those gaps, right, to change those policies, to change those configurations, to adjust permissions, that the agent has. And so that's why I think that observability is so important. And in some cases, we don't have it. In some cases, you can't trust it. This is an interesting world that gets very metaphysical almost because we know what things like... We we know what things like chain of thought, for example, that what the agent says it's doing is not always what it's doing. Sometimes it tells us. it's doing something, but when you actually look, it's not. And so that ability to find those gaps of where it said this, but it was actually doing that, that's another important thing that we need to understand. So it is a... It's a it's a tough problem, but I think that understanding, classifying, securing intent is, from an agentic perspective, is literally the foundation of securing agentic AI. Like, if we don't do this, we can't we can't accomplish this objective, which means our organizations can't trust what they're doing. So I think this is a huge opportunity for us, and I think it's the most important thing that we'll be working on over the next two to three years. Amazing. I wish I could get chain of thought from my husband because I bet you sometimes what he says is very different than what's going on inside his head. And that's probably a good thing. That might be that might be what preserves the relationship, Christy. So so maybe maybe you don't want that. Oh my gosh. Guys, this is amazing. Thank you so much. For everyone who's joined, we're wrapping up here. I would like to point you to the documents portion of the interface because you will see the option to download Forrester's vision report on securing intent. If you enjoyed today's conversation, I highly recommend you check out some of the words for visuals. We decided to keep this as a conversation rather than just PowerPointing you to death. There's also, the ability to request a demo and see Cyera Agent Guardian at work. So thanks so much, Jeff and Yuval, and for all of you for taking time out of your, Wednesday for being a part of us.