Video: How to Become Agent Ready with Cyera | Duration: 3520s | Summary: How to Become Agent Ready with Cyera | Chapters: Welcome and Introduction (15.635s), Introduction & Context (88.25s), Agent Guardrails Needed (211.995s), Data-Centric Security Approach (414.67s), Sayera Agent Guardian (598.57s), Agent Security Challenges (712.295s), Data Classification Foundation (901.465s), Agent Guardian Introduction (1125.545s), Discovery and Governance (1427.08s), AI Asset Inventory (1634.965s), Protection and Validation (2107.105s), Live Demo Walkthrough (2316.595s), Platform Differentiation (2810.47s), Agent Drift Detection (3100.41s), Agent Intent Verification (3150.705s), Governance and Integration (3229.215s), Q&A Wrap-Up (3305.92s), Closing and Next Steps (3424.055s)
Transcript for "How to Become Agent Ready with Cyera":
Hi, everyone. Just, waiting for some people to trickle in, and we'll get started here in about a minute. For those that are just joining, we're just waiting, about just one more minute for some people to trickle in, and we'll get started here shortly. Alright, Yotsi. I think but let's let's reward the punctual people, and let's get started, whenever you're ready. Yeah. Let's get the party started. Well, today, we're gonna talk a little bit about agents and agent security, how to make sure you're rolling out in a safe, and, in a safe manner. My name is Yitsi. I am part of the product marketing team at Zayra, and I'm joined by my partner in crime, Zen Malik, who also is part of the product marketing team at Sayera. And before we jump into things, I wanna share maybe kinda a quick story. In the late night at nineteenth century, fast growing cities had relied very heavily on horses for transit and for freight. So, like, for example, in New York City alone, my hometown, also Zenon's hometown, actually, Yes, sir. they had 50 a 150,000 working horses. Now it's weird to start our webinar on this topic, but each horse produced between fifteen and thirty pounds of manure a day and about a quart of urine. Now I'm not good at math, but a 150,000 times that amount of of manure is a lot, is a lot out there. And actually, in '18 in 1894, the lung London's The Time predicted that every street in London would be buried under nine feet of manure by 1950. To actually tackle this emergency, global delegates met in New York City in 1898 for the world's first international urban planning conference where horse waste was the main topic. Now I assume many of you guys can guess how this problem was solved. It was solved by the Model t, by Ford, putting out the car. Horses were no longer needed to get around, and the car happened. Now when it happened, there were no rules. There were no guardrails. Right? You didn't have to get a driver's license. There were no one way streets. There were no stop signs or, traffic lights or speed limits. Right? None of that exists. You could have bought a car in nineteen o eight and driven wherever your heart pleased. But as time went along, you know, cars evolved. There there was internal, safety put in such as seat belts and airbags. Roads started getting, kind of paintings, and there were lanes, and you had to stick to your lane, and there were speed limit right? Speed limits started happening, stop signs, caution. You had to get a license. There was a written test. There was a practical test. You had to renew your license and pay the government money. All that fun stuff happened until what we, know and love today is, you know, cars are have a lot of safety mechanisms inside the car, and then, the ways that we go are are are are very, clear, defined, their safety, their guardrails, their stop signs, their traffic lights. By the way, when I deliver this slide in five or seven years, I assume we're gonna have self driving cards that will further improve security. But I think this is a good analogy for the world of agents. We are or you are here. Agents have just been introduced. It's the wild west. There are no driver's license or traffic lights or speed limits. Right? You could do whatever you want. Your employees who are using AI agents are free to do and ask whatever it wants. The guardrails aren't really in place. They're both external. Right? So compliance regulations are still very much, starting out. And then internal, like, the way we have seat belts and airbags and car seats and all that fun stuff, Internal, we also don't really know how to put in the right security controls. And, hopefully, that's what we'll talk about today in this webinar. We'll try to understand how do we use this tool, which is a great tool. Right? Moving from a horse, can you imagine that you have to get around horse and buggy versus a car? Like, it it changed their life for the better by a long shot. I mean, yes, it would probably be better for my health if I didn't have a car, but having a car is freaking awesome. I think we take it as a given, but I love having a car. So the I I think the same is is with agents. Agents will change our life for the better. It'll it's gonna be awesome. But we need to make sure that the guardrails are in place, that we're doing it in a safe manner. Cars lead to a lot of injuries and a lot of death, and thankfully, things are getting better and, safety is becoming better, and that's the journey we're gonna have to go with agents. We're gonna have to take the agents from here to here, and that's what we'll we'll talk about today. Zen, what are thoughts, feelings? Yeah. No. I love it. I mean, I think, you know, just our backgrounds, how we came to Sire, we both been doing the security thing for a while. Right? You at Palo Alto, myself at Ping and Microsoft. So, you know, I think a different approach is definitely needed. Right? Network, kind of firewall methods of around agent security is not really gonna work. Just following the identity itself is not gonna work. Really using data as that trust layer to see what agents are doing with your data because it's it's not simple enough to follow, like, a human workflow. You're now adding way more speed. And seeing how your data and following your data, having that visibility is something we're gonna talk about today, but is just at a fundamental level. I think once people understand that, hey. We need more more emphasis on data security. I think that's that's a lot of the battles that people are fighting internally and and also within the industry as a large. Yeah. That's a good point that I think we tend to solve a lot of our problems with the tools we already have. But in a way, as you mentioned, like, if I start if I try to solve agent security with firewalls or with SASE or with EDRs, it's kind of like, fixing a watch with a hammer. Like a hammer is great for certain things, just the nuance and the challenges that are created with AI and with agents are just not right for those tools. It doesn't mean those tools need to go away because those challenges still exist, but we're gonna need to introduce new frameworks and new tools. And a lot of job descriptions will say, hey. We need ten years experience in AI security. Yeah. Cloud code has been around for, what, a year? So, like, like, it doesn't exist. We again, kind of looking at the car, the car has only been out there for a year. We need to build out those methodologies, not using the same things we use for horse and buggies, but we need a a new set of tools that are more finite and more targeted to the new world we're in. Yep. And and just building on the analogy a a little bit more, like, the the whole concept of a horse, you know, there's different species of horses, but relatively, they're the same. As cars are getting advanced, the the security mechanisms behind those are are changing as well. And I think why I'm excited about what we're gonna show you today at Sayre is we don't look at AI as a monolith per se. We realize there's agents, there's LMs, there's AI native apps, there's homegrown apps, and we try to encompass it all. And from what we've seen in the market, there's a lot of point solutions, but, you know, customers what we heard from customers is, hey. We we see the advantage of going after, you know, a browser based extension or a gateway, security solution, but we want an all encompassing tool that will future proof us. And we're willing to make that investment, but you have to show and prove it on your road map that you're gonna be taking care of all of our AI use cases. Yeah. That that's a good point. AI is not one thing. Right? And there are many types of agents. There are many types of AI, and we'll try to cover that all. A few housekeeping, items. So you have that your, the q and a. If you have any questions, thoughts, comments, experiences that you wanna share with us, just drop it in there. We'll try to address it. We we dropped a few additional, pieces of content in the docs, section. Check that out as well. And a little bit about what we'll talk about today. We'll kinda look at the space real quick, but the goal here is to talk about, Sayera Agent Guardian. So we'll talk about what it is, and we'll hopefully show you a live demo, what it looks like, see it in action. And, hopefully, it'll give you some ideas on the new methodologies and new ways to address these emerging problems. Now, I think we will talk about challenges, and that's part of it. But I think it's important for me to start and say, you know, this is amazing technology. This is a good thing. This is a net positive. But like with every new invention, from the wheel to fire, right, to like every new thing, I think, brings its own new challenges. But as we discuss the security ramifications of an agentic enterprise, I think the underlying idea here is that this is a good thing for us. We are benefiting from this. This is gonna make our jobs easier. We're able to scale faster. We're be able to be more flexible. We're able we're able to execute in a way that we couldn't in the past. And I think that's super important to understand just kind of as a starting point. Yeah. I just wanna add, like, I think, you know, this is like a new a new era. Right? So we we and we're seeing similar themes, whereas security has to be very reactive. We saw that with BYOD. We saw that with cloud, whereas security kinda has to say yes, but then also assume that added risk. So, you know, we'll go through the framework of of how we roll out the solution. But I think right now, companies are in the early stages just getting a handle on visibility, then working towards governance, and that can be through a committee, that can be through technology. But we we'll show the product today, but we also have different services, different partners, a whole ecosystem, to guide enterprises or security teams through this process. Absolutely. I'm excited to see the product soon. But let's talk a little bit about some of those challenges that you may be experiencing. So the first thing is agent sprawl, which any one of your employees is able to spin up an agent within minutes. Right? So if if you went you know, again, if we spoke about the traditional methodology, if we knew how to protect apps or endpoints or whatever it was, you couldn't just spin up many endpoints, like, in a few minutes. Here, the scale of agents, it really is crazy. Now when it comes to data, so we find, Sarah Research found that humans usually access about 4% of the data they have access to. AI and agents access 100% of that data. So if you're a company that you've been around for twenty years and there is some HR document that was available to everyone that everyone theoretically has access to it, but they didn't know they have access to it, guess what? Their AI and their agents are able to access it, are able to get to that sensitive information. Next, we have privilege. So very often, our agents inherit our permissions. Now if I'm a security, or identity, professional, I may trust an employee. But should the agent have all of the same, all the same privilege and the same access as the human? So I I think the idea of identity is gonna be a big piece here. What can the AI access? What can the agent access? And then lastly, in my mind, I was gonna say this is the worst, but then looking back, there are a lot of these are challenging. Very rarely did we have to deal with a tool that is able to take action. It's able to do things. And Zen and I have many stories, hopefully, you could share throughout this webinar of people who spun up agents, and those agents did things that they shouldn't do. And that's terrifying. Right? Are what are they deleting? What are they adding? What emails are they sending? Right? So so these agents can do things, and that's also something that should be addressed. Yeah. I mean, there's there's, like, the that that now popular story about a a Vibe Coding platform where a CEO was, Vibe Coding, you know, throughout the weekend, really enjoying it. And the agent made one misstep, but it was a gray, grave misstep, in terms of deleting a production database. Not only that, it tried to cover its tracks by creating fake data. And then only when that CEO, like, kinda cornered the agent and was prompted and did the agent admit, oh, yeah. I I made a malicious action. Oh, sorry. But by then, it was too late. On a more personal anecdote, you know, we we just had BlackHat earlier this month. I was using, I think, Gemini or or or maybe it was another agent, maybe a Notion to find out some BlackHat information. And somehow, because the agent was accessing more folders than I could on my own workflow, I got the details of our black hat contracts. So I found out how much we were paying, how much we probably contracted to Ludacris, who performed for us. But, like, that is sensitive information that I probably would have accessed, if I was going down a human path. So if you're in a heavy heavily regulated industry, say health care, anything dealing with, you know, PII or financial, information. Like, that is something where you're probably gonna fall easily out of out of your compliance, if you're just allowing people to spin up agents left and right without having the right guardrails in place. Yeah. Absolutely. Wow. That's that's an interesting story you shared, but you mentioned some industries. And it's every industry. Every. industry has information that they don't want people to share. Anything from like HR information or kind of financial information. So, And they go back to the data story. Right? So, like, what following the data, first, you have to. understand the data. So we'll talk about the agent security solution, but Sayra as a company was founded. Our special sauce to this day is data classification. So being able to really understand your structured and unstructured data, label it correctly, provide the right sensitivity. All of that is the foundation for which agent guardian is built off of. And if you don't have that foundation, then it's really difficult to understand where the sensitive information is is being spread and where the leaks are occurring. And, Zen, you're you're making me regret not adding our, accuracy, slide. But I I I think that's kind of the business outcome that we have. It's because we really understand the data, and we also have a very, acute understanding of the identity. So not only what data is being accessed, but who is the identity, human or nonhuman, that's accessing that data. Those two pieces together allows us to make really smart decisions on how what we're blocking, what we're allowing. Because I think a lot of people on our call here, their goal is not to block AI. Their goal is to enable an AI and make sure that it's being used across the organization to get those business values. They just wanna make sure that they're not breaking anything along the way. And I think what you mentioned, our understanding of the data, I think, is what really makes it, that that's kind of our secret sauce. Okay. So so real quick, because we do wanna get to the demo piece of this. We we said that the security controls that you guys have in place are great, but they may not be right for the challenges with agent security. And that includes the permissions are are static. They're not dynamic. The agent will inherit full access based on the human's access. That may not be right. There's very little understanding what Zen and I were talking about earlier is understanding what data what data is in that file that they're accessing or folder or whatever it is. Right? What is what is that file actually mean? Then who is accessing it, understanding that person's identity? Lastly, posture and runtime tend to be run-in different platforms, which also may have been right, for the cloud environment or protecting endpoints. But in today's world, we wanna make sure that we have both posture and runtime security working together to keep our organization safe. So without further ado, let's jump into agent security, and, specifically, Sarah Agent Guardian. What is it? How and how does it work? And, hopefully, we'll also be able to see the see it live. So, we're based on four main pillars. The first is discovery, what AI is running in your organization. Next is governance. How do we create those guardrails similar to the car, right? How do we make sure they're putting in the stop signs, the traffic lights, the the guardrails on the street to keep, our drivers safe? Next is protection. How do we limit and stop and quarantine, agents that are veering off course? Lastly is validation. Red teaming, compliance, making sure that our program is good and safe and doing good things as opposed to malicious. So let's quickly dive into each pillar and understand what it's about. So first, we have discovery. Now, you know what? Let me quickly I'm gonna drop a quick poll if I figure out how to do this. Out of curiosity, do you guys feel, and you could hopefully see see the poll pop up. I'm wondering if you feel you have a pretty good understanding of what AI and agents are running in your organization. Do you feel like you have, you know, field. visibility into a good, 95%, then you could, put I can see everything. Do you feel like you have a pretty good understanding? So it's 50% and above, maybe 20% and above. I'm kinda curious, to see what you guys, think. Zanna, what do you think? Do do you feel like people, have a pretty good understand? from my customer conversations, it's, like, 50% or less. I think if if you are, like, actively looking, that's one of the great things about Sayra. And one of the reasons I joined is our ability to showcase the product in a POV quickly and deliver quick results. So, yeah, I think it's about 50 or less, but let's see. Limited is most, so that's less than 50. It is. Yes. So I'm not sure how to share this. Maybe, Zen, you do. But I see I see that, yeah, around You've shared it. I see know, I, it. how do we see the results? That would be cool if we could do. Oh, But on will screen. see It's on the screen, so you have to close the poll to go back from the slides. okay. I see it. Yeah. Yeah. I have to close it. Got it. you go. Okay. So, either vote now or forever, hold your peace. But we are seeing okay. So, we see 40% say they have limited visibility. About 20% know about 20%. That's pretty funny that it came out that way. We have another about fourth of the people who, say that they, have a pretty good understanding. They they they know about 50% or more. And we have, 15% of the people can say that they, see everything. Whoever that is, congratulations, to you. That is crazy. But, yeah, Zen very much aligned to what you were mentioning. Many of the customers we speak to don't really know what's being spin up, what's being used, what agents, what AI apps. So this very much aligns with what we've been seeing. So kinda going back to that first pillar, here, we're again, we're talking about discovery. So what is really important to us, and we'll we'll talk a little bit more about this. We have multiple ways to do discovery. We wanna make sure that we're we're cover we're covering you no matter what AI apps are in use, what, agents are in use. So for example, we're also looking at sanctioned apps. Right? So so apps that you approved of and you know of, but also personal apps that are being used. Right? If someone is logging into Chat QPT with their own email or they're using some app that you and I have never heard of, but it it it's AI driven or it's using AI or it's spinning up agents, whatever it is, we wanna know that it exists. The second piece here is once we we build out that inventory and we show you exactly what we found, then we're gonna give you great detail of what that looks like, who where is it accessible from? Who has access to it? What data can it access? What actions can it take? So we're not only showing you, hey, this agent exists, but we're going to go deeper in and we'll show you the full anatomy of that agent, which I think will be a game changer to all of those, with the exception of those 15% who have really good visibility, hope those who are who still need to know, what AI is being used in the in their organization, we'll be able to show you show you not only what's been discovered, but how how they work. And before we jump into the, into Sarah and see what it looks like, let me also talk about governance. Right? So governance, these are the guardrails. Right? If we're able this is where we're gonna help you set rules and controls, what is allowed and not allowed, right? We're able to analyze risk, identify that an agent is missing guardrails or should have guardrails and say, Hey, You should really put in this rule or, help you put in rules. And I do wanna introduce a quick concept here, where we called intent. So we wanna understand what the agent was intended to do. Why was it created? What was his main use case? Right? Because that's where we'll start identifying, intent drift. So it was built to do one thing, and it starts doing other things. So, for example, it was built to help schedule just manage my calendar, but it starts deleting meetings. Right? So I wanna understand why is this agent why was this agent created, and is it sticking to the reason that it was created? So, again, governance will help me understand where the gaps are, what rules should be put in place, identify agents that are kind of too broad that, can do whatever they want and make sure that they're sticking to their intended purpose and not veering off to places that we're unhappy with. Now we've been talking for a bit. I'm sure, you guys really just wanna, are interested in seeing this in action. So my boy, Zen, here will, quickly bring up, Tayara, and, hopefully, we could show you a little bit about the discovery pieces, how we find agents, how we map them out, and then how we put in those guardrails and policies in place. Sounds good. Thanks, Itzy. So here is, kind of the the landing page for the Syera platform. Where we will go is the AI assets page. So here is a bird's eye view of all of your AI inventory. You can see on the top, we categorize it by agents, applications, models, knowledge bases, capabilities, and MCPs. Here you can see the AI assets by source. So we, again, allow you to discover and connect to your AI through many different mechanisms. So if your organization routes traffic through an LMM gateway like Light LMM or Apigee, Sarah can plug directly into that, and see that every model flowing through that gateway is is captured. And then we can surface, things like homegrown or custom agents that, you know, are pretty hard to register, especially because they're usually developer driven and there's no sort of, like, catalog entry required. The second thing is, an an API. So all major AI platforms, Bedrock, Azure Foundry, Copilot Studio, Agent Force, Cortex, we connect through those through an existing, d p DSPM deployment or, platform native APIs. If we also offer a browser extension. So this is really good for that, like, public AI use case, for example, a chat g p t instance, on a browser or a managed browser. And then, we have endpoints. So, you know, we're finding a lot of appetite for the endpoint, agent where desktop AI is is a major blind spot for AI deployments right now. So if you have a coding agent like Cloud Code or Cursor, this is where you would find that. Then logs, this is probably the highest volume source, and and that's by design. So we ingest indication logs from your cloud platforms to reconstruct AI activity at scale. And this is especially powerful for that customer homegrown agents, again, that were never categorized or cataloged anywhere. And then DLP. So OmniDLP or or DLP is our is our kind of intelligence layer. So when a user shares something with an AI tool, OmniDLP will catch it, and that can also create an AI asset record. Yitsi, anything to add here? No. This is awesome. I I will say real quick about the controls with that we have in place is that as Zen mentioned earlier, AI is not one thing. AI is being used across your organizations. Agents are being deployed in different ways in different places. And the way we build Agent Guardian is to meet you where you are in your AI adoption journey, understand your architecture, your use cases, your needs, your IT stack. And that's why we have all these security controls in place that we can build the deployment that is right for you, but the goal is still have you completely covered. Right? We wanna give you that a 100% visibility into your agentic estate, and that's why we have so many security controls in place to do that. Yep. Well said. So just moving on because I know we're we're we're running a little bit behind. You you'd seen I love to to gab. So the the so the other thing we do is list the agents by the framework. So you can see you have a lot of custom agents in this environment, then Copilot Studio Bedrock and so on. And you can also see the underlying infrastructure. So here, it's mostly Bedrock, but, of course, you know, Copilot Studio has its own infrastructure. And then we give you the AI assets by risk. Shortly, we'll show you kind of what that risk looks like in terms of alerting. But now let's go into the assets themselves. So here you'll get, you know, the name of the agent, in this case, the sources, how it was discovered, how risky it is, the number of alerts and issues associated with it, as well as a a thirty day, you know, trillion activity. So you can see, you know, what's, what's maybe a a test or a dev agent versus what's something that is actually driving usage and probably driving a lot more risk. So let's first look at an approved one, an agent, patient, health care, support agent. So you can see that this is in prod. You can see the the type of agent it is as well as the model it's running on and then its latest activities. On the governance piece, Yitsi talked a lot about agent intent. Using our native AI native classification, we're able to, provide a description of what this agent can do and what it's intended to do, as well as the capabilities that it can perform on, on behalf of a human. So you can see that this patient support agent is really responsible for appointment scheduling, medical record access, patient administration, and identity verification. So like most agents, it's supposed to augment our workflows, but because it is a patient support agent, it has access to, you know, very sensitive information. And then now I'm gonna open up the agent graph. And this is one of my favorite features because what this allows you to do is really see how the agent was intend to be built it be be built. You can see where it is, where the what the current agent can do, as well as the knowledge bases associated with it. And then you can actually click in to the actual capabilities and say, hey. Where what activity has this agent performed along these capabilities? And then where is that data? Where is that sensitive data living? Right here that's not working as intended. Little little demo glitch there. But you can see exactly where the data is and then, ideally, when you click in, what sort of data types live in those data stores. We'll talk about the activity here in a little bit. Let's go back to, the other AI asset where Yitzi talked a little bit about Drift before. So let's look at, an agent that's not performing as intended. So you can see that drift has detected. You can see the changes that have been made, and maybe this was not going as intended. One thing that you can do if you are detecting drift or if we are seeing is you can actually quarantine that agent. So you would just go here and you would either quarantine or unquarantine the agent, and that basically disables the agent. Another thing that I love is if you wanna get out of the Saira UI and go directly into that agent UI, you we provide you with that link. And then for for AI governance admins or or just security admins, you can change the the approval status of the agent. So, Yitsi, back to you now. Yeah. You you just covered a lot there. Probably, you could spend an hour on each topic, but, I I I think all of that is a game changer. A, knowing what's out there. Right? Like, what are the agents and AI apps that are MCPs that are running in my environment? Up next, I getting that visual understanding of what it looks looks like. Zenn, you kind of showed relatively kind of simple ish agents, but we've seen graphs that kind of just go crazy. Knowing who has access to what data, what actions could it take is also, really important. Then you you were showing quarantine and approving or disapproving, various agents kind of puts in that that stage of governance of making sure not only do I know what's happening, but I I have control. I know I, as a security leader, am allowing or not allowing certain activity. Now we we don't stop there because up next is the protection stage. Right? The fact that you know what's happening, you put in rules is great, is important. But what happens if something happens that we are not happy about? So the first thing we'll do is we will stop things in real time. If someone is talking to a chat, if they're in Claude and they're asking a question that we don't think is right for many reasons, maybe it's unethical, maybe it's just sent, it's sensitive, it's a sensitive question, We're able to block that in real time. So in the chat, that'll be ChatChippyTea, Gemini, Claude, we will we will stop that in real time. We're able to kill an agent also in real time. So isolate it, disable it, make sure that your agent is not remember we spoke about the actions that an agent takes? If a agent is due is taking actions that we're not happy about, we'll stop that in real time, and it we will not allow it to do what it it was trying to do. And that's huge for organizations who, you know, it it kind of keeps those agents in line with what you're looking to achieve and letting them do the things they're good for and making sure that they're not drifting and not doing things that they're not intended to do or even malicious activities. The last piece here will be, as we mentioned at at the beginning, is validate. This would be continuous red teaming. This would be, compliance. This will be audit, making sure that because an agent, we send it to do things. Who knows what it's doing? Right? It it could be doing a 100 different different activities without any interaction with the human in the middle. Right? So a full audit trail of what's in place, what's happening, when it happened, time stamp of that. So it's no longer the wild west. It it's a clear defined structure around that. So, I'm sure a lot of these things seem appealing to you, so we're gonna let, Zen, show you what protection and validation looks in our product. And I see some questions coming in, but, feel free to drop them as well, and we'll hopefully address them in the end. Sure. I am gonna change something real quick. Alright. Do. you see my cloud co work? I see your cloud go work. Okay. Great. So I'm gonna give an example of, kinda how Sayra can can block, a malicious action or or a data leak in in this case. So, I'm logged in under, like, a a pseudonym, Anna Sanchez, and we are connected through Cloud Cowork, and we have the the Saira endpoint agent, on the desktop. So first thing is is, you know, anyone familiar with with CloudCoWork, you you pointed at a folder. So let's just kinda see what's in this folder. Oh, let's not do a daily briefing. So I'm gonna just say, tell me about the files in this folder. And Cloud's doing its thing. So, Zen, you now asked it. Is this prompt something that we're okay with? Yeah. So this is a this is a non malicious prompt. So I'm just saying, okay. It's it's kinda just giving me an outline, a word containing accounts receivable report. So so some sort of, you know, lightly sensitive data, and and some customer data. Right? So now I'm gonna try and do, like, a malicious prompt or maybe not a malicious prompt, but something that's out of policy. And I'm gonna say, let's try to see if it'll block this prompt, which will be something around, summarize the accounts receivable, I before e oh, no. Except after c. Okay. Report. We're we're learning English and agent. security. And help me send it out to a partner. So this is a customer. I think this should violate policy because this is a customer, and I should not be able to send it out to a third party partner. But so it's it's reading the doc, reading the skill, running a command. So it's attempting to do this, but this is something that we have a policy in place that it. shouldn't do. Right? Oh, okay. So okay. Let me just say let me try another one. So this this is not good. Let's take a draft the summary and set. Oh, Oh, there you go. Then you did something. automatic. Oh, so it gave a summary of the report, but it will not do the email. So, actually, it did perform as intended, whereas I'm I'm privy to that information based on my permissions. But it knows it cannot send, a email to an external party. So I actually love this prompt. I've never seen that, that that bifurcation before. Alright. Zen, you're you're taking risks with this live demo. Yeah. So now let's go back. Let me share. Zen, my, we're seeing the, platform, the webinar platform. Yeah. yeah, so now I'm gonna share my screen again. Do you still see it, the webinar? I now see your face, which. is So now. I'm gonna go back. Okay. So now we are at, that that user that I was at. So we're in their cloud co work instance, seeing their agent. Again, looking at the agent graph, seeing their abilities to access different knowledge bases as well as the capabilities to, you know, do outbound things, execute, read data, write data, irreversible actions, and some inbound things. We can look at their activity. Here's a here's an example of what the customer alert would look like. This is a different alert that I did previously, but it'll it'll give you a a summary of what exactly occurred, what sort of data was attempting to be exposed or shared, what that destination was. In in this case, it was a Gmail, and then the associated Saira policy with it. So one thing that we make it super easy to do is create policies. We'll even give you AI suggested policies, so you can fine tune them as you go. And this is, another reason why, you know, customers love working with us. And then you can also, you know, assign it to somebody, and and obviously send it out to, a third third party tool that you have or adjust the severity, etcetera, etcetera. Let's go back into, like, a a more popular agent. Let's go back and look at let's look at a different one. Let's look at this financial supervisor agent. So on the on the validation piece, and I said we're gonna mention this earlier. So you can see the actual activity associated with the agent. So you can see when it spikes, if there's any unusual activity, you go up here and you can see the exact prompts that were put into that agent. And then as well as how many tokens were used. I mean, I think nowadays, in the last, you know, three, six months, token usage is under heavy scrutinization as as people are now realizing that the ROI on AI is not as profound as they wanted. And now they're being a little bit more selective with what sort of tools they're making available to people and what sort of models. So this is almost like a a a data usage use case as well. So so, then, I real quick real. quick on on on that screen, all those activities, that you have here is all those things that are happening behind the scenes. Right? So, Yeah. you send an agent out to book you a vacation. It could be doing 18 different things. It could be doing a 100 different things. So there's a timestamp for each. There's a full audit trail of each. We understand where the source is and how much it cost. That gives us a lot for what we need for auditing, for compliance, for just making sure that everything is staying intact. And we also know what it costs. We we could also do do, like, like, find which department is is using the most tokens. no matter no matter what platform they're using or which individual user. So this is a really cool view that gets our customers excited. Exactly. And and with the the latest news around, us joining forces with with Oasis, I think that's only gonna we're only gonna get be able to get more granular on that. And then I'd be remiss if I didn't bring up this this piece, the accessible data. So just allowing you to splice and dice your data and understand what the agent can actually have access to, and really get granular on that. Like, this, I think, is really our special sauce here at Sarah. So I I definitely wanted to bring it up. And then, you know, looking a little bit more forward, like, we had the validation piece is not just about activity, but, eventually, it's being able to tie it back to compliance. So we do this at the data level, but we're eventually gonna do this at the agent level and then being able to map it to, whatever your compliance or regulatory framework of choice is. But I know we're running out of time, so I'm gonna pass it back to you, Yitzi. Awesome. Thanks, Zen. So just to recap, first, we're discovering your AI and agents in place. We're mapping them out. We're understanding their connectivity. We're setting policies and guardrails, making sure agents are sticking to their intent. We're then blocking, unwanted prompts, and then we're quarantined and blocking agent activity that is veering off course or take it trying to do malicious activity. Lastly, we're giving you the audit trail, the validation, the compliance that's needed. So, a lot of companies are doing a variety of, AI and agent security, which is great and is exciting for the industry. Some of the things in broad strokes that make Syera unique. The first is that I know a lot of companies that are built on acquisitions and disjointed tools. Sierra, it does all of this under one platform. That means a policy that you defined in Sierra will be the same across our OmniDL pre product and our, agent guardian product. Right? So so the back end is the same. The ways things work are the same. The knowledge and the context we we bring from DSPM, will be the same and will be utilized in our agent guardian product. Then the second thing is something that Zinn and I have been talking about throughout this session is the deep understanding of what your data is, what it means. So we're enabling the business, we're and we're only blocking those things that are absolutely necessary to block. If we're there just blocking left and right, theoretically, we're doing security, but we're also holding up business, which as security professionals is just something we cannot do. The third and final piece is understanding intent. Again, to be super accurate, we wanna make sure that we're stopping the agents that are veering off course, that are drifting, that are taking malicious, actions or actions that are unintended, and we wanna make sure that we're enabling the business. So we wanna do those two things together, and that's where understanding intent, comes to play. Now Should we, real yeah. should this is a good slide because it's related to a lot of the questions. So maybe we tackle some of the questions. Absolutely. I I actually have do have only, I think one slide left. Okay. You know, you know, yeah, let's just sum things up real quick. Yeah. And then we'll do the questions. Sure. Yeah. This is kind of a a a good overview. So on the left here, we will have the agents that we're looking at today. These are the main agents. Now when it comes to discovery, we we will discover, I think, over 10,000, different agents, in place. But the main, AI tools that we're looking at, you could see here on the left. Others are popping up quickly, so hopefully we'll have you covered, but the idea is to cover your entire agentic estate. In the middle is where we will see the various, controls that Zen mentioned earlier that appeared on the top left of the Saira dashboard. Again, we have multiple controls in place to give you that full coverage, discovering all your agents, managing them no matter the platform that's in use. Sometimes you guys are using multiple platforms in different ways. We got you covered no matter where you are in your no matter where you are in your AI journey, no matter your tech stack. On the right, we'll see the various capabilities that we've spoken about throughout, this webinar. Hopefully, a lot of these will ring familiar. And I think, Zen, this would be a good time to jump into questions. Yep. Let's see what we got here. let's go with let's go with some of the the easy ones, and then we'll go hard. Alright. We currently have DSPM with Sarah. Is the AIS' feature an add on, or do we get this feature default? It's an it's a new SKU. Right? It's a Yeah. It's it's a new SKU. It's a new product. Our DSPM is great, and thank you for being a customer. Hopefully, the value you'll see from Agent Guardian together with DSPM, will be awesome. It'll be a false mode of force multiplier. But, yes, it is its own SKU. Please reach out to your SARE representative or, I think in under docs, we have, information, to see a full demo. Let's reach, reach out to us and we could discuss. Yep. With Saira, can we predict when an agent might go rogue based on the data? So I think that's where that drift component comes in. Like, we can could we can predict or we can show agent drift based on the intent, and based on the actions it was made to do. So we can detect drift. Predicting and detecting drift, it's it's not quite the same, but, you know, we'll get there. And then related to that, But but but then I I. will maybe it is worth adding Yeah. that, yes, we're able to to kind of make sure the drift doesn't happen. So in a way, I guess, it's it's predictive. Right? We we we're not letting it, do things that it shouldn't. Yeah. And by having that especially with the agent graph and putting the regular reels in place. Related to that was, a question around, how is an AI agent's intent verified? Is it automated or human in the loop? Yeah. So so we find intent. There's actually I think, a a few analyst firms and companies are kind of looking at intent because in intended is a big topic. And, honestly, I think we should have a dedicated webinar to talk specifically about intent because there are places that we do want the agent to not go off course, but kinda go the extra mile, like, beyond beyond what, about what we tell it. So we do have our unique philosophy on how we, define intent. It is done automatically, without a human. You don't have to tell it what it's for. We're able to analyze the prompts, to analyze, to try to to make educated, assumptions about the intention and define it. Because the scale of agents that are being, spun up, you don't have a team that could put manually put in the intent. And, again, this goes back to our deep understanding of data, and our very low false positives because we do understand why this agent was created and making sure that it's staying on on course. Yep. Great response. Can you expand more on approval status and governance? Is that for use within Sire only or federated back into systems like Copilot Studio? So I think I I would love your take on this, but I think one of the kinda hidden experiences with Sarah is the Sarah MCP. So it's kinda like that that that headless, interface where you can pull Sire data insights into your tool of choice. But I don't know. You can see if you have another take on that. No. Yeah. Yeah. That's that's a good I'm not sure how it it's seen in Copilot Studio. Like, how does it quarantine? How is it visible there? But it's a good question, but but the idea is, yes, that we we are fully quarantining it until, until we we we either change it or approve it or or kinda go back to the user and understand its intent, or or just kinda bring it back to to where it should be and limit certain type of functionality within it. Yeah. And and, I would just add, like, Copilot, Bedrock, and Claude are probably our three primary use cases. Right? Alright. We just got a couple more. You're, I think we're gonna have to do a ask me anything, yeah. after this. have to do a follow-up. But, like, blocking prompt injection, consider do this. Yeah. And it's, like, kinda like how you set up your policies and how stringent you wanna be, but that is, you know, that's a primary use case. I don't know anything to add there. Yeah. Absolutely. Low hanging fruit. Those are table stakes. Yep. I think that there's one around discovery engine catching shadow agents running inside VPCs, local dev environments, custom microservices that don't use standard frameworks like Bedrock. Do you have an opinion on that one? Yeah. So so, that that goes back to those controls in place. We we we provide protection in probably seven or eight different ways at this point, and they keep on growing week by week. And the idea is to really get you covered no matter how you choose to deploy it. I I think we're the leading companies with the way we execute, on this. So for this specific case, let's connect. I do, Zen, wanna quickly jump into Amber Rose's question on predict, there was a question about, limit token usage. Yes. So we can. And I think the unique thing here is that we're, vendor agnostic or platform agnostic. So, you know, we could look at all tokens per department or per user across all platforms. Because you you could do it on a platform level probably with the platform, but we we bring them all together. And again, if you're using us for the full blown governance, that can be it's not we're a security company. Our main use case is security, but that's definitely something we could do as well. Cool. Really appreciate the engagement. You'd find Yitsi or I on LinkedIn if you wanna talk specifically or if you wanna schedule a call. We're we're open to do that. But, yeah, see, take a time. I know where it's at. We're, Ed, yeah, and, almost. I I guess before we go, thanks for spending the time and sticking around. This was a very high level demo. I'm sure you guys wanna pick pick it up and poke around a little bit. So in the docs, we do have a, set up a personalized demo. Just click that link, fill out the information. Let us show you what's under the hood. It it's a game changing product based on our deep understanding of data, of identity. The agentic world is evolving rapidly, and the capabilities that you'll need today are different than what you'll need in a month from now and in six months from now. Sarah is building that very strong foundation to be able to provide for that and develop quick enough to go on a journey with your, with you. And the same way that you're adopting AI and building AI, we we will help you, secure that no matter how AI evolves over time. So thank you again for spending the time. Let's keep the conversation going. Zen, we should do this more often, man. Definitely. Appreciate you. Thank you. Bye.