How Saviynt Zuma Secures AI Agents With Zero Trust

[00:00:00] Your agents aren't producing accurate answers because they don't have a complete semantic understanding of your data. And Denodo is solving this and solving it through semantic consistency. Through semantic consistency, your agents can start making accurate predictions in real time. So see what else Denodo can do by visiting denodo.com to learn more. But now let me introduce you to today's guest.

[00:00:28] What happens when an AI agent can read your sensitive data, update your business systems and even recruit other agents to complete a task? Well for CISOs it could create an identity security problem with worrying new features like this. Because AI agents they're designed to act even when their answer could be wrong.

[00:00:55] And giving one of those agents permanent administrative access. And as today's guest will put it shortly, it's game over. And I'm fortunate to be joined today by the founder and CEO of Saviynt. And they are an identity company managing over 100 million identities for more than 700 customers. And my guest will explain why every AI agent needs a human owner.

[00:01:23] Temporary permissions and controls that evaluate its intentions before even thinking about allowing any action. And he will also introduce us to something called Zuma, which is Saviynt's platform for discovering, registering and governing AI agents and other non-human identities. And at a time where we're talking about businesses having hundreds if not thousands of agents, if one employee could potentially have hundreds of agents as well on various different workflows.

[00:01:51] I think this is something that's going to continue to rise in importance over the next few years. But rather than wait for that to happen, I want to get you all in at ground level here. And share a practical four-part framework for securing agents right from development through deployment.

[00:02:10] So if your business is preparing to give AI access to real systems, I'm hoping that this conversation today will help you decide what must happen before the agent receives the keys to your business. But enough from me. I've revealed far too much already. Let me introduce you to my guest right now. So thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do? Thank you, Neil. It's great to be on the podcast with you.

[00:02:39] Sachin Nayyar. I am the founder and CEO of Savient. We are an identity security company founded about 14 years ago based out of Los Angeles with 700 plus global customers who have given us the privilege to help them and serve them, manage their identities every day. And we manage over 100 million identities today. Well, thank you for sitting down with me today. Every day on this podcast, I try and dive into a different topic and try and demystify it, put it in a language everyone can understand.

[00:03:08] And identities is such a massive topic now. And I'd love to dive in straight away here. I mean, how do you secure a modern business when identities are no longer only human employees, but also partners, machines, applications, and increasingly autonomous AI agents? You must have seen so many big changes here. But where do you begin to secure all of these things? First, let's step back and define the context a little bit. Yeah.

[00:03:35] What as companies, you know, most of the organizations in the world, all organizations in the world have been working with AI to figure out how AI can add value to their business. Right. And as these AI workloads or AI deployments come to production, the biggest question that is now being answered or being the biggest question that is being asked, not answered, being asked is, how do you manage access of these agents?

[00:04:04] How do you manage access of these co-pilots? So when I start with a simple co-pilot example, you know, where all of this journey started for all of us, and if there is an enterprise co-pilot, let's say it's giving us, it has access to our HR data. Right. And you and me both ask the question, what is the salary of our employees? Right. Now, you should get a different answer if you are a manager, only for your team, if you're allowed to see that based on the jurisdiction.

[00:04:34] If I am the CEO, I should be able to see it across the board, again, based on my jurisdiction. And now that is the simplest answer. But what happened? We both asked the same question to AI, but we got different answers. Right. And that logic is not built within the AI core model. It is built within a gateway or a firewall or a proxy that answer and that questions are going through. Right.

[00:05:04] And there are layers and layers of how that was determined at runtime with the latency that doesn't make it unusable to the end user. Right. So what we are now doing in this example is adding context that you are a manager in the US. You are a manager in London. Maybe in London, you are not able to access that data because of some of the GDPR or some of the regulation. In the US, yes, I can show it to you. So there is context on who you are. Are you a manager?

[00:05:33] What is the policy for that region? And all of this is being applied at runtime or in real time to that answer. Right. So this is this is a simple example. Now extrapolate that to agents where you are asking your Microsoft copilot or your ServiceNow or your Salesforce agent to go and do something on your behalf. What is what is the access of that agent?

[00:05:58] So what we have seen is within a universe like Microsoft, if you're using Microsoft agents to access Microsoft universe, then Microsoft can give you and say this agent acts on your behalf. OK, that means the agent inherits all your access. Right. So all your read access and write access. Again, it might sound very simple answer, very benign. But but as humans, we understand what we are trying to do. An agent is literally taking the prompt.

[00:06:26] And if you give it access to not just read critical data, even though that you have access to, but write critical data, the results could be completely unwarranted that you're looking for. Extrapolate that to now moving from one universe like a Microsoft universe to a copilot agent talking to Salesforce or a copilot agent talking to your ITSM or your Jira or your ServiceNow system. How is that agent accessing those systems?

[00:06:54] That's now outside of Microsoft jurisdiction. So how will you grant access to that agent to access those systems? The moment you say a lot of companies talk about the term application API, which means the agent has an account in the system. That's an administrator account. The one of the key things that your listeners can take away from this podcast with me is if you ever give your agent administrative account to a system game over. Right.

[00:07:22] Do not give your agent administrative access to a system. Do not do that. Right. So how do you how do you grant that access? Then extrapolate that further. We are now talking about one agent talking to multiple agents. So you start something with one agent and all of a sudden that agent thinks that you and all of these agents are biased for action. That's how the LLMs have been created. They are biased.

[00:07:45] If you see in all of our day to day, even if sometimes the agent doesn't have the right answer, it always pretends that it has the right answer. That's what we've been calling hallucination and all of these pieces. But they're biased for action. Now, one agent can use a network of agents to get its job done. So what about that access management? So the problem really is what are the guardrails that we put in place for for a copilot or an agent and agent to agent transaction and agent going across applications or infrastructure.

[00:08:15] All of that has to be put into these access management guardrails that have to be implemented at runtime with with and always have to be available and always have to have very low latency. So it doesn't impact the end user behavior. So that's that's the problem we are trying to solve here.

[00:08:34] And the problem and to step back, that problem starts with discovering where the agents are, because the agents are being built everywhere within the apps, within Microsoft, within ServiceNow, within a Lang chain, within AWS, Google, Azure, Vertex, all of those. So many different places, GitHub that your agents are sitting in. First and foremost, our job is to give you visibility into what your agents, who your agents are and what access they have.

[00:09:02] So your viewers would be very excited to know that Saviynt believes that this is table stakes and every enterprise should have this access. And as of this moment, any customers that are existing customers or any customers that sign up with us, we provide access visibility of agents for no cost or that is included as part of our platform for our customers. Right. So because we believe that's table stakes and customers should be doing that. So I described the problem.

[00:09:31] The answer is first is discover your agents and not just the agents, but what access do they have? Second is now define ownership of that agent. Right. Again, a second key point for your users to take away, an agent without an owner should not exist in your environment. Right. And we are talking about human owner. Right.

[00:09:53] The ownership is a critical concept because this is where we build the bridge between an agent or a non-human identity with a human. Right. Right. Because we want somebody in the organization accountable for the actions of this agent. One person. Right. So the second part is, first part is visibility and access and which results in a registry, which is also a key point. Having a single registry, which farms all these agents.

[00:10:23] Second is ownership, definition of ownership. Once you define the ownership, that's a big step you have taken. And from there, we can now establish governance that the human owner is responsible to review the transactions of the agent on a regular basis. To review, is the agent live or was it temporary or ephemeral? So all of those pieces now can be moved to a human owner who is now clearly accountable for that agent and is signing up on the access and the activity of the agent on a regular basis.

[00:10:53] The third piece is now how am I going to give this agent access when this agent needs to do its job? So what we third describe is just in time access. Like I defined earlier, cannot have standing privileges, cannot have admin privileges. So the most that agent should be able to do at design time is what the user on behalf of that user. Right. That's that's the most.

[00:11:17] But so third is just in giving that agent just in time access for the work for the work it needs to do. The fourth is enforcing access at runtime based on what the purpose of the agent is and the intent of the agent based on the prompt or what what was that agent designed for to limit that subset of access that your user has on behalf of which the agent is acting. We also want to limit that further based on what is the intent of this operation.

[00:11:46] Did I just ask you to clean up something? That means recommend something, not just go and make changes. And if you are making any major changes in my environment or pulling down critical information, let me review that. So four parts to it, discovery and access. Right. Alongside creating a registry. Part one. Part two, ownership, which results in all the lifecycle and governance capabilities.

[00:12:12] Three, just in time access or ephemeral access only for agents. Fourth is runtime control based on intent and tying it all back to all the other components. Does the agent have an owner? Does the you know what what was the discovery and is the human owner signing off on the access? All of that coming together and runtime and having a central kill switch and everything going through one service or a gateway and access gateway.

[00:12:40] Let's not get confused with different types of gateway and access gateway to control its access. That's how we believe today that we can solve this problem for our customers and enable them to adopt AI, move their workload to production in a secure way. Wow. So many big points there. And we will have a few people listening thinking we're in the very early days of this stuff yet.

[00:13:02] But I think it's so important to highlight that we're actually preparing for a world where digital identities could dramatically outnumber human employees, especially when you consider an average employee could have, I don't know, 10, 15, 20 agents themselves to work through their workflow. So how do you see this explosion of human and machine identities? What do you see that meaning for the traditional identity security model? And where are the biggest gaps that you're seeing appearing right now?

[00:13:33] So another important point I want to make for your viewers, we've all been talking about an agent being an identity, which is now factually established. And we are on the same page. I would also bring up another point that an agent is also an application because it is an application, meaning we can control the access within it. We have the ability to code the access based on certain criteria and actions that the agent will take.

[00:14:02] So like you said, 15 to 20, I've heard numbers of 82, I've heard numbers over 100 agents per user. But listen, I think the goal is not to worry so much about the number. We are directionally accurate that we will have multiple agents per human user in the enterprise. That's a given, right?

[00:14:25] I think now, but the way to solve it is with that framework of visibility and posture, access, visibility and posture, which is not all visibility and posture is the same. I'm talking specifically about access, visibility and posture, defining ownership and defining the policies and the enforcement.

[00:14:44] So ownership, the enforcement so that if you don't have the ownership or something has not been approved or something is acting outside its normal behavior is shut down immediately. And then the whole intent-based scoping of access is critical. So back to the same point, but my point is the same. Once we established this framework, we could have one to 20, we could have one to 80. And as long as the framework scales, this is the same we saw with cloud, right? We had very few services on-prem.

[00:15:13] We went to the cloud, the number of services, the number of certificates exploded. But what we did was, but also what these tools do with AI like cloud, it provides us a lot of plumbing and telemetry so we can do things right. I'm actually excited. I believe our AI agents and AI workloads will be way more secure than our human workloads have been

[00:15:38] because we are now going to build security at design time and at runtime and right into the code and we are going to shift controls left. I want to talk more about that in some of the other conversations we have. But I think we are moving towards a much more secure world, which is not a very common thinking.

[00:15:59] But I really believe that AI is giving us the tools to provide security at code time or move security controls to the left. And we will have a much more secure world. So I'm less worried about 20 versus 82. I think we do the right framework and we will give ourselves the opportunity to see the value of the productivity that AI brings to us. And it's so refreshing to hear you talk about this like that.

[00:16:25] And before you join me on the podcast today, I was reading how some of your customers are telling you that they don't need another dashboard and they're driving many of these big changes. And for example, I was reading that with Zoomer, you're bridging together insights, access and governance all into a single control plane.

[00:16:44] So again, for people listening, can you expand on why visibility alone is no longer enough and how enterprises can move from simply discovering AI agents and non-human identities to actively controlling what they can do in real time? Yes, absolutely. So hopefully our customers have seen or your users have seen our announcement on Zuma.

[00:17:07] We are very excited to announce SavientZuma, which is going to be a single platform for all identity needs for AI agents and non-human identities. And to your point, visibility is just a starting point. And we have to move from visibility to posture to fixing the issues very quickly, ownership and registry, lifecycle management, just-in-time access and runtime enforcement. Back to the same framework.

[00:17:36] None of them are enough by themselves. And in the world of AI, all the identity services that we are all used to, the warehouse, which is the visibility, then the posture, then the lifecycle management, access certification, segregation of duty rules, access request, continuous certifications, all of those have to come together for the agents and their owners.

[00:18:01] Then we start talking about certificate lifecycle management with quantum and non-human lifecycle management come together and how AI will access these systems with credential-less and just-in-time access comes there. And then how will we control all of that through one sort of narrow part of the hourglass, you know, through a single gateway, which is access context-aware and authorization and governance-aware. All of this comes together.

[00:18:29] And what we believe is customers will not be able to buy a service or not buy, but implement a service from one vendor, another service from another vendor, another something homegrown, something else they build. Because what Savian's value proposition is Savian provides all of these capabilities fully integrated, built by us from ground up. We haven't put bad solutions together or different products together. All of this was built by us ground up.

[00:18:57] So all of these capabilities that several of us in the world of identity have been doing for maybe two to three decades now has to all come together and work in real time at the speed of AI. Visibility is definitely not enough because the moment you see visibility, the next moment you will have to fix it. The next moment you will have to define ownership, build your registry. The next moment you'll be like, how do I enforce this at runtime?

[00:19:25] So I need my enforcement capability. The next question you will ask, and all of this will happen over a span of five minutes, right? And the next question you will ask is, how are these agents getting access to all the systems at runtime, which is all just-in-time access? So all of the identity services that we talk about have to come together. Again, the good news is, these are all available through Zuma, single platform, lightweight, usable today for agents.

[00:19:55] And it connects with all the downstream and meets all the needs of the CISO, of the chief risk officer, and the chief compliance officer. Exciting times. It really is. And one concern I do read about with rapid AI adoption now is the emergence of eventually unmanaged or orphaned agents, where people may have left the organization. These agents are left and nobody clearly owns them anymore.

[00:20:18] So how can businesses better discover those identities, establish who is accountable or who should be accountable for them, and manage them right the way through from creation and deployment through to eventual decommissioning? Again, big part of IT, this. A lot of people have not thought about it. There are concerns there. But how do you see this evolving?

[00:20:39] Yeah, first, there should absolutely be no agents running in an enterprise that do not have owners or whose owners have parted ways with the organization, right? That's the recipe for disaster. And that's where we now start with us having automated lifecycle management of the owner. So when one owner leaves, the agent already moves based on certain rules to the next owner.

[00:21:09] Now you have human in the loop. That owner can accept it or suggest somebody else. But there is very little period of time. And none of this is humanly triggered. It's all automatically triggered and happens at real time when one owner leaves immediately. That process goes into account.

[00:21:29] But also we have controls on the enforcement side that the moment that agent tries to make a call or do something and that agent doesn't have an owner or the owner left, that action has to be blocked because now you're running an unmanaged, ungoverned action in your enterprise. Remember, all of this will happen within seconds or minutes or in real time.

[00:21:51] But the point is that we have controls both at the lifecycle of the design layer that if a human leaves or somebody creates an agent, it just cannot start working without an ownership definition. And at the enforcement point where the moment the agent makes a call to something and some of those capabilities of an owner or the lifecycle management are missing,

[00:22:14] that agent is stopped from taking any action and the right team is notified immediately. So I think through all of this, I want to keep repeating the point that we will be much more secure as a community based with AI as we move, because we have the ability to put all these controls and we are designing the future of IT and identity controls belong at design time and need to be shifted left.

[00:22:44] And before you join me on the podcast today, I was looking for a way of maybe trying to bring to life what we're talking about here with real world use cases. And when doing a little research, I came across your work with Langchain, which looks at securing AI agents right across their lifecycle from code to decommissioning. So what does that secure by design look like for an AI agent and how early should identity permissions,

[00:23:10] governance and audibility enter that development process right from the very beginning? So listen, we are very excited about our partnership with Langchain and with other agent orchestration and build platforms like AWS, AWS agent core, Google vertex, Microsoft Azure's agent, agent build gateways or build platforms, crew AI and several others, right?

[00:23:38] Also the agent build platforms orchestration platforms from ServiceNow as well as Salesforce. The point with all of this is, again, I think it's third or fourth point for your listeners to take away from today is identity controls need to shift left. Shift left, shift left identity controls. We are all seeing vulnerability shift left. Identity controls need to shift left. So as organization or developers are building these new agents and applications,

[00:24:06] we need to embed enterprise-grade identity controls within them. So what Savient has with Zuma is all of, again, this entire identity set of capabilities working together in unison and now available for the agent platforms to embed this at code time, at build time, right? And that's where the world is going and that's where the world has to go. So we are very excited about our partnership with Langchain,

[00:24:36] where we are working with them to put these controls at build time. And similar controls are being worked on with some of the other leading agent build platforms in the industry. Yeah, I was also reading you work with so many big names. You're also expanding partnerships with companies like Zscaler to bring identity security and inline enforcement all closer together.

[00:25:00] So why is this combination becoming more important for identity-centric, zero trust? And what can integrated platforms maybe achieve that separate security tools that organizations have just cannot meet? Yeah, great question. So listen, we are very excited and proud of our partnership with Zscaler. AI agent management is a zero trust problem. There has never been a problem bigger than AI.

[00:25:28] There's never been another bigger opportunity for zero trust architecture to solve than AI agents. So as part of Zscaler, zero trust architecture, Zscaler has great capabilities, inline capabilities, right? At the network level. What Savian does is provide Zscaler a lot of identity context. Back to the point, what does that agent have access to? Who's the owner of that agent? What is the agent doing in the environment? Are there any access or activity-related behavior anomalies?

[00:25:58] And what should that agent access be within that data? If the agent is accessing your SAP environment or if the agent is accessing Snowflake or Databricks, which has your SAP data, what should the access be? Because within SAP, for folks who know, we have very, very well-defined access permissions called authorization objects. Like very complex, very deep. Saviant understands that.

[00:26:22] So when an agent accesses SAP or the SAP data is moved over to Snowflake or to Databricks or another data warehouse, they do not, that warehouse doesn't understand that permission model from SAP. So Saviant understands that. So at runtime, Zscaler does the enforcement of what the agent should be doing. Saviant provides Zscaler all this context of the agent at runtime through that Zscaler is able to call.

[00:26:52] Also, Saviant is providing all the front-end access requests for privileged access users, human users, as well as agents, and all the backend compliance and certification of access for the agents, as well as human users to Zscaler. So we have a very strong integration and an integrated capability with Zscaler for our customers and Zscaler customers that customers should try it out.

[00:27:17] And I think it will really move the zero-trust concept to the next level and make it applicable for agents, non-human identities, and privileged users. And finally, my research on you guys also uncovered that Saviant has now surpassed $300 million in annual reoccurring revenue. So, I mean, what is that growth telling you about how seriously enterprises are taking AI and non-human identity risk?

[00:27:44] And for any CISOs listening and board members, what should they be doing to prepare for this world where autonomous identities are operating across every part of their business? What would you advise those people listening? Yes, thank you. So, listen, we are very excited and honored to have crossed $300 million in annual recurring revenue while still maintaining over a 30% growth and maintaining over 96% customer retention.

[00:28:13] That's a testament to the work that Saviant team members have been doing. It's a testament to our customers who have supported us and our partners who have supported us. So I want to start by saying that. Now, what does that mean? It means that our customer, but remember, we are building a startup at scale.

[00:28:30] Saviant, we run the company startup at scale, which means the same energy, focus, speed of a startup, but the resilience, the quality, and the depth of a company at scale or an enterprise-grade company.

[00:28:46] So with the combination of these two, in addition to the $300 million ARR, we also raised over $700 million at close to $3 billion valuation recently. So the combination of these two things should tell our customers that Saviant is a very strong financial company. The industry, the investors, and the financial industry believes in Saviant's vision and our execution, and we are a safe bet.

[00:29:15] But now add our innovation, the whole Zuma launch with it, of what we are doing across the board. One of the most comprehensive identity security capability for AI and NHI of a company of our scale or a company at scale, a vendor at scale. Customers should feel very confident and comfortable looking at Saviynt, evaluating our tool sets, knowing that you have the financial stability of a strong company, but the innovation of a startup.

[00:29:45] That's what we've been trying to bring in. And I'm very excited and humbled to say that we've been successful at that and want to continue down that path. Right now, everything I talked about with Zuma, we are absolutely providing the A-plus level capability for complete end-to-end identity management of AI agents and non-human identities from build time to design time to run time. Right? And customers should take us for a spin. Wow.

[00:30:14] So many big takeaways there. And I cannot thank you enough for taking the time to sit down with me and talk about how at Saviant, you're tackling the explosion of human and machine identities. It's something that we see on our news feeds every single day at the moment. But hearing how you're managing this, it just delivers so much value. And I expect people listening will want to dig a little bit deeper. Even though we've covered a lot, we still only scratched the surface. So anybody wanting to learn more about Saviant, some of the big announcements that have come out recently,

[00:30:43] and how they can keep up to speed or even contact you or your team? Where should they go? Start with our website. If you need to reach out to me, LinkedIn is great. I'm very responsive on LinkedIn. Or reach out to your nearest Saviant representative. We are everywhere, and so are our partners. So reach out. But listen, I want to end by saying we all know these are exciting times.

[00:31:03] I want to say that we have an ability to build a much more secure enterprise by moving a lot of controls left at build time and reinforcing them at design time and making sure we incorporate them at runtime.

[00:31:22] And we have thought about this end to end to end. Please reach out to us. Please reach out to us. Work with us. We are practitioners. We are hands-on practitioners, including myself. We are keen to partner with you. We've built a lot. None of us have all the answers.

[00:31:48] But I think we have enough to put the right framework in place from where we can all feel quite good about our next steps together. Thank you. Thank you, Neil, for the opportunity. Well, thank you for sitting down with me today. There is so much noise around human and machine identities at the moment. But it was so great to hear you talk about a language everyone can understand. And most importantly, the excitement and the opportunity to build a more secure future right from the design phase at the very beginning. So many big takeaways.

[00:32:17] I'd invite people listening to feedback. Let me know how you're managing this and what you're going to be taking away. And you can connect with me at techtalksnetwork.com. But a big thank you to you for starting this conversation today. Let's hope we can keep it going and use this opportunity to build that more secure future. But thanks for sitting down with me today. Brilliant. Thank you. I'm excited. Thanks all. Thank you for the time. I think our guest today left us with a direct warning. Yes, it is exciting times. And there's a lot to be excited about.

[00:32:47] But to begin with, never give an AI agent permanent admin access. And a big thank you to him for offering a solution, a useful framework for every security leader listening today that will allow you to discover every agent, register that agent, assign one accountable human owner, and provide temporary access for specific tasks and enforce the policy while that agent is working.

[00:33:13] And I think it's that ownership point that deserves particular attention. Employees leave, projects end, and agents can quite easily remain behind without an approach like this. And Saviant's approach automatically begins reassignment when an owner departs and blocks actions when the agent no longer has an accountable human ownership. So there isn't any AI ghost employees quietly rearranging the company on a Sunday afternoon.

[00:33:44] What I love more than anything about this conversation today is it was refreshingly optimistic. Because identity controls can be embedded while agents are developed, AI workloads could eventually become better governed than many human ones in your organization today. So remember, you can learn more about Zuma and Saviant's wider identity security work through the company website or connect with Sashing over on LinkedIn.

[00:34:13] But before I go, question for you. Could your business identify every single AI agent and its owner and its permissions? And can you do that right now? Let me know your thoughts. That's techtalksnetwork.com. Same as always. And I'll be back again tomorrow with another topic. Thanks for listening. Bye for now.