RSAC 2026 Conference | Jeetu Patel, Cisco

Clips
More from RSAC 2026 Conference

Jeetu Patel

President & CPO

Cisco

play_circle_outline Trusted Delegation in the Age of Autonomous Digital Coworkers: Addressing the Wisdom Gap at Trillion-Agent Scale

play_circle_outline Agents' autonomous duration expands from minutes to months, increasing token and compute consumption

play_circle_outline Workforce shift: AI-fluent versus non-fluent divide; urgent need for large-scale AI education

play_circle_outline Action Control Over Access: Interface-First Agent Deployment with Continuous Full-Stack Observability

play_circle_outline Agentic SOCs required—security operations must operate at machine scale, not human scale

play_circle_outline Agent identity requires temporal, granular, just-in-time least-privilege permissions and revocation

Info
Transcript

Jeetu Patel, Cisco

Jeetu Patel

President & CPO Cisco

In this interview from RSAC 2026, Jeetu Patel, president and chief product officer of Cisco, joins theCUBE's Dave Vellante to discuss why the rise of autonomous AI agents demands a fundamental rethink of security architecture — from access control to action control. Patel frames autonomous agents not as software tools but as digital coworkers — capable and fast, yet prone to consequential mistakes without proper guardrails. He explains why the scale of the agentic era defies prior assumptions: with 10 to 1,000 agents per human operating around the clock, orga... Read more

explore Keep Exploring

What is the impact of the shift from chatbots to autonomous AI agents on how organizations must delegate work, build trust, and redesign their systems and architectures? add

How will autonomous agents that can execute tasks for long durations (from hours to weeks or months) change the nature of work, the role of humans (managing vs. doing), and the infrastructure/token requirements needed to support them? add

How can society close the gap between people who are AI-fluent and those who are not, to prevent widespread job displacement? add

What is an “agent” (as distinct from an LLM or chatbot), and what are the deployment, security, and operational considerations for using agents (including access/privilege management, human accountability, trust boundaries, and full‑stack observability)? add

How should organizations and security operations centers evolve their defenses to protect against AI-enabled, agentic adversaries and other AI-driven threats? add

How should agent identity and permissioning be handled differently from human identity and permissioning? add

bolt Powered by CUBE AI

Jeetu Patel, Cisco

search

Dave Vellante

>> We're back at RSAC 2026. You're watching theCUBE's live coverage. Jeetu Patel is here. He is Cisco's president and chief product officer. Great to see you, my friend. Thanks so much for making time.

Jeetu Patel

>> It's great to see you.

Dave Vellante

>> It's just awesome having you. We've had you at this show several times. We just saw you in Barcelona. .

Jeetu Patel

>> I don't think I've been to this show without being on your show. There's a couple of these shows where it's like my memory of RSA is sitting with Dave and talking to him about what's going on.

Dave Vellante

>> Well, we expanded the set too, give you a little bit more elbow room instead of a little round table. So thank you for making some time. We just met at Barcelona talking about all the innovations that you guys are doing, the edge. It's really exciting. I mean, the cloud is just expanding. It's no longer a place up in some remote areas. I mean, it really is amazing. You have said that we're moving from copilots to autonomous agents. It's ironic at this event, the theme is power community, but all we're talking about is machines talking to machines. What does that all mean? What fundamentally changes in your view?

Jeetu Patel

>> I think there's a profound shift happening right now. And if you really think of what we are seeing, it's not just the ... Agents had a ChatGPT-like moment with OpenClaw. And so what you've started to see happen now is as you move from chatbots intelligently answering your questions to these agents that are going to be able to conduct jobs and tasks for you autonomously, you can't think of these agents as software tools that are going to enhance your productivity. You have to think of them as digital coworkers that are going to be augmented to your team and they're going to be team members. And the problem, Dave, with these agents is they're kind of like teenagers. So they're superbly and supremely intelligent, but they have no fear of consequences. They do stupid stuff all the time.

Dave Vellante

>> They lack wisdom.

Jeetu Patel

>> They lack wisdom, and so ... And I have a 15-year-old and there are days when she feels like she's 45 and then there are days when I feel like she's five. And there's a range.

Dave Vellante

>> And you learn a lot from her, no doubt.

Jeetu Patel

>> .

Dave Vellante

>> A tremendous amount. Impressive, right? But then ... Hang on.

Jeetu Patel

>> On a serious note though, what happens is that what we are going to be doing is we're going to augment these agents into our teams and we're going to need to delegate work to them. Now the difference between delegating work and securely delegating work in a trusted way, trusted delegation, that the difference between delegation and trusted delegation is equivalent to the difference between bankruptcy and market leadership. It is literally that stark. So this is a very important conference because where we are at this juncture is we're at this point where you have to fundamentally rethink the architecture because the scale proportions and the scale volumes that these agents are going to behave at is going to be very different. For every human, let's say that you have between 10 to 100 on the lower end and 100 to 1,000 on the higher end of agents within a workplace, and so that's close to a trillion agents. Now you think about it that they're going to be working 7 by 24, which means it's three times a human capacity, so that's equivalent to three trillion other humans worth of work that got added on. And now think that they're probably at least twice as productive as us, that's six trillion.

Dave Vellante

>> At least. Yeah.

Jeetu Patel

>> Right? These are all conservative numbers when you start thinking about it. Think about the infrastructure that's going to be required to go out and support them. Think about the level of trust that needs to be put around them so that they don't actually go out and do things that could be in harm's way. Because in the chatbot era, what happened when they did something wrong? The risk of something going wrong in a chatbot is that you got the wrong answer. The risk of something going wrong in agents is they've taken the wrong action that potentially could be irreversible damage. Like I've taken out millions of dollars from your bank account and transferred it to Jeetu. It would be great for me, but bad for you. And so those are the kinds of things that we need to make sure that we actually start to think about. And I think security and safety is going to be kind of really important. And one last point I'll say is, in security, we used to think about zero trust for humans as largely access control. With agents, you're going to have to think about zero trust as action control, which means every step of the way, you will have to make sure that you inspect every step. And if there's anything that's going wrong, you have to be able to have the ability to intercept it.

Dave Vellante

>> Continuous ... I think you've made this point before. I think I might have heard this, you talked ... We might have talked about Barcelona or maybe in your AI summit, but you've got to have-

Jeetu Patel

>> We're spending too much time together.

Dave Vellante

>> Never enough. You've got to continuously monitor these agents and monitor their actions, you said you showed a chart last year at Cisco Live, and you showed the difference between a request and retrieve chatbot and the network traffic versus what happens with reasoning.

Jeetu Patel

>> I'm grateful that you remembered.

Dave Vellante

>> Oh, I paid close attention. I said, "This is the most important chart of the session." And it was just amazing because you saw this little there-

Jeetu Patel

>> ....

Dave Vellante

>> and then all of a sudden just this wave of color. Now we're entering the Agentic era, and Jensen made this point last week. The most important chart he'd showed, I think was his Pareto. He showed the three ... And it's interesting because you guys are riding that curve. One of the most important things you did, I think, in terms of just getting ... First of all, you got the Cisco team really focused, but then you developed a deep partnership with NVIDIA. And if you don't have a deep partnership with NVIDIA, what's the point here? They're driving the innovation and the performance, and you guys play a critical role in that value chain. But he said that you've got to figure out where you are in that Pareto. And to your earlier point, on that horizontal axis, we're talking about throughput versus, let's call it customer experience or responsiveness, really latency. If you're on the left side of the chart on that horizontal axis, it's free, freemium. If ChatGPT makes a mistake, no big deal. On the right side, this is your new revenue model. It's a monetization model. So what does that mean ... And the other thing he said was, "If I hire a half a million dollar engineer, a software engineer-

Jeetu Patel

>> If they don't use a quarter million dollars of tokens-

Dave Vellante

>> I'm going to be really mad." They're a $5,000 token, no, they have to have a token budget and they have to use it. What does that mean generally, but specifically for security, for the business model, how the world rethinks tech?

Jeetu Patel

>> Let me tell you my perspective on this, because I feel like the ability for an agent to have autonomous execution for long durations of time is going to be superbly important. It used to be when deep research came out, which was like the first agent that you and I had seen in its very crude rudimentary form, go out and conduct some research and give me back a report by compiling things from a bunch of different areas and give it to me in a neat form, but with the level of depth, right? It would take about 15, 20 minutes for deep research to come back to you in the early days to give you back a good 20-page document or something. And now, when you look at this, these agents are going to have ... The duration for autonomous execution is going to be not 20 minutes, not two hours, maybe two weeks, over time, two months, maybe two quarters. Imagine the kind of work that'll be able to be done when you tell an agent, "You and a basket of other agents, go solve this problem and come back to me when you think you have something meaningful." One, think about the change and the shift in the workforce that's going to happen. We're going to have to think entirely differently. We are no longer going to be doers. We are going to be managers of agents and we are going to delegate tasks and jobs to these agents and we have to do it in a trusted way and then they're going to go out and do it and we're going to come back and the bottleneck will almost become our consumption rate of what they have done. And so the next layer of review will also need to happen with the agents. But the second thing is, imagine the level of token consumption, imagine the level of infrastructure required, it's going to be of epic proportion.

Dave Vellante

>> Highly capable worker bee agents and we've seen-

Jeetu Patel

>> Conscious worker bees. They might be strategic thinkers, they might be scientists over time.

Dave Vellante

>> Yeah, and you're going to build that up because you're going to need-

Jeetu Patel

>> They're going to keep building that up all the time.

Dave Vellante

>> Because we've seen glimpses of this, Jeetu, I know, personally, my backlog of really interesting projects is growing exponentially because of AI and I'm the bottleneck again.

Jeetu Patel

>> The bottleneck, yeah.

Dave Vellante

>> So I need, like you say, more than just ... I got plenty of worker bee action going on now, I need more senior and strategic agents. I got to pyramid myself out.

Jeetu Patel

>> But this is a very important point, which is people are worried about the jobs getting lost to AI. I think there's a couple of things that are going to happen. Humans will be divided into two major buckets or categories. There's going to be highly AI fluent people, dexterous in AI, the ability to go out and really understand what's going on. And then there are going to be people that are not AI fluent. I think it'll be, you will have a massive shortage of AI fluency in the workforce with humans and you'll have a massive surplus of people that are not AI fluent and they will experience a fair amount of displacement. And so I think what we have to do as a society is get on this task of making sure that we can make more and more people fluent in AI. And if we can do that, you'll actually see high rates of development, progress, employability, all of that. But if you don't, then you'll actually see a fair amount of pain and suffering.

Dave Vellante

>> I'm actually quite impressed with the number of people that I talk to in our industry that are actually using OpenClaw.

Jeetu Patel

>> Yes.

Dave Vellante

>> And then of course, many people, most people outside of the industry haven't even heard of it. And to your point, there's going to be a huge asymmetry-

Jeetu Patel

>> There's a huge asymmetry....

Dave Vellante

>> between the haves and the have-nots. So how do we close that gap?

Jeetu Patel

>> Well, I think the first thing you have to do is these technologies have to get simpler because OpenClaw, while it's easy, it's still not as easy for my late mother. If she were alive right now, she wouldn't be able to just go out and start using OpenClaw. And if she did, it'd be a disaster because she would give it all kinds of personal information that she shouldn't be giving. And so there'd be no kind of semblance or guardrail. So we have to make sure that we get an education cycle going for everyone to get to a base level of understanding. And once you get to that base level of understanding, I think you have to make sure that there's constant level of encouragement from companies to say, "You have to use these agents in order to make sure that you get your job done." Because the throughput capacity of those two groups of people I talked about, the AI fluent group is not going to be 20% more productive than the non-AI fluent group.

Dave Vellante

>> No, much, much higher.

Jeetu Patel

>> It is going to be 50X more productive. It'll be hard to justify. There's no analogy in the past that I can even equate to it because imagine getting a knowledge worker today that doesn't know how to use a smartphone, doesn't know how to access the internet, doesn't know how to use Microsoft Excel and wants a job, has never done email in their lives. So I think those are the kind of things that'll be huge challenges for us as society to make sure that we get eight billion people to a base level of dexterity in AI.

Dave Vellante

>> I want to come back to your point about simplification. You know Jensen's five layer cake, it's really infrastructure heavy. You're in infrastructure. I've been in infrastructure my whole life, but I want to come at it from top down. Will the LLMs, which are advancing so rapidly ... I like to think of the layer cake as starting with the intelligence layer, let's call it the LLMs, and then there's this cognitive surface that is emerging that's not really fully baked yet, which harmonizes data, call it semantics, it's forming. Look at Palantir, what they're doing. Okay, but that's within their little world. You need a forward deployed engineer to use Palantir. You're talking about the everyday person. So something has to evolve there to make it simpler. Will the LLMs be at least a big, of course there'll be a big part of that answer, but are they going to start to really innovate in that simplicity zone and who will fill the other gaps? What does that look like?

Jeetu Patel

>> I think it's not ... The simplicity doesn't come from the LLMs. It comes from the agents that use the LLMs and use the models. So I think the way that you have to think about it is what is the deployment simplicity for getting an agent activated and how well can they be trained and coached to actually do the entirety of the job? What is the classical definition of an agent? Someone who has access to your tools, right? That is the one difference between a chatbot and an agent.

Dave Vellante

>> And has the authority to act on your behalf.

Jeetu Patel

>> It has the authority to act on your behalf, has access to tools and systems, and can go out and conduct a job almost autonomously, if not fully autonomously, and come back to you if at all there's something needed, but otherwise they'll just go out and make judgment calls. A few things to keep in mind. It has to make the right judgment calls. We have to make sure that we have trusted guardrails. We can go out and have an ... There has to be a human accountable for an agent's behavior. So the infrastructure and the apparatus around security will completely change because if you think about the way that ... Like I said earlier, if you wanted to have zero trust, which is least privileged access for agents, it can't be only about access control, it has to be about action control. What does that mean? That means that you have to give the agent the right level of privileges just in time. It has to be just enough privilege, not over privileged, and it has to be for just enough time, and then you have to revoke it. Get the task done, now I'm going to take away your permission. That is a whole different architecture than what was needed classically. The other thing is the endpoint is no longer something that you can fully trust because you can download anything to the endpoint. Where trust gets established is where there's connectivity that's occurring. Where does connectivity happen, on the network. And so you have to make sure that that connectivity is happening on the network, and as that happens, you're monitoring these agents with round the clock observability through the full stack from, how are the tokens working and the GPU to get manufactured? How's the model performing? How's the app performing? How's the agent performing? That entire stack has to be observable all the time. And every new app that gets built has to be born observable.

Dave Vellante

>> So I feel like the first line of defense five, seven years ago, even four or five years ago was the cloud. What's the first line of defense now? Is it agents? Is there no first line of defense? Obviously you have to have layered defenses. Is it the network?

Jeetu Patel

>> The granularity is going to be at the task level, the action that you're taking as an agent. That's going to be the area that you have to protect, and every single action has to be observable, interceptable, and applicable to dynamic guardrails that can be provided. So because here's an example. Just because you gave an agent permission to do something doesn't mean that it should do it at bulk scale. So for example, if I give an agent permission to say, "You have access rights to delete my email." Right? I don't want it to delete 10,000 emails in a second. It's okay for it to delete one email that they might have processed, but I don't want it to delete 10,000 emails. So when it's starting to do something like that, knowing that something is not going right and being able to intercept, you're going to need to have an agent at agent scale that actually does that because you cannot do it at human scale. So if there's an agent that's starting to do things that are going sideways, you need to have another agent to be able to intercept at human scale. And the transitive property that we have actually so long relied on for guardrails with humans does not work with agents, which means if you come to me for something and I'm the CFO and you're an employee in the company and you say, "Hey, can you give me what the earnings look like before they come out?" You as a CFO say, "You know what? I can't do that. Even though you have access to talk to me, I can't do that." And an agent, if you have access to talk to that agent and say, "Give me this," the agent might give it that. So you have to then be able to intercept and say, "No, I can't give that to another agent." So there's a whole lot of complexity here that needs to be solved for, which I don't think we have actually fully . And that's precisely the reason why we announced Defense Claw and Defense Claw was actually piggybacking on the announcement that Jensen made at GTC-

Dave Vellante

>> NemoClaw.

Jeetu Patel

>> But it was more around OpenShell. So what Jensen announced was OpenShell.

Dave Vellante

>> Oh, yes. Right, right.

Jeetu Patel

>> What is OpenShell? OpenShell is a secure container that can run OpenClaw. What we have done is we have actually taken a bunch of services, skill scanners, MCP scanners, AI bill of materials for inventorying AI agents, model provenance, all of these services, we've open sourced them. And then what we did was we put it all under a single umbrella called Defense Claw and that can connect with a hook into OpenShell. What's the benefit of that? Every single time an agent runs, an OpenClaw agent runs in OpenShell, you automatically have all of our security services activated against that agent. You don't have to worry about it. And that's the beauty of it. And then you can expect us to have in our platform a Defense Claw enterprise edition that can provide the management capability for Defense Claw, the open source edition.

Dave Vellante

>> So this is, if I'm in a room with you with a whiteboard, I mean, obviously Cisco's got a phenomenal business, networking, collaboration, security, et cetera, but this is a whole new vector of opportunity for you.

Jeetu Patel

>> Agent security is a multi hundreds of billions of dollar market. And by the way, we have no concept right now on how grave the consequences can be if you don't deploy this. And the beauty about this is Cisco is at the forefront of this. Historically, Dave, you've given me a hard time in the past saying, "Hey, Cisco had stopped innovating." Five years ago, I remember when you met me, when you and I saw each other, you were like, "Jeetu, what are you doing? Cisco has stopped innovating." No one accuses us of that anymore. Right?

Dave Vellante

>> True. You're right.

Jeetu Patel

>> We are by far and away ahead in the market. We were the only ones over here in seven days from the time that OpenShell came out, there was DefenseClaw for OpenShell, seven days. 100% of AI defense is written by AI. Half a dozen products by the end of the year, '26 will be written with AI, no human lines of code. And by the end of next year, 70% of all Cisco products will be written with AI. And the implication of that is there's no such thing as a legacy company anymore. You will either have a modernized stack or you'll be out of business and we are refactoring our entire code base.

Dave Vellante

>> And DefenseClaw is you said open source.

Jeetu Patel

>> DefenseClaw completely open source.

Dave Vellante

>> And so-

Jeetu Patel

>> Available on GitHub today.

Dave Vellante

>> Available on GitHub, when did you start writing code? When did AI start writing code?

Jeetu Patel

>> So we started writing modules of DefenseClaw, like Skill Scanner and all of those about two and a half months ago or so. The day we launched it, we had 1,200 GitHub stars, which by the way, in open source is massive. We bundled all of these pieces together. DefenseClaw was actually written in seven days. And so I just want people to process the fact that ... And by the way, it's not like, "Oh, this is a concept kind of thing." No, this is a full scaffolding for OpenShell for an OpenClaw agent working with security guardrails in place in seven days.

Dave Vellante

>> Amazing speed. I wasn't able to see your keynote, I was doing this.

Jeetu Patel

>> .

Dave Vellante

>> But I talked to some people who saw it, but there's these unknown unknowns, and you sort of alluded to that before, and I think you talked about unmanaged risk as well. And then there's of course known unknowns, but it's those unknown unknowns that really we have to be careful about, isn't it?

Jeetu Patel

>> And by the way, on that front, what needs to happen is we have a whole kind of life cycle of services that we provide now. So we can look at every model and know how a model is performing against algorithmic red teaming attacks that we have actually done and given a model of score, which is a leaderboard. We have an explorer edition of AI Defense now, which allows you to, on a free of charge basis, explore AI defense with your own data and your own models. And then what we've got is all of these other AI defense capabilities we've put in place. But the most important thing right now that we have to think about is this notion of an agentic SOC. If agents are going to be used by adversaries to go out and create attacks on society, the only way you can combat this is at machine scale, not at human scale, which means that every SOC will be agentic. If you don't have an agentic SOC, you don't have a SOC, that's number one. Number two, once you have an agentic SOC, every alert and every signal will be evaluated. In the past, SOCs used to go out and prioritize because a SOC analyst didn't have time to go out and look at all the alerts and all the vulnerabilities and all the signals. Every alert and every signal now will be evaluated. So detection and response will happen at machine scale and speed.

Dave Vellante

>> I think this idea of moving from access control to action-

Jeetu Patel

>> Action control....

Dave Vellante

>> control, is how you said it. And I think that's profound. And so you're extending essentially identity to agents.

Jeetu Patel

>> Yes. And by the way, agent identity needs to be thought of slightly differently from just human identity.

Dave Vellante

>> How so? Explain that.

Jeetu Patel

>> Because when you're thinking about an agent identity, the permissioning that you do against that identity has to be far more temporal than a human identity and has to be far more granular. And so like I said, it's just in time permissioning, just enough permissioning and for just enough time permissioning. Just in time means as soon as the agent wants to do something, you have to give it permission, you can't be delayed. Just enough permissioning because you only give it permission for the scope that they're going to do something and no more. Least privileged access. Just enough time is the moment they've been done with the task, take away the permission. That's a completely different way to go out and think about the action control for an agent identity versus a human identity where I can give you something for the term of your employment, unless you have otherwise proven that you're not deserving of it. But in agents, you have to assume it the other way around. They're not deserving of it for more than the time that you need to give it to them.

Dave Vellante

>> Right. Well, Jeetu, you've awakened a sleeping giant. They say you can't move a steamship overnight, but you're actually doing that. And so congratulations on that. Your AI summit was the best event of the year so far. I think it topped even GTC, which was wonderful, but it was-

Jeetu Patel

>> Oh wow, that's big praise.

Dave Vellante

>> But the reason I say that is GTC was awesome, but it's all NVIDIA all the time, whereas you had a much wider sort of observation space across the industry, and it was just a real pleasure to see that .

Jeetu Patel

>> We wanted to make sure that we brought the ecosystem together to have the important conversations without just peddling products. And I will tell you that the conversations you do in these midform, it's hard to do these in five minute broadcast interviews. So I appreciate you taking a little bit of the time so that you can understand the details and congrats on all the success that you've had. Can I just give one last plug to my team?

Dave Vellante

>> Yeah, please do.

Jeetu Patel

>> DJ Sampath, who runs our AI team and people in his team, Akshay, Arjun, all of these guys that have been working tirelessly round the clock, Matt Caulfield who runs identity and others, I think we need to make sure that the level of credit that they deserve for what they're building right now, it's nothing short of extraordinary what they're doing. I get all the credit oftentimes from people like you, but literally it's those folks that deserve the credit.

Dave Vellante

>> Glad you mentioned that because again, I think you've awakened Cisco's engineering culture and so congratulations on that.

Jeetu Patel

>> We're having fun.

Dave Vellante

>> You're a builder and you inspire other builders in your organization. So thanks so much for taking some time with us. Really appreciate it.

Jeetu Patel

>> Thank you, my friend.

Dave Vellante

>> All right. Thank you for watching. This is Dave Vellante. We're right back, RSAC 2026. You're watching theCUBE.