Umesh Mahajan, Broadcom | VMware Explore 2026

[music] Hello, I'm John Furry, your host of the Cube. This is the Cub's VMware Explore coverage. Of course, we got two days of live cube coverage from the show floor. I'm here with the broadcast executives to unpack the innovation at news and news at the show. Now, here up here in the cube now is Zumesh. John is VP and general manager of the application networking and security division. And we're going to talk about the agentic AI era and its impact on the attack surface areas in the enterprise. We're seeing new threats from autonomous infiltration of LLMs, agents, data stores, MCP services as the new infrastructure is being deployed. These are huge opportunities to to nail down. You got to nail these challenges or agents cannot scale. Great to see you again. Thanks for coming on the cube. >> Thank you, John. >> So, you know, we've had this conversation every kind of cycle. The perimeter is gone. What's the surface area? LLMs bring in a lot of opportunity and intelligence in the enterprise. Obviously, the domain expertise and the data is unlocking. >> What is the attack surface area look like now with AI being injected into the equation? >> So, it's very interesting time, right? Uh news gets old very quickly in these days. Like last year, it was autonomous attacks. Oh my god, anybody can become a genius and start attacking rapidly. Then mythos showed up and we saw okay, it's very good at finding vulnerabilities. has got everybody worried. How are we going to protect against that? And lately the agent AI stuff gets out of your environment, starts hacking other customers and goes all over the place. So it's really a scary time for our customers. Anything everything can go and anybody can become an expert and unknowingly also things can happen with these agents. So I think this is the time where you have you can't put off security any longer. Oh, I got a parameter firewall. I'm good. No, no, no. not good. It can be bypassed. So now the security gurus or experts are saying you know you can't take two years, three years. You have to deploy lateral security. You have to think all the stages of security and properly tackle them because you have to put all the prevention in place because these attacks are going to come it's going to be too late. So you better put segmentation in place. You better start thinking of more than even the lateral firewalling. You start thinking about uh advanced threat protection, zero day tax. What are you going to do? And you'll have to patch your environment. Both things are customers. The infrastructure you buy from vendors like Broadcom, but their own applications and these days is big harming. They just could be patching their entire lives, right? Because there so many vulnerabilities are being found. So we also suggesting to our customers we can help you with virtual patching we can give you signatures while you're waiting to do your own application patching we can patch for you. So lot of exciting stuff but lot of the same security tools you can enhance properly you can provide the visibility and you can guide them to deploy much faster and that's the environment which is is the security is moving to and so our customers have to select the vendors properly. Yeah, I mean your point about um the the attacks that are happening, it's not hypothetical, they're real. We're seeing play out and the velocity uh is is a huge deal. Speed of attacks um but now you have autonomy as a factor. Um you mentioned lateral security. Explain the strategy of how people are thinking about this because it's happening. People are taking a more cautious agentic approach. You're seeing more deterministic workloads end to end and they're well scoped. So it's not you know truly autonomous but that's that's the where the entry point is in the enterprise how what needs to be in place for agent to agent as as these proliferate in the infrastructure >> you almost you almost need to think of uh zero trust and zero day properly and not just peace meal so there in lies the challenge right uh everybody knows you have to do zero trust and zero day today it's a must you can't I'll just put a little >> it's not optional you not an option is it has to >> it's not an option but our customers have bought multiple security products they can't put it together it's like buying Swiss cheese yeah you have pieces of security because you have plenty of holes which people can drive through so you kind of need somehow an integrated stack and more a software stack where all the elements are talking to each other and that's exactly what we have with red defend and not only that you need to give them the full visibility and you need to walk them through prescriptive kind of thing that's what we did with DFW1234 last year we have released ATP1234 so it's in a very rigorous stepwise fashion made it very simple they can deploy they can look at things and then they can go about you know securing the environment and then with our IDS IPs we have lot of signatures available we're coming up we have a signature pipeline which we are enhancing with agentic AI of course you use AI to further advance Yeah, mythos could be a double-edged sword. One hand is good and other hand's bad. You >> But we we have to use it. We have to use it to make our security better, right? Otherwise, we won't be able to fight. >> All right. So, take us through just detail out. I want to get nail you down on this because I think people want to know what what's in the box as they say. What's on by default? Sounds like you guys have zero trust as an always on >> configuration. Yes. >> What else does the customer need to do to maintain that control or manage the security? Um uh >> so I I think you know they they need to make sure that they turn on the options right some options we will obviously take them through but as they get to as they get more comfortable with it you know we have very high scale right so we have very high scale because in AI environments two things matter of course what is your security stack but uh the throughput matters so we have very very high throughput we are doing like 75 terab per uh v center cluster uh for firewalling we are doing 17 terabit for IDS IPS and the other aspect is also latency because in AI workloads latency matters so our security is done at the hypervisor level much like AS6 so we make sure the packet doesn't go left right you're not copying it 15 times so we are delivering all that at the same time but the customers need to be able to you know say okay I'm okay you're gonna turn this on by default at least they need to agree to that. If they agree to that, I think we can secure their environments in a very nice fashion and as they get hit with new vulnerabilities >> or CVEes, we can help them with IDS IPS signature pipeline so that they can do virtual patching quickly because we can intelligently push that from a central location in a distributed fashion and we won't push all the signatures everywhere a little bit more intelligent. That's where AI comes in again. we'll use the IIO what is running in the host we'll only push the >> those signatures into that host and if they upgrade the software we'll get rid of the old signatures so all that kind of thing is inbuilt into infrastructure so it scales >> and at the same time intelligent they are not bewildered oh my god I have to learn 50,000 things how will I ever do it >> well this is where I think AI is interesting because you mentioned segmentation at the top of this interview and you know when you look at things like oh we got a vulnerability how do we quarantine it the the process of doing security now can be applied with agents. So a lot of these configurations and management you can actually you know segment something pretty quickly. Yes. >> So talk about that impact because I think this is a new reality in the security world is that >> there's a lot of that blocking and tackling you're doing on a whether you're a red team or whatever you're doing you can actually move faster. What what specifically is AI doing better now than that humans had to do before and what's the role of the human? So I so I think the human is still there to do some audits and checks and bounds. What if uh you know I'm using agents in my solution like security services platform to provide the visibility and quickly do the operational stuff like you just click a button and boom the rules will go in because I'll calculate them rapidly or those changes. So so I think there at some point there should be checks and bounds a little bit because you want to make sure like the stuff didn't go crazy. Other than that I think the humans just need to decide the top level policies right what are the guard rails what's going to happen and leave the busy work if I can call it to AI and the machinery over there and that's exactly what we are doing with agents in our security services platform we want we want to take all the busy nittygritty work out where humans a can and make mistakes and then they take too long take too long >> we're seeing here at explore a lot of ecosystem partners are engaging a lot of activity people have good visibility on what the AI infrastructure is going to look like at least from a a mechanic standpoint AI factories etc coming out but the number one question I get and I'd love to ask this question so if I'm your friend and I say tell me the impact of agentic AI we're seeing LLMs we're seeing MCP servers all these new tools it's a great environment for AI development um agents are here what's different what what should I be aware of with Agentic infrastructure and agentic applications. What's the number top three things I need to look at and pay attention to? >> So, first of all, I need to be able to discover them, right? MCP, who's talking to who? What are the tools? I need to be able to figure that out that picture because in traditional IT everything was written down someplace that doesn't work in this environment. These will come and go very fast. So, you need to be able to discover them in real time. And then the same things which are used for traditional work uh machines you know virtual machines and Kubernetes clusters you want the same kind of policies to be enhanced to work with MCP and all these other uh tools and agents which are talking to each other. So and you as you discover you'll be able to tell these are real this is shadow IT what the hell is going on and then then your administrators can say I never authorize all this what the hell is all this running so I can >> quarantine this and not let it talk to the rest of the infrastructure. So that stuff is very easy for do for us to do with a few tweaks to our existing enhancement we can figure all this out and show it to the end customer. This is this is what you allowed. This is what this shadow somebody without your knowledge is cranking out all these things and doing something. >> Yeah. Yeah. A is running wild. It's kind of like a TV show. Um these days uh modern apps are running on Kubernetes. We've seen that in all the reference architectures on the big AI AI scaled systems because you got to run control planes. You got to run the resources. >> Um APIs are very dynamic. Um that in cloud native is all API based and microservices. How do you guys protect the API side of it? Okay. So we we have this other product called the AI load balancer. Now the and it's it's a wonderful load balancer but it's specifically very strong in the Kubernetes in the ingress side right uh because in the ingress side you have to load balance and to manage it and it's an integrated load balancer layer 4 layer 7 gslb waf so customers love it. Uh we have lot of penetration with kubernetes and the VMware kubernetes vks. So already WFT is there as a security element and it's becoming popular. W VA waf seven eight years ago was not that hot but in the last two three years with all these attacks coming in >> and all these attacks which come in at the web server level it's it's getting really hot. So we have that product over there. So we thought when you're talking Kubernetes and agent take a agent agent AI workloads are all Kubernetes they're all talking API. we have this wonderful platform which customers are deploying in large number why don't we take the w and add the API protection on top again principle crack open the packet looking inside the packet why not look at the APIs too so we are looking at the API and providing the AP we have releasing we have released that code and where that is your w becomes w so we have bot management API production and web application firewall all in one tool And it is going to be able to provide API visibility, API vulnerability detection and how do you you know stop what is not allowed at an API level because a lot of people come in the through an API they can make you you know execute some code at the end of an API which can cause a lot of trouble. So it'll stop all that and it's sitting right there with the Kubernetes workloads. So it is the right place >> and obvious sits in Kubernetes. Where is that? What is that? So it basically it's it sits at the ingress level. >> English. Okay. Okay. Okay. How does that evolve? Because I was talking with a friend from the cloudnative compute foundation CNCF coupon. We've been every been to every show since its inception in North America and Europe. And we were riffing on the idea that you know Kubernetes and microervices um were a big thing. And of course we that's cloud native. He said agents are a lot like microservices. You got to look at it. Mhm. >> Take us through the evolution of Avi and this this piece because >> it's almost the same movie you're seeing because they're out running around and they're standing up turn terminating their relationships but they're doing reasoning and work. They could be coding. I mean agents not just like a search paradigm. They're they're multi-step process. They could discover something write some code on the fly. >> Yes, >> that's a cool thing but also it could be dangerous. >> It could be very dangerous. So, so you know again the thing the good thing is we don't have to start from zero in any place right we talked about kubernetes and uh workloads running over there but all these agent AI workloads are also kubernetes workloads so same environment we have the auv load balancer ingress load balancer doing vap now we thought about it hey this this whole notion of AI gateway right AI gateway kind of does three things it discovers all these agents uh and who are they talking to etc and it also provides visibility and logging and so that you can you know track what happened over time and it does the usual every AI gateway does like route to the cheapest model or open source model okay that's the easy part but the third part is the tougher one how do you provide security to these agents we've been in customer conversations without naming them at our EBC's >> and when the CISO shows up they keep asking not explain to me how exactly my agents are talking, how you going to protect me and all that. Yeah, identity is one thing but beyond identity, how are you going to do the blocking and tackling so that they don't go rogue and they >> what is the blocking and tackling? What does that entail? >> So, so I think some of it is the same principle, right? First you have to be able to see what's happening then you have to pro you know lateral segmentation etc that you you wouldn't have reached hugging face if you had lateral segmentation in place. Right. So you're saying hugging face wouldn't have happened, the open AI thing wouldn't have happened. >> Yeah. If you if you had if you had done the lateral segmentation at least would have protected the propagation from happening. But going to our AI gateway with which again we are taking the AI tool and we enhancing it's already a pseudo gateway with the API protection and all that. What we want to add to there is also the Xfiltration right your PII information your credentials your credit card numbers they don't go out. So we want you know there's something notion of oh vast 10 in waf so we going to for agentic stuff there's another incarnation of that so we're going to take the most important stuff of that enhance it because that's what people are worried about right enterprises like I wrote some agents to do some payroll stuff and low all my information social security numbers are out there no so you want to be able to protect that you don't want to allow it you just don't want to uh you know rely on some other tool here there somebody forget get to connect it you don't want to do that so security for us is a huge part of uh the AI gateway and we'll be doing that and then with v defend you have all the other parts right your lateral propagation zero trust zero day so that together I think we can provide a very comprehensive and complete security solution >> well umesh this the AI story here at explore is strong obviously it's part of it it brings intelligence into the enterprise a lot of unlock block. Um, final question. As as a senior leader, uh, and in security, you're seeing a lot of change. It's funny, the networking is moving up the stack and software is moving down the stack. So, you're starting to see security become a real critical AI infrastructure component. It's always been there in other paradigms. >> Um, and you have the traditional security, zero trust, but it's coming down into the core, >> you know, around the density of the systems. uh you mentioned you know virtualization layer is a big part of it you have the the ingress piece with a what's the biggest change with AI that that's changed in the security world is it the notion that it's coming down closer to the resource and embedded from day one does it change the the uh design and deployment of infrastructure and software for these AI systems so >> I think it's sum or all of the above in some ways right Yes, it has to be at the infrastructure level. It has to be at scale. Otherwise, when are you going to do it? 2 years from now, by the time you'll be compromised, you know, may not have your job or if the right stuff gets stolen, you'll be in big trouble. So, I think the realization >> right from the top, the CISO, CIO, head of infrastructure level is now suddenly we go into a meeting, all of them are there in the meeting because they are worried, right? So they are worried, they want to listen, they want to learn and they want to make sure they are picking the right architecture, the right solution, will it work at scale? If they are picking VCF, are they picking the right security offering and the right load balancing offering or not? And as they thinking of AI to run on Kubernetes, VKS, do they have correct solution? So I think all those conversations that they're happening, they are realizing they have to deploy it at the same time. Can we provide them the tools to do that now? Yes, we can. As soon as they understand that, I think some of the customers are moving quickly forward. >> It's a very interesting time on the computer industry because you have horizontal scale and vertical integration, but it's not like a just the the workload. It's happening everywhere. You can have an end to end. It's got to know everything end to end is essentially vertically integrated, but you got to have horizontal scale. Yes. >> And vertical or domain specific end to end >> at scale. In the old enterprise, it was great. You just build a stack, put some hardware there, there's the workload go. Yes. Yes. And with hardware prices going up, I think scale and throughput, not just scale, throughput has become very important and that's been an area of focus for us, right? We want to provide twice and thrice the amount of throughput. Server prices have just gone through the roof. Uh and the traffic is increasing. The poor customers, they need to buy more servers. Traffic is increasing. So if we can provide them a scalable solution and not just scale out scale out as you put it on most servers but also scale up per core I want to give them more scale >> then it becomes very interesting for them and we are really giving them value with our product >> large scale open distributed computing and the edges we even get to the edge next year I think will be a big edge conversation that'll that'll play in um you're busy you know application and security is a hot area >> you feel really interesting times I I think uh I'm excited about this whole domain. I'm a you know vendor so it's good for me that there's so much happening in this area and there's so much demand >> uh for our product >> and the role of networking is elevated to up again up the stack security's coming in network is the key >> network is the key with AI traffic is very heavy right and and it has to be secured >> yeah I mean some of these racks you have you know north south east west traffic going all over the place you have you know this density around these systems but that's just conventional networking meets new AI networking. It's a whole another ball game. >> Yes, >> thank you for taking the time to unpack all these issues and news at Explore. Appreciate it. >> Thank you so much, John. >> All right, I'm John Furry with the Cube. We're here for our VMware Explore coverage, of course, two days of live coverage. Check out the Cube at the cube.net. Thanks for watching.