How Broadcom And VMware Protect Dynamic AI Workloads

Agentic AI is showing up everywhere, and it is not waiting for your security program to catch up. From the VMware Explore floor, we talk with Umesh Mahajan, General Manager for Broadcom’s Application Networking and Security Division, about what changes when AI agents become real workloads that talk constantly, call tools through APIs, and evolve by the week. The headline is uncomfortable but useful: AI helps defenders ship faster, but Frontier AI also helps attackers find vulnerabilities and drive more zero-day behavior than many teams are ready to handle.

We dig into the security model that holds up when agents are dynamic and mistakes travel fast. Umesh explains why you have to solve a dual problem: protecting agents from attacks while also preventing a compromised or misdirected agent from attacking others. That leads to concrete priorities like lateral segmentation, distributed firewalling, IDS/IPS, and NDR-style behavioral detection, plus network-level controls to reduce the risk of sensitive data exfiltration. We also talk about discovery, because shadow AI is already a reality. If you cannot identify which agents, MCP tools, servers, and models are running, you cannot confidently separate authorized usage from risky sprawl.

Finally, we tackle the fear every operator has: security that slows the business down. We cover why performance and latency matter more for AI workloads, how enforcement close to the hypervisor can reduce overhead, and what it takes to deploy protection quickly so “bought but not deployed” does not become your weak spot. If you’re building enterprise AI on VMware infrastructure, this is a practical map for getting safer without stalling momentum.

Subscribe for more VMware Explore conversations, share this with the person owning AI rollout, and leave a review with your biggest question about securing agentic AI.

Transcript

Copy Transcript Copied

SPEAKER_00 0:01

And

Welcome From VMware Explore

SPEAKER_00 0:02

it's Evan here, and I'm back for another VMware Explorer. It's been a blockbuster opening at the event. And I'm really excited to chat with Umesh from Broadcom. Umesh, how are you? I'm doing great, and thank you for having me here. Well, thank you. Congratulations on another incredible year at VMware. Before

Meet Broadcom App Security Leader

SPEAKER_00 0:21

we dive in, maybe introduce yourself, your team within Broadcom and the mission.

SPEAKER_01 0:26

I'm Umesh Mahajan. I'm the general manager for Application Networking and Security Division at Broadcom. And our primary products which we manage are the VDefend security, full security stack, and the RV load balancer.

SPEAKER_00 0:41

And it sounds so simple, but you've been busy the past year. So much capability today, so much more on the roadmap. Maybe talk a little bit about what's driving that, particularly with all the excitement around Agentic. So much more security is needed and you've delivered. So tell us what's new.

Frontier AI And New Vulnerabilities

SPEAKER_01 0:59

AI has been fantastic for us. Of course, we use AI in our products, including coding and doing advanced security and load balancing. But I think primarily Frontier AI found vulnerabilities. I mean, they are finding tons and tons at a scale much higher than normal tools, and then the agent tech AI, right? Security for this dynamic kind of environment is very important. So in all these environments, new attacks are coming in at a higher level of security where you need to deploy advanced threat protection much faster and much sooner than prior times. Prior time, okay, you do some kind of firewall, you're good enough. No longer, that's no longer good enough because with these attacks, you can you know penetrate hosts and workloads very quickly. And then zero-day attacks are coming in. AI is also good at zero-day attacks, which you've never seen before. So how do you deal with them? And when it comes to agente AI workloads, first you're they're dynamic in nature, they're all built-in Kubernetes, they all talk to APIs, you know, API level security. And here you have a dual kind of problem. First, you have to protect these agents, agent, and then they go rogue and they start attacking. So you have to make sure you don't let them attack anybody. So there's a dual prompt security problem. The third aspect is you you have to be very sure that nothing is getting key or sensitive information is getting exfiltrated. So you have to stop that. And so you have to have some lightweight DLP at a network level, network security level in place. Otherwise, your credentials will go outside the organization.

SPEAKER_00 2:41

Yeah, it's it's it's a fast-moving space.

Finding Shadow AI And Unauthorized Agents

SPEAKER_00 2:43

And how do enterprises even know what's running in their environments and how do they become aware of what's authorized, what tools and agents are running, and get a get a snapshot of where they are?

SPEAKER_01 2:54

So with BDFend, we've infused some AI intelligence into it. So now it can discover the agents, it can discover the MCP tools, MCP servers, and the LLM models. So it can show, hey, this is what we are seeing, we've discovered, this is authorized, and all the stuff is not authorized. These guys never took permission for you. So there's a whole shadow AI kind of stuff people have spun up. So then the IT folks can easily block that, right? Not allow it to go anywhere. Because these users or lines of business, they just spun up things without proper permission, and it's there. And then with our we defend one, two, three, four, what is authorized, we can protect it day one in a comprehensive fashion, right? With our full lateral security at a distributed firewall, IDS IPS, NTNDR for zero-day attack. So we have that tooling in place, easy tooling, you can deploy it very quickly, phase one, two, three, four in that manner, and protect it from day one.

SPEAKER_00 3:56

So

Rogue Agents And Lateral Containment

SPEAKER_00 3:57

as these agents get more sophisticated, there are new use cases emerging. What are the kind of threats that you are seeing in real life? I I see media reports every day about an agent escaping or taking control of my clawed account and and uh stealing credits. That was the one from this morning. But what are some real scenarios that you're seeing in the field?

SPEAKER_01 4:16

Aaron Powell So we saw an OpenAI did a study, case study on how that OpenAI case was like three, four weeks ago, it went rogue and then it compromised Hugging Face, and it was in action for a few days, and nobody even knew. Because I think what happens is these these LLM models are very powerful, and you give an intent, but you know, the tool understands slightly differently, or they are true to the whatever he said, and they don't reason this could be wrong or this could be this. They are true to whatever he said, and then they take it to heart and then they start doing all kinds of funny things, and that's where it went rogue. It was looking to do something extra because you had said do this, that. And uh, and if we had lateral segmentation in place, the initial compromise of you know the two the LLM could have gone crazy with the agent tech AI workloads, but it wouldn't have spread all the way to Hugging Face and gone there. So I think what we are saying, security is security and the concepts we have, right? You have a pyramid firewall, but more importantly, I think lateral security is way more important. You have to implement it day one, you have to deploy it. That's what I mean. And also you have to deploy the advanced threat protection security, and then you have to be able to see all these open AI, not open AI, agentic interfaces and solutions so that you know who's talking to what, what is happening, what is shadow AI. And all this has to be in your pipeline, the security intrinsically, when you spin up these agents. It's not afterwards I will come and deploy it. I think that paradigm, if you do, the anything can happen. Either these agents can be attacked from the outside by autonomous attacks, or these agents can go rogue and start doing a lot of damage. So, as like in Broadcom, as we are building pipelines to deploy agents right there and then we will deploy security, otherwise you are at a very high risk. And you have to have the right architecture also for security. Has

Protecting Prompts APIs And Data

SPEAKER_01 6:23

to be software defined, and all the elements have to work. Example, RV also we are enhancing. We talked about VTFN, but there also we are enhancing, right? In the RV side, we are looking both sides because RV has the WAF and now the WAP API protection. So it can do prompt inspection, it can do other levels of security detection, and it can prevent the special items, your special crown jewels from getting exfiltrated, like PII and credit card information. So it can act as a security element in both directions, and then also it can look for any anomalous behavior for your agents, LLM models, et cetera, and prevent zero-day attacks on them. Because with Frontier AI models, it's not only a regular attacks, but zero-day attacks increasing. And those are highly dangerous because you haven't seen them before and you don't know. And the typical security models won't work unless you're doing something like NDR, NT, NTR to determine through behavioral detection and protect against that.

SPEAKER_00 7:31

Fantastic. It's amazing to think we weren't even talking about agentic AI a year ago here. Agents weren't even on the radar. Uh maybe in the lab, perhaps. But it's incredible how fast things are moving. Of course,

High Performance Security Without Latency

SPEAKER_00 7:43

at VMware and and Broadcom, you guys think performance all the time. How do you think about security with not slowing down your operations, your network, your employees, et cetera?

SPEAKER_01 7:54

So I think very good question. And you know, and I can also add uh server prices are going up, so everybody wants more for their dollar. So scale is important, especially for AI workloads. They are very heavily throughput intensive because they are very chit-chatty. A lot of traffic goes back and forth. And even the AI A6, there's a lot of networking component part to it, right? So same here. So we have enhanced the performance of DFW by almost 240% now in our latest releases, and we scale up to 75 terabits per second for a VCF instance, which has a thousand hosts. IDPS, we went from like 8 terabytes per second to 17 terabytes per second. We've doubled the performance. And we are doing the same thing on RV side, we're doubling the performance. But there's another component to throughput and all for AI, which is latency. And customers have started asking us, okay, you'll give me all the security, but are you gonna add a lot of latency? No, is the answer because we run at the hypervisor level. So we are we are running at the lowest level where the packets are cracked open, so it works like an ASIC pipeline.

SPEAKER_00 9:06

Wow.

SPEAKER_01 9:07

So we add the lowest amount of latency when we provide security. Unlike third-party security vendors, which we call bolt on security, they add quite a bit of latency when they apply security.

SPEAKER_00 9:19

Oh wow.

Fast Deployment With Security Services Platform

SPEAKER_00 9:20

So talk about scaling up. How do I deploy your solution with VMware infrastructure that I may have in place, or maybe next to it, other infrastructure? How quickly can I get your solutions up to market?

SPEAKER_01 9:32

So what we've done is we have the security services platform, and we have found that some customers can deploy a solution where other customers are slow on it because they don't have the proper staff to automate everything at a rapid fashion, right? They just don't have the skill set. So they do little parts of it and then they tend to move slowly. So we invested heavily in this tool. So last year we rolled out we defend one, two, three, four, for distribute firewall or lateral segmentation. And this year, a couple of weeks back, we added the threat defense or ATP because we feel proactively you have to deploy the entire stack and not just part of it because you're exposed at the upper layers and attacks are moving to that. So I think with this rollout, we are finding and at the CTAB, which was here on Saturday, Sunday, some of our customers told us that, hey, I was able to deploy 4,000 VMs security using the SSP 1234 in a matter of weeks. So that's how we are operationalizing security at a very rapid pace, because even if people bought a product and never deployed it, they're not secure until you really, really deploy the product.

SPEAKER_00 10:47

That's exciting. Congratulations on that. So, you know, so much great content here at Explore around security, really important stuff.

Where Agentic Security Goes Next

SPEAKER_00 10:55

If if without committing or talking about roadmap specifics, if we were to fast forward to next year, this time, where would you see the industry being? I think AgenTic will be rolling out at scale. What else do you do you hope or envision for the industry and for VMware by this time next year?

SPEAKER_01 11:10

I think you know we have really good features. My our biggest focus will be we have the proper roadmap and the mind share as this as this area evolves rapidly, because you know, who knows? They could be agentic AI too. And so you need to be ready for that. Because anything in this area is like four months old, slides are obsolete, the way things are moving these days. One is that. Second is we get the customers what we've delivered as innovations deployed on this feature. We get them at least safe to this level, right? And then we can talk about new and more advanced things showing up. We'll continue on that journey. So getting the customers embrace our architecture, embrace our technology, and get it deployed and safe is our primary objective in the coming year. And then we'll grow from there. We will not stay still. We'll c we have ideas and we'll continue investing in a very forward looking roadmap and security area.

SPEAKER_00 12:11

Well, as always, congratulations and uh onwards and upwards. Thank you so much. Thank you.