How Broadcom And VMware Protect Dynamic AI Workloads

Evan Kirstel (0:02)

And it's Evan here, and I'm back for another VMware Explore. It's been a blockbuster opening at the event, and I'm really excited to chat with Umesh from Broadcom. Umesh, how are you?

Umesh Mahajan (0:13)

I'm doing great, and thank you for having me here.

Evan Kirstel (0:15)

Well, thank you. Congratulations on another incredible year at VMware. Before we dive in, maybe introduce yourself, your team within Broadcom and the mission.

Umesh Mahajan (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.

Evan Kirstel (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. Tell us what's new.

Umesh Mahajan (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're finding tons and tons at a scale much higher than normal tools. And then the agentic 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 agent AI workloads, first, they're dynamic in nature, they're all built in Kubernetes, they all talk to APIs, in API level security. And here you have a dual kind of problem. First, you have to protect these agents, and then they go rogue and they start attacking. You have to make sure you don't let them attack anybody. So there's a dual prom security problem. The third aspect is, 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.

Evan Kirstel (2:41)

Yeah, it's a fast moving space. 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 snapshot of where they are?

Umesh Mahajan (2:54)

So with VDefend, 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 this stuff is not authorized. These guys never took permission from 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, the lines of business, they just spun up things without proper permission, and it's there. Then with our VDefend 1, 2, 3, 4, what is authorized, we can protect day one in a comprehensive fashion, right? With our full lateral security at our Distributed Firewall, IDS, IPS, NT, NDR for zero-day attacks. So we have that tooling in place, easy tooling, you can deploy it very quickly, phase 1, 2, 3, 4 in that manner and protect it from day one.

Evan Kirstel (3:56)

So as these agents get more sophisticated, there are new use cases emerging.

What are the threats that you're seeing in real life? I see media reports every day about an agent escaping or taking control of my Claude account and stealing credits. That was the one from this morning. But what are some real scenarios that you're seeing in the field?

Umesh Mahajan (4:17)

So we saw an OpenAI did a study case study on how that OpenAI kept a case for like three, four weeks ago.

It went rogue and then it compromised hugging phase, and it was in action for a few days and nobody even knew.

Because I think what happens is these LLM models are very powerful, and you give an intent, but the tool understands slightly differently, or they are true to whatever you said, and they don't reason this could be wrong or this could be this. They are true to whatever you 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.