Networking and Cybersecurity Implications of AI Agents with Anirban Sengupta

Hello, and welcome to the latest edition of the Text Arm AI video series. I'm your host, Mike Zora. Today we're with Anand San Gupta, who CTO for Aviatrix.

And we're talking about the overlooked implications of networking and security challenges that will occur as we all deploy thousands, hundreds of thousands, maybe even millions of AI agents one of these days. Hey, Anand, welcome to the show. Thank you.

Thank you for having me. So describe for us what it is that is at the core issue here when it comes to networking and security. Because all these AI agents, as far as I can tell, they're just really gonna be additional endpoints on the network, and they're all gonna start generating traffic.

And well, what could go wrong? Well, um, uh, that's a great way to put it. What can go wrong?

Uh, so, uh, before I, uh, point out what can go wrong, what I would like to, uh, kind of cover for the audience is what these agent or agents for the AI really do, right? Uh, that might, uh, help, uh, align the needs and the shortcomings and how, uh, the, how we are trying to fix those issues. So, uh, there are four key steps that the AI agent really does.

Uh, so the first piece is, uh, what, uh, we call perception and data collection. So what these thousands, and sometimes as you said, maybe millions of or million of, uh, these agents do is what is called perception and data collection, which is gathering and collecting data from multiple places. And now this data can exist virtually anywhere.

It can be in clouds, it can be on-prem, it can be on edge, it can be in QSRs, right? And it can be anywhere across the world. It's not just in particular location or geo, but it can be across the world.

So that's the first step. The second step is about decision making. Once these agents have collected all this data, right, they use these models, right?

And tune models to make the decisions. That's the next step. Number three step is about action, right?

And execution. So once they have made that decision, then these agents, right, act on that decision. And last but not the least, it's that it's learning from all the data it has gathered and from all the, uh, decisions it has made, right?

It tweaks and adds to that learning, right? And that step is basically called learning and adoption adaptation. Uh, so now if we really look at it, let's try going from the first, second, third, fourth.

So for perception and data collection, most of the data, right, has to be collected from all these different places, right? Which can be actually not just where like thousands, as you said, it can be millions of places, right? And to collect that data, what do you really need?

You need like high bandwidth network, which needs to be intrinsically secure, right? Because a lot of this data is sensitive. It's PII data, right?

It can be also high value data, even if it is not PII, right? So all these data has to be accumulated and gathered, right? On a very secure way, right?

And also it is very high, uh, bandwidth. So it has to be high bandwidth security. And that's what we also help out, right?

With, uh, aviatrix, we provide very high speed encryption right from end to end. And, uh, it's completely secure. The second part, which is there for agents, is decision making.

In this case, what happens is that once the data is gathered, right, it talks to the models, right? And figures out what it needs to do from the decision perspective, right? And this is where right, our, uh, the whole firewall service comes in, right?

And make sure that, um, the right decision is, uh, the right agents get to make the right decisions with access to the right models. The third piece, which is action, is very interesting. Now, once you have, uh, thousands or maybe millions of agents, right?

What decision those agents are making needs to be identified, right? What agent is making what decisions and how they're communicating with each other. And this is what I call traceability, right?

Or else, like, these are some things. Sometimes I say these are robots, fighting robots that you need to figure out that who did what and why, and things like that. So that is where observability really is important.

And, uh, it's not just network observability, but figuring out which endpoints, right? Are talking to which endpoints and when, so that there is a traceability or they can be tracked right? When these actions are being taken or else it can basically get into a very chaotic situation where you don't know who is doing what and it's not traceable or trackable.

Uh, that's what we don't want to happen, right? And, uh, the last, but not the least is about, um, about the learning and adaptation. So one thing to think about is that this, almost all companies, they spend millions, if not hundreds of millions and few companies.

They even spend billions on tuning this model and learning and adapting these models. But, uh, if a bad actor gets hold of that, right? And exfiltrate that model, the things which you have spend millions, hundreds of millions of dollars, right?

Can be theirs in minutes, right? So this is what is really crucial. And in all these four areas, right?

Aviatrix really helps out to provide a solution. To your point, it seems like we're a little obsessed these days with protecting the data, but that's only one part of the equation. And ultimately, the bad guys, maybe they just wanna steal the agent and the model and everything they went with it.

'cause that's the intellectual property that matters A hundred percent, right? So, and also if you look at the attack surface, there are three, four areas which I talk about why, why it has changed the whole landscape. One is, the first is that the attack surface has become way bigger.

It's not just the on-prem data center is not just the cloud, it's edge, thousands, millions of endpoints that can be there. So the attack surface is higher. And the next part is that, uh, with gene AI and the money, which is pouring in, and many of them are nation state funded, right?

I maybe you already know about the salt type phone use case, right? Mm-hmm. So in these cases, the bad actors or the infiltrators right, has become really, really sophisticated.

So your attack surface has grown, your bad actors have become way more sophisticated, and you have third pieces a lot to lose. Mm-hmm. Right?

So all of these three things has created the perfect storm to bring in a solution like aviatrix, which provides built-in security and connectivity. You don't have to think differently that you have a connectivity, then you put security on top, and maybe they're not working together. It's in silos.

Your configuration can be wrong. All of that is taken away. Unfortunately, in my experience, every time we have some new advance in technology, the cybersecurity aspects of it wind up being an afterthought until something bad happens.

So how long will it be before that something bad happens? Uh, really great question. So, uh, this is how I put together.

So the thing is that, that enterprise, all enterprise, they have a huge pressure to bring in all this cool technology, right? So because of the pressure, right? It is really needed to increase the velocity to bring those applications.

It's not, it's from the market, it's from the c uh, CEO, it's from the developers. Everybody wants things to be fast, right? But from the other side, there is need for security and connectivity and compliance and audits, all that thing, right?

The controls. Now, uh, the part which is really interesting is that how to balance velocity with security and compliance. Now, the thing is what your question about when that bad thing happens, actually the bad things are happening, right?

And, and like think from today, right? Just because of salt typhoon, nobody really knows what is compromised. So in security, while I've been working on networking and security for more than two decades, so we have a meme, uh, we say that there are two types of companies, one, which knows that they have been compromised and one who doesn't.

So the meme part of it is that it's not like it's a big event, which happens, I'm sure, like we all get these letters once in a while, which says, Hey, your PIA data, your social security number, your credit cards have been compromised. And these companies will say, oh, for one year you'll get a credit report. Personally, I have got like at least five of them in last one year.

These are all cases of breaches, right? It is exfiltration right now. A lot of it has happened already in grand scale, right?

I'm sure you guys know about large amount of breaches that has happened in the United States itself, that it is happening worldwide, right? So it is not that it, we are waiting for such a case. It is happening every day more than once, right?

So the, the part is that, uh, I can't talk about certain customers, but there is a customer whose network and really, really importance, uh, important, uh, production in healthcare that went down for a week, right? And they had to really put something immediately, and they came to us, right? And we had a really good solution.

We were easily deployable and things like that. But it is happening every day. We, it's not like we are waiting for something bad to happen.

Do I need to go find somebody who's a cybersecurity expert who happens to also know about how AI works? Or am I trying to train the AI people about what it means to be more secure? I mean, who ultimately is gonna take responsibility for this?

I mean, I know who's gonna get blamed, but I just wanna know who's responsible. Yeah. So, um, this is how, uh, I explain to my customers the thing is that it is about the technology and about the processes at the end of the day, right?

Uh, now if you really look at it right, over time, what has happened, and even like if you look at on-prem, there were networking companies like Cisco's and, and there were security company leads like Palo Alto network, checkpoint and et cetera, right? And it was always, security was a afterthought or a bolt on technology. So the cri critical pattern, it, it used to work pretty well.

When you have a building and you know exactly where your egress point is, where your ingress point is, you have a big door in front and you put a big lock on the big door, right? But with today's architecture, with cloud and edge and on-prem, right? What I call it as a disappearing perimeter.

So data can go out and in from any place, and you don't even know that place exists. So what we really need from the technology point of view is a inbuilt secure network fabric, right? Every communication is secure by default, sometimes also called a zero trust, right?

So if you have, or if companies have that technology, that's the baseline, right? And on top of that sits the processes, right? So all the processes that is needed, like for example, today, if you look at the DevOps pipeline with Kubernetes and containers, right?

It's a process and it cre makes sure that there is reputability, there is audit, there is compliance, there is drift management, all of that, right? So what my, what how I tell customers is that if you have the right technology and if you have the right processes, then everybody does not have to think of it. You want your developers who are like cool gi uh, developers not to have to think every day about security because it is already taken care of by the technology and the processes.

Mm-hmm. And same way, right? The, the platform team does not have to think about it every day because they have set the processes in place so that it is already taken care of.

They just have to make sure that it is properly audited, right? Which the system is going to tell you, right? And it should just work without anybody's, um, manual intervention.

So are we gonna have to fund A Major upgrade cycle of our infrastructure to get to the faster networking that you described? And of course the servers and GPUs and everything that goes with that. But are people overlooking the networking side of that equation?

Uh, yeah. So, uh, that's a great question. So actually you have the technology right?

There is fast connectivity. Whether you look at CSPs or the um, or, or the service providers, there is fast network, which is already there, right? The part which needs to be done is what I call it as secure by layering.

So all it needs is to put a software defined networking and security layer on top of the base connectivity so that the end-to-end across all clouds, across all edge and cores and on-prem, right? That's software layer, right? A distributed software layer, right?

Takes care of it as a layered approach. We don't have to upgrade, right? The networks or upgrade the infrastructure.

We need to just put a distributed networking and software solution right on top of your connectivity today with a single or unified pane of configuration and visibility and observability, right? And that can be easily layered in. So for example, the uh, solution which aviatrix has, right?

You don't have to change or you don't have to upgrade everything. It just goes as a layer on top of your current deployment. Not just that you have to change everything.

You can roll it in in phases, right? And transparently add that layer on top the other parties that we just released, something called a pass solution avia, strict PAs, which is that we manage everything. So your operational cost actually goes down, right?

And you get this whole service integrated security, right? Which is completely managed by Will you be adding your own AI agents to your service? Who in turn will have to talk to all my AI agents?

And how is that all gonna get orchestrated? 'cause it seems like there's gonna be all these, uh, control playing and data playing conversations that need to take place between different classes of agents. Is that right?

Yes. Actually, uh, we already have certain agent, uh, agents in the product itself, right? So we have started rolling them out.

It, uh, we have agents which figures out, um, how bad the scenario is for every VPCs and their deployments, right? And on the other side, we are also working with, um, the CSPs to integrate with their agents to make sure that they can, we can provide the information they need from end-to-end. Um, we need to be way more collaborative in this space, right?

So that we all work together to achieve the outcome, which we are looking at. But it has already started. And, uh, we are working with, you know, like top three CSPs in that area.

So what is your best advice to folks about how to have this conversation with their teams? 'cause I think every organization you talk to is AI happy. There's a lot of maybe occasional or irrational exuberance, but I don't think they're really understanding all the, the fundamentals that you need to get addressed.

So how do I kind of get in there and kind of bring everybody down in reality? Yeah, great question. So what I generally advise customers and prospects is, number one is to embrace multi-cloud infrastructure.

So to design with multi-cloud and the infrastructure in mind. So have a infrastructure like, which is basically networking, security, and the full stack, which is across clouds, across edge, and it is primed, right? For supporting any kind of gene AI applications on top, invest in it, right?

Invest and embrace, right? Uh, full multi-cloud edge kind of, uh, infrastructure. Uh, number two is that security needs to be embedded in the infrastructure, not a afterthought and not a bolt-on, right?

When the cloud architecture is being designed, it should be embedded into that, not what I call is, is built in security versus bolt on security. And last but not the least, is provide the processes so that the developers and these gene AI experts, they can develop and deploy their applications with no friction from security and infrastructure team. I think maybe we're overlooking one small, but important point is, aren't most of these AI agents, and maybe even eventually the models themselves gonna be running at the network edge, because I need these things to be closer to the point where the data is being consumed and created, rather than having some sort of round trip to a cloud somewhere where I may have trained the initial model, but now I need to get the inference engine from that out to the edge.

So, um, does that require a level of, of finesse that we're not thinking through? Yeah, so, uh, it is a hundred percent correct, right? So the thing is that, uh, the base models are generated generally in the cloud because it needs lots and lots of, uh, GPUs and infrastructure and data to be generated.

But, uh, these agents basically tweak those models, right? And these models go out into the edges and all the different, uh, places, and they get tweaked, and then that, uh, tweaked data finally comes back and then makes those base models more, um, enriched, right? So, uh, that's the main reason why I ask, uh, enterprise customers to provide a fully distributed, uh, secure network across clouds and across edge, right?

Because once you have that, you can run these models anywhere. And, and you don't have to think, like, for example, if a developer wants to deploy it in cloud, they deploy it in cloud. If they want to deploy it in edge, they deploy it in edge.

Um, I had in my past, um, company QSRs, one of the largest QSRs, right? And they used to de they would deploy all these, um, agents in their small QSR with two little boxes, right? But if the infrastructure is robust, high performing and secure, you don't have, like, the developers don't have to think where to deploy, Right?

Folks, you're heard here, I think it was back in the eighties the first time I heard the phrase that the, the sy the network is the system. Well flash forward all these years later in ai, and the network is still the system. Hey Riman, thanks for being on the show.

Thank You very much. Thanks for having me. All Right.

ai video series. You can find this episode and others on our website. We'd invite you to check them all out.

Till then, We'll see you next time.