Jazz CEO on Rebuilding DLP for AI

Hey guys, thanks for the throw. We're here with Ido Livne, who's the CEO of Jas, and we're having a little chat about, well, data loss prevention. It's not quite living up to what we need in the age of AI.

Ido, welcome to the show. Thank you for having me, Mike. Great to be here.

So on the plus side, more people are paying attention to DLP than any time in recent memory, and a lot of that has to do with they're concerned about what data might be leaking in through these AI tools that everybody seems to be using, a lot of which is unsanctioned, shall we say. But we're discovering that yesterday's DLP technology may not be up to the task. So mark us through what's going on here.

Yeah, 100%. And again, thank you for having me. So, the story of DLP, and everybody that's touched DLP knows this intimately, it has been a story of failure.

Okay? And that's not by accident. It was fundamentally built the wrong way.

So DLP is data loss prevention, just to make everybody be on the same page. Data loss prevention, the way it was set up is basically, it's a dumb framework that was designed for a different era, right? It was never equipped to understand the context of modern work, how we work today, the data that we're operating with, and the system we're operating with.

It was not built to understand that. And in that sense, it's failing. And when we're talking to a lot of security teams, and we find them to be stuck in one of two states, okay?

Two, or sometimes we name three, but basically, we have the paralyzed. There's a lot of people out there, a lot of companies that are out there, that are completely avoiding DLP altogether. And they do that not because they're not understanding the risk, but rather because they're smart and they're avoiding the solution.

So in a sense, the pain of having a DLP solution is bigger than the pain of having the problem, right? Because the solutions that exist out there today just don't solve for the problem. If they don't have to do that, they're not compelled to, for legal reasons, for compliance reasons, they'll avoid that altogether.

The subcategory there is that, we call them the burnt or the failed, right? The people that have tried and failed and were not able to operationalize a DLP program. We see those, a lot of them as well.

And we obviously see the last category, which are the trapped. Those that have a legacy program, and they know it's, in a sense, sort of a compliance theater more so than anything else. They still have a lot of data loss.

They have a lot of coverage gaps. And that thing does not protect their company's data. Yeah.

So I'll pause here for a second for more questions. So walk us through what is it that is the obstacles that people are encountering that's leading them to conclude that DLP might be one of those things where the cure is worse than the disease? Yeah.

So there are basically three ways in which a DLP today fails, okay? So the first one is that it's noisy, right? Every DLP program out there that has actually started and taken on this task of setting up the program knows that, right?

9% of them are false positives, and team get bogged down with alert fatigue, need to close out the false positives. This becomes busy work more so than anything else. Again, if you're actually trying to implement that program and walk through closing out all those false positives, you have sometimes 10, sometimes hundreds of analysts doing that work, documenting everything for compliance reasons.

But basically, it's very rare that they actually find the thing, and they just get fatigued and even bad things that do happen just go across. So that's the first method of breakage, right? The second one is that legacy DLP is blind, right?

So it doesn't see all the new modern ways in which people are working with data. So, GenAI prompts and complex clipboard activities, screenshotting. There's a lot of different ways in which data is being processed today.

And not only that, the modern landscape of IT tooling mandates also that there's a lot of tools that have a corporate version of them and a personal version of them. Obviously, uploading sensitive information to your personal Google Drive, very different story than uploading it to the corporate version. So not being able to distinguish between the two is really hard.

Also, remote work is changing that environment as well, creating a lot of blind spots within that, and a lot of personal use of company machines, et cetera. So this is a big blind spot. And the third piece, and that's probably the biggest point here, is that those legacy DLP tooling is basically, they're dumb tools, right?

They don't understand context. They're based on a rule-based system that is very technical, right? You put in the patterns, you put in the regexes, you put in the rules, and every time there's a match, you say, "Let's send that match to a human that will actually understand that," right?

So they're completely lacking of deep understanding of the business operation, how the business operates, and who's who, and why things are happening. That's completely not part of what the tooling is supposed to do with legacy tooling, which is obviously a big gap. That's part of what the humans are supposed to do in the DLP program.

And the human part just never scales. So how do we kind of address that in the age of AI then, so that we don't have a situation where we've got a bunch of, for lack of a better way of describing it, dumb tools that lack any context that are just kind of getting in the way? Yeah.

So, I'll start by saying that basically AI is both a challenge and the solution, right? But it makes the problem with DLP just bigger, right? Like all these gen AI tools, obviously a lot of agents that have popped up in the recent months and already like a year at this point in time, are increasing the challenge because they offer employees new ways in which they do work, right?

To accelerate their work, getting to high productivity, using these tools, it's really hard to say no to using these tools after you started using them. And we all know that. It's really easy to put company data at risk while doing so, right?

Either by sharing data yourself with those tooling or having those tools work for you and put that data at risk. So that's a big part of the challenge increasing and why you can't wait today in finally implementing a DLP program that is good, right? So that's the challenge side.

But it also offers a solution, right? So, the main insight that we had in Jazz when we started out was that, as I kind of like alluded to before, if every DLP program has the machine side and the human side, and the machine side is where you put all the patterns and registers and rules, and basically every match goes to a human to investigate, understand, and take action on. And the human side just never scales to the level that is required by the noisy DLP physics here, right?

Like data is moving all the time in the enterprise, and getting everything under review is just impossible for humans, right? So our insight was, let's build the human, right, into the machine. And obviously, what enables that is AI.

Like we're able to mimic human-level reasoning, nuanced reasoning, in-context reasoning, that allows the machine basically to understand why things are happening and not just what. And we've embodied that by implementing basically what we call Melody. Melody's our investigator, DLP investigator, that deeply understands, not only how data is moved, but also why things are happening in context of the business.

So, AI is what is enabling that. And there have been a lot of other next gen DLP providers that kind of like started using AI, but we believe that they've been mostly sprinkling a little bit of AI on top of the same old rule-based framework at its core. And the rules are the problem.

What you want to have is basically a nuanced, a reasoning engine that understands how the organization is using data, why things are happening, being able to make those decisions in real time to actually protect data in context. So how does the platform know what's the right context? Because theoretically, I shouldn't probably be using the consumer-grade version of OpenAI, but maybe there's a reasonable excuse for doing that.

So how does it kind of understand what's legitimate versus what may be something malicious? Yeah, that's a great point. So, the way this needs to work, right, again, from first, if you want to mimic how a human will reason, because like if you would have a human that understands security review what you're doing with data, they would know if this is okay or not.

And the question is, how do we encapsulate that and re-implement that perspective into the machine or the system, as I kind of call it? And the answer is the same way that we mimic humans in reasoning with other AI tooling and that you personally experience in your personal life, or professional life. Like basically it's a natural language policy engine, right?

So, sorry about that. It's a natural language policy engine where, basically Melody, our investigator, reviews data flows at scale and in depth. It understands the data that's moving, the systems in between which the data is moving, the people that are engaged with that, the first party actor, who's sending the data?

Who are they sending that to? The second party, third parties being related to the activity, and also the business processes of the data movement is actually generating it. Through that, Melody basically creates that full story of not only what has happened, but also why things are happening and the intent of the actor.

And only then it reasons through, is this thing that I just observed, that is happening, is that okay or not? And the way it reasons is with a natural language policy engine that basically mimics how a human would think about this. So it's a very nuanced, sort of blob of text of what is considered to be sanctioned activity or unsanctioned activity for this specific organization.

So, and the way we've set it up is that it's really easy to set up the initial policy based on existing policy documents, but then there's also a built-in feedback loop in the product. As Melody surfaces unsanctioned activity, you're basically able to engage with that and learn with it what's okay and what's not okay, what's actually happening in the business, and basically get to a very precise definition of what is the actual policy for your organization. Mm-hmm.

I get that. Is that going to change the relationship between the security people and the folks they're trying to protect? Because I think a lot of the conversation over the years at least, has been People like the concept of DLP, but the means is somewhat, shall we say, abrasive in some ways, and it kind of makes the security people into some kind of police force for the organization rather than somebody who's there to protect people and help people.

So will that conversation change as we get more of this context? Because I think a lot of the times, there's this tension between the security folks and the end users they're trying to protect. Right.

And that reflects itself in many different ways in the day-to-day security work, right? " Or how things are configured and sometimes feel like they're saying no to a lot of things that a business wants to. And that doesn't need to be the case, right?

The goal of every program set up correctly is to enable the business to succeed and not be a deterrent, not be a detractor or a slowing factor, right? And in a sense, that is also what a good DLP program is set out to do. So like, and you can look at the big banks and massive healthcare organizations, like when they implement a very strict DLP program with legacy tooling.

Basically, that's why there's really a bad rap to DLP programs, because there's blanket blocks and prevention capabilities are being put in place that basically make the employees feel like everything is... It's hard to do work, right? It's stopping business conducive activity.

So in a sense, having somebody that deeply understands what's going on and is only blocking actual risky activity, that is a big shift and that's a big change to actually an enabler to the business. That's one point of that. And second point is, the way we've set up the solution, the way I think that a modern DLP program needs to work in, is by also helping the employees understand when they're putting the company at risk.

Because a lot of times, and basically business side employees, they're not thinking about security day in and day out. What they care about is doing their job, right? And basically, a solution that will help them understand how they can do their job in a way that's more resilient towards the business, like they would want the company to succeed.

They won't want to put the company at risk. So something that is an advisor and a guide to how to better do this business workflow, without impeding the results. And by also securing the data while doing so and making the data more resilient for the company, that is something that the entire workforce would strive to accomplish.

So the question is, how do you balance things and how do you present things to the end employee to make everybody work together towards achieving the business goals, but also keeping the data safe? Of course, there's a new type of end user out there, or at least maybe there are extensions of the existing end user, depending on what their use case is. But AI agents don't necessarily have a whole lot of insight into context, and from their perspective, they are just trying to accomplish the mission by any means necessary.

So how do we kind of apply DLP to an AI agent that is by design going to do everything and anything it deems feasible to accomplish a mission? Yeah, and that's a great question. " And then there's a lot of lacking guardrails.

Like it will do crazy things that will not be stopped. And not only that, we believe that when you look at all the ways in which agents could be run by employees, the endpoint or the machine they work on is actually, that's the Wild West. That's where employees are using all different types of platforms for agent use.

All the agents that are running on the machine are running with the employee's credentials for everything, right? And that's a lot of access to a lot of different systems, a lot of different data. And that's where most of the risk is in that case, right?

And the question is, how do you rein that in? How do you make sure that you're allowing your employees to accelerate the workflows and business activity, without putting the company at risk? And the answer to that is, that you need the right solution to put the guardrails on those agent activities.

Obviously, we'd love to demo what we have there. But the goal is to make decisions that are in context of that specific business, understanding why things are happening and what is the human trying to achieve, what is the agent trying to achieve in allowing activity that is not destructive, but stopping an activity that is. Mm-hmm.

So as we go forward here, are we getting a new appreciation for data security that maybe we've ignored for the better part of three decades or so in the age of AI? And I know there's more risk at the moment, but ultimately might this prove to be a good thing because we're finally going to address some issues that we've kind of let slide for way too long? 100%.

This is why this is exciting, right? AI allows us the opportunity to solve very old problems. And as I said at the beginning, many companies are avoiding data protection altogether because in the CISOs, the chiefs information security officers across the board, they have past experiences of working with other companies that maybe they led a DLP program, maybe they were engaged with that, maybe they saw that from the side.

DLP Ah, does not work today. So there's many experiences where people have been burnt by DLP. But this is different.

And when you apply a Jazz, and when you apply AI in the right way, I will say, not maybe just Jazz, then basically that gives the opportunity to help organizations finally protect their data in a way that was not possible in the past, and do so easily, which is critical to make it work for everybody. So what do you see security teams doing today as it applies to DLP that just makes you shake your head a little bit? And frankly, I'm not even sure it's the security people who are in charge of DLP.

Sometimes it's just the IT folks. But are there things that you wish that we would just do better or be a little smarter about? Yeah.

I feel like people are either avoiding progressing their DLP, or taking the same old DLP and, as I said previously, sprinkling a little bit of AI on top of the same old solution, using AI to better classify the data, or using AI to automatically set the rules, or using AI to automate what analysts would do with the alerts that come in from the legacy solutions. So, we call that garbage in, garbage out. You're not getting a lot of context, and you have an analyst making the decision without the relevant context, and now you're automating that across 1,000 additional alerts.

That's not the right approach. The right approach is deeply understanding the data flow, why it's happening, what is the intent, what is the user trying to achieve, and through that, making the right decision on the fly, and making the right decision on detecting and responding to that situation, versus just automating a decision that was made with lack of context. So, yeah.

All right. Well, folks, you heard it here. Whatever biases you have about DLP, maybe you might want to put those aside in the age of AI because, well, the game is changing, and like it or not, the bad guys are after your data.

Hey, Ido, thanks for being on the show. Thank you so much. Thank you, Mike.

All right. And back to you guys in the studio.