Episode Transcript
KB [00:00:10]:
What's up everyone? It's KB dialing in from the US for VeeamON Sydney at Ilumina. This is Veeam's global tour touching down for one day only, and the brand has come a long way from where it started. Veeam made its name on backup and recovery, the stuff you were quietly glad you had when something broke, really. Today, the room splits into 2 tracks, resilience on one side and security on the other, and both of them keep circling the same worry. By Veeam's own numbers, something like 90% of enterprise data is dark, unclassified, and ungoverned, and now AI agents are being handed the keys to it, moving faster than any team can audit. Which means the thing that you are backing up might already be the thing that's being quietly compromised, and you find out when you go to recover. Across the day, I'm sitting down with the people who run resilience and security for a living to work out what's a real shift and what's just another agentic conversation bolted onto last year's pitch. This is From the Field.
KB [00:01:20]:
Let's get into it. Joining me now is John Jester, CRO at Veeam, and today we're discussing why resilience is now a business survival conversation. So, John, thanks for joining me and welcome.
John Jester [00:01:36]:
It's great to be here, KB. Thanks for having me on the podcast.
KB [00:01:39]:
Okay, so John, you opened up the keynote today, so I want to get straight into it. What's one message you want every board and executive to really stop ignoring about resilience?
John Jester [00:01:51]:
So the key message we talked about is how do we enable safe AI in the enterprise? And let me break that term down. So I very much believe we're now in the agentic era. Companies have AI agents interacting with their data, making recommendations, triggering actions. It's incredibly important that we've got the right governance in place. And this is a board conversation now. I have so many customers reaching out to me. And their board is asking them questions on cyber resilience, or if you're enabling AI, how are you gonna protect the business? How are you gonna make sure that you've got continuity plans if something goes wrong? And we talked about this concept of AI drift as customers implement AI agents. And what drift means is an AI agent took an action that was unexpected.
John Jester [00:02:43]:
It was wrong. It made a mistake. And there's really 2 categories that that can happen in. The first category is what everyone talks about today. There's a cyberattack using AI agents or a rogue AI agent, or something as simply as an employee brought in AI tools, shadow AI, gave it permissions to datasets, and it caused unintended consequences. That's what we talk about all the time. But the second category is actually the one that customers bring to me all the time. And that is, hey, we had our own AI agent.
John Jester [00:03:15]:
We sponsored this, it was approved, but it took actions that corrupted data or made wrong recommendations. And that's 'cause they trained it on older obsolete datasets. They over-permissioned the AI agent, so it had access to data or the ability to write and edit data beyond the scope of what that AI agent was. So what we're doing is providing our customers with tools to implement safe AI. So how do you govern the AI agents operating in your environment? How do you understand all the data that they interact with? And if you track that properly, that enables something we call precision recovery. So I know exactly the actions an AI agent took, and I may have hundreds of AI agents operating. 99% of them are doing exactly what I want. I'm getting high business value from them.
John Jester [00:04:05]:
But that 1% that's corrupting data, making wrong decisions, precision recovery lets you go in, undo those actions, and put your business in the state you need it. And ultimately what that means is an enterprise customer can deploy AI, they can run pilots, and they can do it with confidence. They know they've got a recovery plan if something goes wrong while they can still run their projects and get the benefits of AI.
KB [00:04:33]:
Okay, this is interesting. So, talking about precision recovery, is it kind of like, you know, when we're back in the day, like, building websites, we could just do a quick rollback if something happened, or when we would deploy a feature, we could do a rollback because something had an issue? Is it kind of like that? Talk to me a little bit more, like, how quick is it? Is it a process? Is it convoluted? Is it complicated?
John Jester [00:04:53]:
I love your example. So, here's why your example fits. It's a rollback. And here's why it's so much more complex today. I've got the pace and speed that they're operating in. And then as I mentioned before, you have these AI agents that are maybe over-provisioned. They have permissions to data they don't need for their job. They have permission to edit, delete data.
John Jester [00:05:18]:
All that makes it incredibly important that you can track everything they do. So when I say precision recovery, It's exactly the rollback you talked about on a website, but it's doing it by monitoring every AI agent, every piece of data they read, they deleted, they edited, they created, and then the ability to just undo the actions that you don't want. That's precision recovery. If you don't have precision recovery, you might have to roll back your entire data estate. Well, guess what? That takes hours or days of time. Companies can't operate in a way where they've got to hit the undo button and recover their entire data estate. They just want to undo the actions that were unintended.
KB [00:06:00]:
Okay, so John, what was coming to my mind, going back to precision recovery, so it's sort of that it allows companies to give them like a little sandbox to do what they want to do to innovate, because now everyone's like, you've got to innovate, you've got to use AI, because you can't not use it because the downside is obviously there. But then if there's an issue, we can go back to how it was before. And would that sort of reduce a lot of that risk? Because I understand a lot of boards are nervous about that, whereas they didn't want to— they want to leverage it, but they're also hugely concerned about risk. But you're sort of saying this is like a halfway sort of mark to enable people to innovate safely with AI, but giving them that option to roll back effectively. Yeah.
John Jester [00:06:43]:
Yeah, that's exactly it. So what this does is it gives you confidence to pilot and to use AI in your enterprise because, you know, I've got the AI agent governance, I'm tracking exactly the data they're editing, writing, creating, and I've got precision recovery to undo it. And maybe that's the simplest way to say it. We can undo unintended consequences while keeping all the actions that added business value. That's what we're providing. So maybe that's a really nice way to summarize it. Undo unintended AI agent actions. When you have that ability, you now have confidence to deploy AI, to get the business benefits, to drive the innovation, but do it in a way where you can confidently sit in front of your board and say, hey, we're implementing AI, we're driving innovation, we're getting business benefit.
John Jester [00:07:37]:
But we're doing it in a way that's safe. We're doing it in a way that we've got the right governance, data security, identity controls, and we're protecting the business.
KB [00:07:48]:
Yeah, that's interesting because, I mean, look, boards just often ask like a very binary question around like, are we doing AI? And so, it's about, well, we are, but these are our parameters to do it, right? Because I think that that is something that is not just such an easy question to answer. So, are you sort of seeing that at the sort of top end of town, or how are these sort of AI conversations really unfolding at the board level whilst they want to be able to drive innovation? If they don't, obviously there's other consequences for that, as we know. But how are people balancing all of this, would you say?
John Jester [00:08:21]:
Yeah. So, I think 12 months ago, the conversation was either a CEO or a board basically putting pressure on the business. Hey, are we taking advantage of AI? Are we getting left behind? What are we doing? And we saw that occur and we saw these pilots spin up. And that's where, frankly, we saw a lot of these unintended actions happen with AI agents. And that's where the conversations with the CIO, the CISO in companies really started to mature. And they were feeling the pressure maybe from the board or the CEO or different business owners saying, we want to implement more AI, we want to pilot more things, but they needed that confidence. Hey, how can we do it? With the right governance of AI agents? How can we really understand exactly the actions they're taking on our data? How can we secure that properly? And then the undo, as we've simplified it to, that ability to undo with precision, that gives them so much more confidence to say yes. You know, when the CRO comes to CISO and says, hey, I want to run an AI system in our CRM implementation, et cetera.
John Jester [00:09:24]:
So that's what we're seeing a lot. And, you know, Veeam has a different point of view than the industry here. If you think about what an AI infrastructure looks like, you know, the foundation is infrastructure, your compute, your CPUs, your GPUs. On top of that, you've got all your data estate. And then the third layer today is AI models. So it's Anthropic, it's OpenAI interacting directly with your data estate. And then of course, AI agents above that, line of business applications above that. That's what the AI stack looks like today.
John Jester [00:10:00]:
And Veeam's point of view is you should insert another layer between your data and your AI models. And we call it the data and AI trust layer. But what that does is provide security, identity controls, access permissions, so that you don't get those unintended consequences. And because you started with your data, your data is the control plane here. Your data's protected from rogue agents, from your own business-sponsored agents. You've got controls in place. So that's kind of our point of view that I think is unique in market. Protect the data, start with the data estate, and then the rest of the stack can interact with that data in AI trust.
KB [00:10:44]:
So talking on the provision side of things, I remember back in the day having to manually do this, call people up, hey, we've noticed you've got this level of access. Do you still need it? And it's like they probably had this access for who knows how long.
John Jester [00:10:56]:
Yeah.
KB [00:10:56]:
So do you now see these agents doing this, I don't know, every month, every quarter to say, hey, John Jester, we need to make sure, you know, he's got this role, he needs this level of access to make sure. Because again, that's how we've had issues in the past when you're dealing with, you know, contractors and vendors and, you know, system integrators and all sorts of people rolling through big enterprises, it's hard manually to keep track of these things. And then to your point, if some of these provisions aren't correct, then it allows agents to go rogue and all sorts of things. So is that then going to be a, as a byproduct, engender that trust then from a data point of view?
John Jester [00:11:36]:
Yeah. So you just nailed the exact issue. So 2 years ago, 5 years ago, what did we worry about? Hey, there's all this data on SharePoint sites, there's unstructured data across the enterprise, you know, should we run an audit and make sure the right employees have access to this data, et cetera? Honestly, it was a manual process. The risk was not that high because data discoverability was harder back then. Someone had to go to the SharePoint site, they had to click around, they had to find a file that they didn't have access to, that they shouldn't have access to. The challenge now is it's actually the AI agents that are going out and discovering information. And those AI agents are way more effective at finding, discovering information across your data center than any search index was, than any other tool was in the past. So that's the challenge if you have data with the wrong permissions, with the wrong identity controls.
John Jester [00:12:32]:
And we have a tool where we scan and we call it the toxic combination. We know sometimes when you have multi-layered identity controls, you actually do very unintended things like give an AI agent permission to an entire dataset, whereas you meant to just let them have a couple records. That's where this blast radius I described. So an AI agent taking actions far beyond its design intent impacting your business, it's because they had access to data that they shouldn't have. They were over-permissioned. They were scoped in a way that was much broader than what they needed to perform their function.
KB [00:13:08]:
Okay, so John, I want to talk about when boards are sort of asking again, like, can we recover? Like those sort of questions. What are your thoughts then towards that, given the space that you're in? Are people giving sort of evidence-based answers or just, you know, like hope, which is obviously not a strategy, but What are your sentiments here?
John Jester [00:13:31]:
So I think that's what we're seeing right now. Everyone wants to pilot, they want to innovate, they want to take advantage of the amazing business benefits you can get from AI. But we haven't yet built out the right tools and the infrastructure to help customers do this in a safe way. And like I said before, safe for me is I'm governing the AI agents, I'm monitoring every action they take. And then I can understand every piece of data they edited, deleted, created, and I can do precision undo. And that precision undo, bringing data back to the healthy state, that's what gives me the confidence I'm doing AI the right way, I'm taking a safe approach, and if a mistake happens, I can undo it and keep my business running.
KB [00:14:17]:
Yeah, this is interesting. So, okay, I want to get your thoughts then on where are people, companies that you're sort of seeing across the world at in terms of their data journeys? So I know that in order to have really good AI, we need to have really good data. And then where is that data? Is it structured? Is it unstructured? Is it on a piece of paper? Where is it? Talk to me about where sort of are people generally at? Are they at early stages? Are companies more mature in this approach in order to really see the full advancement of AI, right? Because obviously now that's what everyone's working towards to have that competitive advantage, et cetera. But a lot of it, as you know, or majority of it is really stemming from that data. So obviously that's the foundation to propelling AI and getting the best bang for buck from it. So I just, I just want to get you to paint the picture there, John.
John Jester [00:15:07]:
Yeah. So you actually hit the nail on the head on exactly what's important. You have to have the right data feeding your AI models. There's a quote from Sam Altman, the CEO of OpenAI that I really like. And in his quote, he says, hey, everyone can do AI now. We have the tools, we have the technology. What really matters is, do you trust the data that you're training your AI models with? And if you trust that data, then you can trust the outcomes and the recommendations that the AI tools are taking. So that is the issue.
John Jester [00:15:39]:
Do you trust your data? And there's a couple of things that you need to do to make sure you're confident. One is companies have old, obsolete, redundant data in their estate. It's been there forever. They've been backing it up. They've been protecting it simply because they don't know what they have and they're afraid to lose things. So step one is to analyze that data, understand where the old data is, exclude it from training your AI models. Because if you train those models with old obsolete data, you're going to get old and obsolete answers and recommendations from it. So that's step 1 to a healthy estate.
John Jester [00:16:15]:
Step 2 to that healthy data estate is what we talked about before. Do I have the right identity, access, and security controls across that data? Because again, I want to get the business benefits of AI. I want to enable my projects and tools, but I need to keep them scoped. They need to be permissioned to where they only have access to the data relevant to that AI agent role. And they only have the permissions to make changes that are relevant to that role. And when I scope it that way, now I'm confident I'm doing AI the right way and I can be confident. But just coming back to your question, it is a healthy dataset. So you've got your old and obsolete data out and you've got the right identity, access, and security controls across that healthy data estate.
John Jester [00:17:02]:
So the AI agents are working on it the right way.
KB [00:17:06]:
Okay, so I have a rudimentary question for you. How do you sort of know, like, do I trust this data or not? Like, what is a good barometer?
John Jester [00:17:14]:
So a couple things that we look at. So we can do scans analysis, how often data is accessed. A great trigger to tell you you've got old and obsolete data is no one's accessed or interacted with that data for years. So that's a great trigger to start with. So doing that analysis. The second is who is accessing it? So look at it by roles in the company or look at it by line of business applications. Again, that gives you a good indication the health of that data. Is it fresh? Is it being used? So those things I think are the key ways you kind of look across your data estate.
John Jester [00:17:49]:
And then there's just that practical reality of having someone who owns these applications, who owns that data, really refine the access controls, the permission of the AI agents to make sure that they're interacting with the right data. Because I do see this over and over. If you don't have a clean data estate, if you don't get rid of old and obsolete data, you don't get the right answers, the right recommendations from your AI tools.
KB [00:18:16]:
Okay. So, John, final question. Do you believe people are resisting to potentially own agents and data because of the consequences? We've seen it happen in the marketplace. Now there's so many conversations I'm having, who owns it? The vendor, the business, the AI agent themselves as their own accountability. Would you say people are apprehensive here because in case something does go wrong?
John Jester [00:18:42]:
So I would describe it as there's a huge appetite to leverage AI to accelerate everyone's business. And that's being driven by business owners, CEOs, board pressures. So everyone's pushing, everyone wants to do it. What I see is the CIO, the CISO are the ones that have that accountability to enable access to AI, to enable AI for the business, but they're also accountable for keeping the business up and running, business continuity. They're accountable for not leaking data to employees who shouldn't have access to data. Think about acquisitions, HR data, or even worse, data leaking outside of a company due to the wrong security controls, the wrong access controls. So your level of apprehension is on a spectrum of, am I accountable for protecting data and security? Then you need to be careful. You're gonna be apprehensive.
John Jester [00:19:35]:
But that's countered with this very strong demand from business owners, from CEOs, from boards to push companies to take advantage of the innovation, take advantage of AI. And like I said, a great example is someone like me. I'm the CRO of a global company. I want to enable AI to have better customer success for my customers. I want to enable AI to take advantage of making my sales team more productive with CRM systems. But of course, there's a CISO at Veeam who I work with, and they want to make sure, hey, we're not going to leak the wrong data. So that's kind of that counterbalance. You're going to have business owners who are very aggressive.
John Jester [00:20:13]:
They want the business benefits. And you're gonna have people that are accountable for securing data, making sure it doesn't leak to the wrong people, making sure AI agents don't corrupt datasets and break a company's ability to operate, et cetera. So that's how I would, I would explain that spectrum is aggressive advocation in some parts of a company. And then this healthy, we need to do it right. We need to enable safe AI. We need to be careful in those roles who are accountable for it.
KB [00:20:44]:
Okay, so joining me now is Rehan Jalil, President of Products and Technology at Veeam. And today we're discussing how AI is only as safe as the data it can touch. So Rehan, thanks for joining me and welcome.
Rehan Jalil [00:20:57]:
KB, thanks for hosting. Excited to be here.
KB [00:21:00]:
Okay, so I want to start with, can organizations honestly say that they are ready for AI, if they still don't know where their sensitive data lives or who can access it. So talk to me a little bit more about what's your thoughts here, because people are saying that they're ready, but sometimes in reality, they don't actually really know like where their data's sitting. So I think this is a really important question to start off with.
Rehan Jalil [00:21:28]:
Yeah, I think if you first like imagine an enterprise, which has lots of data, not in one place, but in all kinds of places, like their on-premise systems, in the cloud, in SaaS services, data warehouses. When we talk about data, it's not like one thing. It's like thousands of different kinds of things and places. And on top of this data, it's only useful only if people can use it. So, different people can actually see different kinds of data. Different people are producing different kinds of data. And of course, mixed with it, completely jumbled up with it, is confidential information, sensitive information. Depending on the person and the privileges, they're the only ones who should be able to see or not see certain kind of information.
Rehan Jalil [00:22:16]:
That's how typical enterprise data is. It's not like a monolith. If you think about, on the other hand, AI in the enterprise, is only useful if it can dip into this data and create some value and of course generate more data out of it. So it's, on one hand, it's a very messy kind of an environment and very complex kind of environment. And the AI can only be useful if you can actually let it talk to this data and extract things out and generate new information. So how do you make sure that when you do let AI come into it, and AI is also not one thing. When you talk about AI in the enterprise, it is various different kinds of models. It's a very, pretty much bottom-up adoption desire inside the organization that people, employees inside the company, they're all very familiar with consumer AI.
Rehan Jalil [00:23:13]:
They wanna bring that into the mix. So now, if there's no one AI, there could be thousands and thousands of different kinds of agents or type of systems or models. How do you know what AI is even trying to use your already very complex environment, data in a very complex environment? And how do they streamline it so that, first of all, it is the blessed AI that comes into the mix, not nefarious or malicious AI or— basically unreliable AI. And now the how do you make sure this data actually is used in a sensible way? And that is the fundamental challenge in enabling AI inside the enterprise to first create that hygiene, the layer of hygiene above this data so that when AI comes and touches data, you actually know that there's not gonna be explosion of issues that will pop up. as things go forward. That's the problem statement that every organization is dealing with. And of course, they have to methodically figure things out, how to go on this journey of using this AI in the enterprise in a safe and sound and reliable manner.
KB [00:24:29]:
And then, Rehan, would you also say that companies perhaps undercook the complexity to this sort of stuff? Like, because, you know, you and I talking about it today, it sounds easy, right? Like in theory, but then in practicality it's a lot harder. And there's regulation, and then we talk about governance, and then we talk about certain executives have an opinion on certain things where we have to go faster, but we have to go slower from a cyber point of view. So where, and given your role and experience, where do you think, like, how are people sort of managing this perhaps when they're talking to you? Anything that sort of comes up?
Rehan Jalil [00:25:03]:
Yeah, I think there's no one path. Some organizations, especially if they're in regulated industries, they have to treat it and start with the hygiene first. Although there is a very high-level desire to actually use AI rapidly in all kinds of use cases, and some use cases are probably much more controlled and they can be used faster. But as soon as you try to take it to all team members, all employees, and all And by definition, then on all data, that's where the rubber hits the road. In the regulated industries or companies who are large and they actually, their brand is at stake, they first of all, they wanna make sure that they have clear understanding of what data exists, what is sensitive, what can AI touch or should not touch? How do I put monitoring? How do I put guardrails? So it means they are enabling their companies to actually go fast by going slow because they want to put these things first. Because the reverse of it could be very dangerous that you go fast and then some problem, which is bound to happen, we will talk about what kinds of problem, but some problem, whether it's some AI violating your internal rulebook or some AI actually breaking compliance that you must have an obligation towards the regulator, or AI actually destroying your data. Because one thing you'll see that AI is generating way more data than humans. It's modifying the files that only humans could modify.
Rehan Jalil [00:26:41]:
So it could overwrite, corrupt, or outrightly delete data. And how do you know what is deleted across your entire estate. That's something that organizations have to deal with.
KB [00:26:56]:
Okay, so there's a couple of things in there that I find interesting. When you said before, go fast to go slow. So even like 6 months ago, what I was hearing from people like yourself saying, oh, we just gotta go really, really fast and sort of jump straight in. Now I'm seeing it's almost like, well, we went too fast and now we have to retrofit potentially some of the issues, problems, complexities, things that we could never foresee. So are you seeing in the market now people have— the word burnt is not the right word, but I'm sort of saying it as a bit as if to sort of people to slow down a little bit more in terms of, well, actually we may need to measure twice and cut once, so to speak.
Rehan Jalil [00:27:39]:
I think it's just human nature to experience something to actually understand what it really means. And frankly, sometimes it's just not in their control because on one hand you have a large swath of team members and employees just bringing AI or trying to bring AI. And there's smaller teams who are actually designed to basically create the hygiene layer, the protection and security governance layer. So it's creating basically a dynamics where as problems are arriving and surfacing, aha moment is coming where we say, no, I can't continue to operate like this. Which means I need to put the guardrails into the mix. And if I do put guardrails, I will simply be, by the benefit of it, will be enabling actually more use of AI because now I know what this basically, what the guardrails are, in which lane the AI can stay. Let me paint those rails and then I can say, okay, go fast.
Rehan Jalil [00:28:31]:
I've already defined those rails, right? That we are clearly seeing happening in the industry right now, where many customers have kind of dabbled into it full throttle and learned through the experience that guardrails are super important. And there's those initiatives you're seeing at broad scale. Those initiatives are actually growing. It's a very common ask to actually understand the data hygiene of the data, as well as now guardrails on AI.
KB [00:29:01]:
And then going back to your comment before, like at the moment AI is destroying sort of your data. So talk to me a little bit more about that and how that looks in your eyes?
Rehan Jalil [00:29:10]:
Yeah. So AI inherently is non-deterministic. And if it was your self-driving car, which was non-deterministic, it's bound to make mistakes, take wrong turns and maybe sometimes have an accident, right? So it's pretty much the same thing here. If you have agents, that are allowed to take actions and bound by the fact that they're non-deterministic and the data that they create, information they create, is not always right. So either they can override with some information which is not right information, or it can outrightly delete information, which has also happened, uh, also happening actually much more recently, that the self-driving AI agents can corrupt, modify, or outright delete the information. So that needs to be accounted for. And to do that, it is, first of all, the first preference always is to make sure it does not happen, which is monitor and protect it from happening. But if it does happen, then at that point, the choice really is that you're prepared for it.
Rehan Jalil [00:30:21]:
And the way to be prepared for it is to be resilient, like back up your data with precision, because backing up just data en masse is not sufficient. You have to be very, very precise in terms of what are you backing up? Are you backing up the things that AI agents are modifying? And if some destruction happens, can you recover very precisely back from only the data that was destroyed by the AI agent?
KB [00:30:50]:
So, how do you know, or how does a company know, like, hey, something's been deleted. How do they start to explore that? Because sometimes when, like, I'm looking for something, I don't notice until I'm actually going looking for something and figure it out, I've lost the thing.
Rehan Jalil [00:31:05]:
Yeah. So, it's about monitoring. So, think of this way, if you actually have all your production data and if it is getting monitored, and not only the data, but on it, you actually have understanding of identity. And why identity? Because agents will have a certain kind of identity, or you, and along with that, you're seeing an activity. So if you have data understanding of who's, who are the users who are active on it and could be agents, and/or you infer them as agents, and then there is clear activity that was modification, deletion. And if you can actually see that with clarity, or you can find it with clarity, or if you can alert it, clarity. At that point, you know that actually this destruction or corruption happened by an agent. At that point in time, it gives you the, the power, first of all, to go ahead and prep it at best.
Rehan Jalil [00:32:00]:
If you start seeing such activity which may convert to destruction or, or corruption of data, you may want to start backing up this data. And if something goes wrong, you should have the ability to bring it back. From your resilient systems. But again, their point there would be, KB, is that they have to be very precise because if I modified, my agent modified 100 out of a million files, I don't want to bring all the million files back. I just want to make sure those 100 files actually are being brought back to life just at the right time, just before the agent had destroyed or corrupted them. that version, you wanna bring it back. And that's some of the tech that we have built also, that a full understanding of the live data, full understanding of backup data, and having the precision of knowledge, like what is an agent doing, and have the precision of recovery, that what do I need to recover if an agent has destroyed it?
KB [00:32:58]:
So what happens when you're aware something's gone wrong, another agent's deleted a bunch of critical information. How quickly can people sort of intervene with that? Because what if that information is used to sort of build another model for something else? So how do you sort of, how quickly can you go, okay, we've got to stop this because that then stops the bleeding because like the blast radius could really impact a company.
Rehan Jalil [00:33:24]:
Yeah, I think there are 2 things. Deletion, finding something is deleted for certain systems could be very quick. Like, if not instantaneously, you can, you know, delete action is happening. But you're gonna also find that some delete action may be very desirable. I'll give an example that somebody may be cleaning up the data. It's old, raw data and somebody's cleaning it up. Right. Or you have created automation to clean up your data and move it from somewhere to somewhere.
Rehan Jalil [00:33:53]:
But the question really becomes is that, can you find, can you have alerts Or can you actually have logic that tells you, oh, this sounds like malicious? So there's one way to actually figure that out much more quickly if you're seeing patterns of certain kind. The other thing is that when it does, often people will not find out till quite some time back. Can you go back in time further? Even if it happened a few days back, can you go back in time a few days back and still be able to take and recover back. Both are important. Like, real-time understanding or near real-time understanding is important, or able to go further back in time and recover back whenever you know it is strategically free to recover back. The beauty of having a resilient solution is that it allows you to do both.
KB [00:34:42]:
So would you say, and I get the resiliency piece, but just going back to the mindset, do you think companies are now armed with, hey, because we're experimenting with AI and agents, et cetera. This may be a thing that happens. You know, we don't want it to happen, but it could happen. Something critical could be deleted. But again, we gotta have a good resiliency plan. I think I spoke to your CRO a couple of weeks ago about the rollback, what I called it sort of the rollback capability that you guys are doing. Mm-hmm. Is that the mindset that's now being engendered into companies? Like, fear is not the right word.
KB [00:35:13]:
It's just more having this in the back of their mind. It could be a thing. And if it is a thing, this is our plan.
Rehan Jalil [00:35:20]:
Well, clearly, I think that we were seeing that realization is kicking in for 2 things. One is an agent can make a mistake. An agent who was actually allowed to be used actually made a mistake. And then what is your plan to basically recover back from such a mistake? The other thing, if not agent, but frontier models with some advanced agents, that is actually a real fear, real, more than a fear, I will say, is it'll become reality at some point in time that these frontier models and the harnesses around it are so smart, they can not only find the issues and vulnerabilities, in fact, multiple issues and vulnerability, and stitch them together to actually exploit it, right? So if that happens, which it's— I think it's just a matter of time, you have a flurry of these things coming. At that point in time, it was not the act of a blessed agent. It is the act of actually some malicious agent that attacked you and came with a blazing fast speed, found the issues in your environment, not one of them, multiple of them, chained them through in a way that it can actually exploit it. And at that point, because it's so fast, the only defense left is that you're resilient in the sense that even if attack happens, you can bring your data back or your processes back and systems back, right? So from AI lenses, both is actually quite, quite important. Um, a blessed agent making mistakes and, um, and real advanced frontier models with agents actually deliberately, uh, stealing your information or ransomware, having ransomware And your ability to recover back from it.
KB [00:37:09]:
So what happens if you're not resilient? Now, I caveat that with saying, we talk about it a lot, but it's still the reality that companies are not because they can't back up. I think there's certain reports out there that certain businesses take 24 days to get back up and running after something happens. So there's clearly the resiliency piece just isn't there for some people. So what do you, what realistically happens then?
Rehan Jalil [00:37:34]:
I do think that at this point, there won't be much of a choice for the companies who want to continue their business operations. Unless you say, it's okay, I'll be down for like a few weeks and maybe I'll figure something out. Even with that, if your data is gone, at that point, what do you do? For instance, if you're a financial institution or people actually give you money to basically hold in your bank accounts. And if somehow you lose that data, you're actually losing money, right? Or if you are down with a service for extended amount of time because you just can't recover fast enough, you're actually losing money yourself, right? So I don't think it, especially in the face of the advanced frontier models and the speed with which they can exploit, the second layer of defense, one is the protection, other is resilience, is going to become more and more important. And that realization is kicking in. Remember, we are in the very early days. And that tsunami of attacks that people do expect, including myself, that when it does come, people who are not ready, they will actually have a significant downtime or major disruption for their businesses.
KB [00:38:47]:
And so what, at the moment, with the people that you're sort of speaking, you sort of can talk in general terms, What is it that you think is really top of mind for people at the moment?
Rehan Jalil [00:38:58]:
The top of mind is, again, the actual advancement of a company, whether it is from how they work, and everyone wants to stay very relevant in this age of AI. So generally, if you go across companies, this would be a very common topic, that how they are transforming themselves. to use AI for coding or their legal or their finances or marketing or content creation, you name it. Like across everything, organizations actually have that desire to kind of go fast and that's what's driving it. But to do that, let's say even if you want to build the fastest race car, you need the brakes to be in place. Without brakes, you can't put that car on track. The car could be very good looking, car could be amazing, engines inside it, but you need the brakes and steering wheel. That's what we're seeing, the desire to actually have that AI, really cool AI stuff, like a race car.
Rehan Jalil [00:39:58]:
But the guardrails, AI guardrails and data guardrails, which is like brakes and steering wheels, they have to go into place. So that's what we're seeing.
KB [00:40:07]:
So now I just want to pivot for a moment and talk about the problem around AI security, would you say it's A, the model, B, it's messy and over-permissioned data, or C, a state that sort of sits underneath it? And I know that's sort of a big question to answer, but I'm curious to hear your thoughts on this one.
Rehan Jalil [00:40:32]:
I think it's both. Why? Because they're very interrelated topics, KB. Because data on its own doesn't mean much till a human can use it. So it's a permissioning on the data for a certain human to read, write, edit, and all that. And for AI, it means not much if it also can actually read and write. And for AI, it doesn't mean much till actually it can benefit a human, an identity on the other side, human identity on the other side. So it's a very interlinked topic, not separate topic. And layered above it is, of course, your rules and compliances and other things that you have to adhere to.
Rehan Jalil [00:41:10]:
The way we have thought through is to not look at these pieces in silos, but look at them in unison. What I mean by unison is in one graph, one knowledge graph, that not only these, so to speak, nodes or files or data systems and your AI models and agents, they exist independently, but how they correlate to each other. What is the relationship between them? And if you can capture that properly, and if you can capture the activity based on this relationship map, based on, of course, the graphs, knowledge graphs, then it gives you a lot more power to define what is called like a bad behavior or toxic combination inside the graph. And then use that to actually find out what is the good from the bad and prevent the bad from happening right along the good to happen. So I would say it's an interplay between data, AI, identity as one. That's why we call it DataAI. DataAI, of course, artificial intelligence, but AI also for access and identity, pretty much combined as one unit.
KB [00:42:18]:
And talking about that combined unit, do you think that that potentially causes 'Cause if we go back for a second, people back in the day would say like, oh, like IT is its own, like it used to be its own silo, and then cybersecurity is kind of its own silo, and now it's sort of becoming more integrated. Do you see this unit becoming more integrated then as we sort of progress and move forward?
Rehan Jalil [00:42:42]:
I do see that the source of truth could be common across different departments, like the cybersecurity department in AI and infotech, or even privacy, because source of truth is the same. You have data, it has something sensitive, it has some people using it, AI is trying to touch it, and there's some activity. That source of truth is same whether it's for cyber team or privacy team or governance team or AI team. How they utilize it, what aspect of it is of interest to them, and what aspect they want to control and govern, That's different based on which lens you actually fund or which lens you look at. So point being that if you can establish that source of truth, then you can service multiple personas inside the organization without having a disjointed and frankly misaligned source of truth. Establishing the source of truth could be a big service to actually go, again, go fast on your AI journey.
KB [00:43:43]:
And how would you sort of recommend people sort of start to establish that source of truth?
Rehan Jalil [00:43:49]:
It has to be the least amount of human effort because data is not static. It's like a massive river that is changing all the time, right? New data getting created. Now AI is creating probably 10x more data than a human has been creating. So there's no influx of entropy in this ecosystem. So, I think that really is the crux of it, that you actually have data that is changing all the time, and you're gonna make sure that you have a good control of that first.
KB [00:44:27]:
So, I wanna talk now for a moment around, people are talking to me a lot about AI governance, and that sort of without data context or pedigree and access visibility, for example, is like governance theater. I keep hearing this governance theater sort of word. So talk me through maybe your version of this.
Rehan Jalil [00:44:53]:
Yeah. So if you're, let's say you bring up an AI, you give an AI agent to an employee. At that point in time, what is important? The important part is that when you turn on an agent, or AI system in? What data can it see? So privileges management becomes very important because remember, this AI system and agent can be fast moving. So if there is a mistake in the permissioning, it's gonna get access to the data, it's gonna use it before you go, before you figure it out and fix it. So establishing that clarity upfront and maintaining that right privilege management is actually important. Right. So, and if this AI agent gets access to something sensitive that was not supposed to get, and then generates some 10 documents and spreads that sensitive information without you knowing, you're basically creating more problems along the way. The point being that you definitely need to know what data the AI can touch.
Rehan Jalil [00:45:57]:
You definitely need to know if there's anything sensitive inside it. We definitely need to know what it generated and where did it place it because a new data got generated. And if there is anything sensitive inside it. And over time, I do think that people would need to know what data is generated by human versus AI with clear separation. Because there used to be a time when one person would write a file and many people would read it like a book. There were fewer writers and more readers. It's kind of changing a little bit. Anybody who has an agent can write like 50 files in a day, and often not everyone reads it either now.
Rehan Jalil [00:46:38]:
It's just that content flux is so high. It's going to go even higher where AI generates like massive documents, we'll read a little bit of it, and there's a new version gets created of it next week or the next day. So you have to think about how to manage that fast flux and fast increase in data volumes because you may have to retire it faster also. Because remember, next AI agent can get hold of that file and start using it, right? So which means start feeding it data on itself, or data generated by one AI agent starts getting fed to some other agent. So management of the data and retiring of the data. And knowing what actually got created for little use, ephemeral use, disposable use. There's a lot more disposable data at this point in time. Previously, I think there was less and less human-created disposable data.
Rehan Jalil [00:47:33]:
Now there's going to be more disposable data. It's how do you manage the waste much more quickly and effectively so it doesn't get used and get mixed up in your content is another thing that people are dealing with.
KB [00:47:46]:
Okay, so just staying with the retiring of the data, do you think that now companies have generated something off this retired obsolete data that may be counterintuitive to what they're going to do? Is it sort of they have to go back and think, well, we better go and have a look at that data because actually that was pointless?
Rehan Jalil [00:48:07]:
Yeah, I think so. That's all about intelligence. How can you infer intelligence, what's in the content and what's useful and what— and it ties to— intelligence is often tied to how many people are actually reading and writing it. So intelligence ties to not just the content, but importance of it in some ways inferred by how many people are actually using it actively, reading it, writing it, and so forth, right? Those signals become important for you to say, should I preserve or just dispose this data? because nobody's touching it either. And when you dispose, when I say dispose it, it's better to actually go archive it somewhere, right? In a very cheap place, just archive it, wait for a given amount of time, say 6 months, a year, whatever. And then you feel comfortable that nobody is asking for it and it's better to just be gone, right? And it's just that now it's going to be, that cycle is going to be faster, much, it has to be faster. Because a lot more disposable data is getting created by agents. And frankly, that pace is just simply going up.
KB [00:49:13]:
And do you think people know, like, so people are obviously in this retiring of the data phase at the moment. Like, how will they know, like, what's sort of worthwhile? Because then there's going to be, like, different versions that morph from the original data source that then is like, hey, we're using it for this, but then we've actually built something on top of that. Like, will it be hard then to sort of keep track on what is actually the most valuable data source, or there's ways of mapping that in terms of more fidelity?
Rehan Jalil [00:49:43]:
So there's no one answer to it. It's not going to be easy. So there's signals you would pick upon to understand whether it's still useful or not. And what are those signals? Signals, if data has been sitting for the last 2 years, nobody's ever opened it. It's been— nobody's ever written on it. It's just sitting there. It means it's of no value creation. And if you inspect this data and if it doesn't have anything sensitive that's required by GLBA kind of regulations that require you to preserve it, retain that data, at that point in time, you can feel comfortable to say, maybe I just need to dispose it off, right? So, you have to pick upon certain signals and the type of content and your own rulebook to make a judgment call to say, I can archive it for right now.
Rehan Jalil [00:50:31]:
And then if nobody's complaining for a year or 6 months, at that point, I'm very confident that this data is gone, right? It ties to your retention policies and so forth. So, yep, but it requires intelligence, intelligence on data and what's— and the content inside it.
KB [00:50:48]:
So do you believe that companies historically have been holding onto data that they don't really need for like a protracted amount of time?
Rehan Jalil [00:50:57]:
When we work with many companies, it becomes very clear that a lot of data is sitting in their more expensive NAS or more expensive live data systems, and nobody's used it for years. Nobody's used it. So not only it's actually expensive, it increases the risk of the— if a breach happens, your blast radius or the impact of the breach is actually way more because you just have so much data that's sitting there, right? So, part of the reason is, as I mentioned, KB, is about intelligence. If you have intelligence about what's in the document, what's in the data, how much is being used, if you can infer that reliably, then you can tell which data should be gone. So, organizations do need the tech where they can reduce the cost significantly by moving data from very expensive live data systems to actually archival systems. But more importantly, you can reduce your compliance risk, exposure risk, and access risk, and all kinds of risk because more data sitting there really means you have potentially more exposure happening. So, both ways actually it helps cost-wise and risk reduction-wise.
KB [00:52:10]:
So, just on that, do you remember like even like 10, more than 10 years ago, like in the industry, companies were collecting all this data 'cause we wanted to personalize stuff to you. Maybe it was a home loan or mortgage or, you know, You know, whatever it is. Then people, all these breaches started to be, rose to prominence. And then people are like, well, we gotta get rid of all the data because we're holding too much PII. So how do companies find the right fit between we need to hold enough data in order to have effective AI, we don't want enough data where perhaps we're holding random PII that we don't need in case there's a breach, and then of course looking at our data in terms of, well, maybe we're holding onto something that we A, don't need to your earlier point, and B, It's expensive.
Rehan Jalil [00:52:49]:
Yeah, goes back to intelligence and also understanding. If you, let's say if there was a magic to tell you like all your data historically, what exists, how many have duplicates, how much you've just obsolete, you've never used it. Can you just get rid of the duplicates first? Because you have still the same information, it's just that they're not sitting in 5 different places. So, can you cross cluster them together. And then if there's old data, you can make judgment calls. So most companies will define a bit for themselves. There are certain things without losing the value of it, can I still remove a bunch of it, right? And if data is very old, if PII is very old, it may not be useful to begin with because their customers, they actually evolve over time. Right? So that's what we've seen there.
Rehan Jalil [00:53:42]:
If you are a mortgage company and you have contracts and leases sitting for the last 10 years and that's not even active, or insurance company that your claims or your agreement, insurance agreements, which are not active, that employee, that customer's gone for a long time, that may not be very useful information to begin with, but there still is a huge amount of risk. And if some breach happens, you would need to still disclose it. You will still need to pay fines because you actually have a breach that happened, and within the jurisdiction of the state, the resident state for that particular persona, you'll still have to pay fines based on where the person resides. So we've clearly seen that it has to be much more methodical, and the tools exist. We provide some of these capabilities where you can really have these shades of gray and define where do you cut your line. But every company has to define where do they want to cut the line, but not without having enough intelligence.
KB [00:54:45]:
And would you say as like a general consensus, people are at that, where do we cut the line sort of period at the moment?
Rehan Jalil [00:54:53]:
Very much so. AI is driving some of it. And frankly, some regulation drives it. But now AI is driving and she's half of it. People want to get rid of old data. Part of the reason is if AI gets hold of your old data, it's going to assume it's the right data. It's going to create basically old answers or irrelevant garbage answers. So, for the quality of data and the quality of answers it creates, it's also better to get rid of, or at least know what is old data and label it to tell the AI that please don't touch this data because it's just, there's no use or no good at this point in time.
KB [00:55:30]:
So just a quick question on that. Do you envision that companies may lose revenue if they're using something that's based on old data and maybe it's telling customers something that it shouldn't be? And it's like, oh, well, that's actually the wrong answer because it's using data from 10 years ago. That's outdated.
Rehan Jalil [00:55:48]:
Very much so. I very much so think that if your source of truth guides you to run a wrong marketing campaign, address wrong people, or send something to old email addresses, which doesn't really make any sense, it obviously is a misfire, right? You are not targeting the things very well. Or if you're doing messaging based on certain things and you're messaging it wrong, there's so many things that you can actually do wrong if your source of truth, which is the data itself, is still not relevant anymore.
KB [00:56:21]:
So now, Rehan, what do you sort of think now moving forward? We've already discussed a lot of terrain today, but what is— what do you think is going to happen between now, end of the year, into next year?
Rehan Jalil [00:56:34]:
KB, the biggest thing we all have to grapple with, the insane amount of intelligence in the frontier models and the agents that can utilize this model to basically pose a threat which we have never, none of us ever seen before. Not only because of the level of intelligence, it's just the speed. And the other thing is these agents is not one agent. Clearly we have seen in recent examples they can swarm. So which means there's not one agent. You have a bunch of agents that can collaborate to increase the intelligence about achieving a goal. If the goal is to penetrate and do ransomware against a company, this swarm of agents can be out there finding your surface area, learning from it, and frankly educating each other and forming basically a unit which is so smart inherently as a model and then as a swarm of agents and the infinite amount of attempts it can make because now GPUs are available, it can just keep going, that the penetration risk is very, very high.
Rehan Jalil [00:57:47]:
We have not seen that as yet because many of these models and agents are just not in the wild with everyone. They're very controlled. But it's just a matter of time that this will happen. And when that happens, the importance of very strong defense in depth, which actually has to be different also. It has to be very automated. It's not going to be old-style manual remediation. And at the same time, it'd also be resilience. Your last stop is, can I recover back if something goes wrong? That's what we— I sincerely believe we're heading towards that near future that this is going to be happening.
KB [00:58:31]:
And there you have it. This is From the Field. Stay tuned for more.
