Sensitive data no longer stays where it was created. It is copied into lakehouses, SaaS applications, test environments, and AI pipelines, and accessed as often by services and agents as by people. Most organizations have responded the traditional way, by buying another tool: discovery, encryption, masking, monitoring, posture management, each with its own policy model and its own idea of where its job ends.
Data Security Platforms promise to replace this patchwork with policy enforced wherever data is accessed. The market is full of products that can tell you exactly how exposed your data is, and rather fewer that can do something about it. This webinar presents the results of the 2026 Leadership Compass on Data Security Platforms and helps tell the two apart.
Alexei Balaganski, Lead Analyst and CTO at KuppingerCole Analysts, will present the key findings of the 2026 Leadership Compass on Data Security Platforms. He will explain how the products were evaluated, from discovery and posture management to crypto-agility and data security for AI systems. He will show where the market has matured since the previous edition, where vendor claims still run ahead of the evidence, and which vendors lead in product capabilities, innovation, and market presence. Finally, he will offer practical guidance for buyers on selecting the platforms that actually enforce policy instead of only reporting on exposure.
Hari Prasad Mariswamy, Director of Product Management for Data Protection at Skyhigh Security, will bring a practical, real-world perspective to the conversation. Drawing on more than 20 years of experience in data protection and cybersecurity, his expertise across DLP, CASB, email security, and cloud data protection will help connect strategic trends and emerging risks with the operational challenges organizations face every day.
Who Should Attend
This webinar is designed for CISOs and security leaders, data security, platform, and enterprise architects, privacy, risk, and compliance officers, IT decision-makers evaluating or consolidating data security tools, and product and strategy teams at security vendors.
Well, hello and welcome to another KuppingerCole webinar. My name is Alexei Balaganski. I'm the Lead Analyst and CTO at KuppingerCole Analysts, and my guest today is Hari Prasad, who is the Director of Product Management at Skyhigh Security.
Welcome, Hari. Hello, Al. Great to have you today. And before we dive directly into our presentations, we would have to say a few words about housekeeping rules. Everybody except us two are muted, so you don't have to worry about your microphone. You can always submit your questions using the dedicated questions tool of this Livestorm platform. Please do it at any time, but we will answer your questions at the end of the webinar session. We are recording this, and we will make the recording available on demand on our website, along with all the slides latest tomorrow.
And of course, everybody who registered for the webinar will receive a notification. And I guess I also have to apologize in advance for my probably slightly weird voice today, because I'm starting with a cold, but I will try my best. And without further ado, let's dive into the webinar.
Formally, our agenda is split into two parts. First, I will be giving you an introduction of the notion and the market segment for data security platforms.
And later, we will have a panel discussion with Hari, discussing the real practical implications of those developments. But we actually wanted to make it a little bit less formal, so we will be breaking up the presentation with some informal discussions as well.
Again, unfortunately, you cannot join, but you can always submit your questions. So where does it all begin? For decades, the very notion of data security relied on three major assumptions. That we know where our sensitive and precious data actually lives. We know which systems control access to that data. And we always know who is accessing it. Unfortunately, these assumptions are just no longer valid in the age of AI and everything else. First of all, the data exists in multiple copies, which are constantly on the move.
They're being transformed and transported across lake houses, test and analytical platforms, cloud platforms, vector stores for AI workloads, and so on. Second, we have so many different new interfaces for our data, including but not limited to MCP and other agentic interfaces. And finally, so many agents, so many entities are now accessing the data that it's simply impossible to keep up with them using traditional tools. So the traditional, again, individual controls we had in our apps, in our databases, in our APIs, and everywhere else, just, they're just too many of them.
And they are not working together. So somehow, we have to move the enforcement somewhere else. Enforcement has to follow the data. Most enterprises actually already have a lot of, quote, unquote, data security tools. And they put those quotes around them, not because I disrespect them or anything like that, but simply because, again, they are tools. They are disjointed. They are specialized. They're owned by very different teams. You might have data discovery classification tools. You might have encryption and masking solutions. You might have activity monitoring.
And even, you might even have a thing in place which I personally quite dislike, a posture management solution. And if you ever followed my published research so far, you might know that this is actually my pet peeve. The agency and SPM is just a term I actively dislike. And this is why we are pushing this notion that something else has to replace it. Every one of those tools arrives with its own policy model and its own idea what exactly it's actually supposed to protect. The result is that you have too many gaps.
You have too many opportunities to miss a drift between the supposed state of your data security and the actual one. And the favorite report of the entire internet, the IBM Cost of Data Breach Report, recently stated that global average cost of a data breach, a single one, mind you, is almost $5 million. And more than half of those organizations affected by a breach never actually had even the most basic encryption controls in place.
This year, they also investigated specifically AI-related breaches. And they found that about one-fifth of affected companies actually had an AI-related breach. And the vast majority of those companies simply had and still don't have proper AI access controls. So something definitely has to be done about it. These breaches did not need to have some clever attacker, a sophisticated nation state agent. They only were possible because of those gaps.
Harry, is it something you would agree with me? Or do you have a different opinion? I agree with you. That makes total sense, Alexey. And these are some of the trends that we are seeing as well amongst a lot of our customers and hear from them as well. So 100% agree. All right. So as I said, security posture management is effective of mine for a long time. I've been saying many times, if a solution exists only to tell you that your house is on fire, without actually giving you a fire extinguisher or just like a water hose, that solution is not very useful.
Because you are left with a report, but no means to actually act on that report. And at the age of AI, when the decision must be made within seconds, not weeks or even days, just the entire security posture management as a notion becomes kind of pointless, if you will. I'm not saying that it is useless. It is necessary. This is the first prerequisite for every data security solution. But it's definitely not sufficient. A finding still has to be routed first to the team that's actually responsible for that It has to be prioritized against other findings, which are plentiful nowadays.
And then somehow it has to be fixed manually or automatically. And then that fix has to be evaluated, analyzed, put into the overall updated posture management, if you will, and finally closed as a risk finding. So nowadays, the vast majority of platforms, and I'm kind of running ahead of the story a little bit, because I'm referring to the numbers from our leadership compass, which I will be talking in detail later. But basically, the point is here, the vast majority of solutions we've analyzed can give a continuous posture score.
So they're even doing better than the minimal viable posture management product. But only half of those solutions can actually help you remediate that finding automatically. So there is still a lot of growth opportunity in that area, if you will. To put it simply, a finding you cannot act on is not a security outcome. It's just homework, if you will.
Harry, is it something you would argue with? No, I think the analogy you've given is a great one as well, Alexey. The smoke detector to a fire extinguisher absolutely completes the whole circle of how discovery monitoring to a protection should be looked at. So that's the right way. Right.
Okay, so what is a data security platform? Well, to put it very simple, it's, well, it's a platform. So it is a combination of tools, but a combination that unifies those tools into a single underlying framework, a single pane of glass as a control pane, a single analytic solution, if you will.
So it combines all those capabilities we have named earlier, discovery, classification, protection, access governance, monitoring, and compliance into a single control model, which means that the policy you create for protecting your data is defined only once, centrally, and then somehow enforced throughout the entire data layer, regardless whether it's in the cloud, on-prem, in a database, in a file storage, or an object store, doesn't matter. So somehow it should work across your entire data estate.
One important addition here is that data security platforms do not replace identity management solutions, data catalogs, SIEM tools, key management and encryption platforms, and so on. They all depend on and integrate with all those things. And those integrations have to be two-way, if you will, because on the one hand, for example, identity providers obviously give extremely important context for every data access decision, like who is accessing it, and do they actually have an entitlement for that, but also data catalogs. They give you business context lineage for that data.
The actual data platforms, the databases, they give you, potentially, the native controls you could reuse. Of course, key management systems give you cryptography primitives you can use for data protection, and so on and so forth. And you can exchange on signals with all those tools, and even better, if that exchange happens in two directions. So a platform whose integrations can only run one way, it doesn't actually solve the fragmentation platform. It only adds more complexity.
Again, a couple of numbers for you. Those integrations on their own are, as I say, now table stakes. Ninety-five percent of all solutions we analyzed can do that.
But again, only half of those can actually give you policy as code, meaning an opportunity to define a policy in a declarative way, which can be apportioned, analyzed, tested, but more importantly, automated for the age of AI, if you will. So what are we actually observing?
Like, what are the forces? What are the developments which actually shape and reshape the market? First of all, as I mentioned, enforcement replaces visibility. AI was a great force multiplier for this. Nobody has time to review PDF reports. Nobody has time to respond to Slack alerts, especially if you have thousands, if not hundreds of thousands of those. Security findings have to be turned into controls and then automated.
Second, there are so many non-human principles now for accessing your data, and contrary to what a lot of people are saying, they are not just AI agents. There is a lot of other non-human identities out there, but still, there are so many of them now that the sheer number kind of changes the entire existing balance in the market. You now have to be prepared for a much higher level of automation, even if your industry for decades has opposed it before.
Third, you don't want to spend a lot of time on post-quantum, but it is unfortunately a very real deadline. You have to take control over your entire cryptographic estate, and you have to know where it still lacks the post-quantum hardening. So you have to know how your data is actually protected with cryptographic algorithms, and you have to know that by a very specific deadline. And finally, the regulations.
With all the talks about data sovereignty nowadays, every country, every industry, every state has their own regulations, their own requirements, and the same data, the same data set might undergo completely different requirements depending on actually what data you are storing in there. Is it data about, say, German citizens or Indian citizens? And of course, is it actually stored physically in Germany or in the US somewhere? So each of those forces pushes the control closer to the data.
And again, Harry, what's your perspective on this? You're looking at it from a somewhat different point of view, I imagine. Do you see the same? We see the same, but understanding what are the forces which are actually making these other things come up to the fore are the emergence of technology, as we know. So let's say the advent and invention of these supercomputers now, which is becoming a norm, and those supercomputers are able to break the cryptographic rules much faster than before.
And hence, it's also given birth to a lot of AI right now, and AI is accessing everything that we know in a much different way, right? And because of which, all of these things are more important now in order to understand what is a human identity versus not a human identity, and then how do we actually respect the regulatory compliances that are coming up so that enterprises can somehow take control of these things which are shaping the future. So they are absolutely very relevant.
Right, right. And again, the most fascinating aspect of all this to me is that things like five years ago, AI was science fiction. Quantum computing was science fiction.
Nowadays, it's not just reality. It's something you have to, like you are forced to deal with because laws and regulations, there is no way around it. So we have to start today the latest.
One thing, of course, everybody is now talking about is agentic AI, and we have to mention it as well. The problem is actually pretty simple. Everything used to happen, everything regarding data access control and authorization used to happen at the application layer. You had a business app. You would log in as a user Alexei or user Harry. The application would know what entitlements you possess, what's your role within the enterprise, and give you appropriate access or deny access to specific data. Unfortunately, this worked when the application was the only way to access data.
Like in the hotel, there is a special guy at the front desk who has the master key to all rooms. But the guy is trained, he's very strict, and he never leaves the desk, if you will. Now we have agents that use MCPs and other existing interfaces. They kind of go in completely around the existing business applications. They ignore all the existing business logic and access rules. So now you have so many more doors and the existing locks are just completely gutted and the doors are broken and all rotten and rusty.
And there is this interesting idea emerging now and discussed that you have to actually, they explain that AI security means you have to build all your security controls around the AI. But this is so obviously and blatantly wrong. Because again, there are so many different ways the AI can access your data, your sensitive data. You cannot put all the guardrails just on the model. Because first of all, they are kind of voluntary. A model can easily ignore them, those guardrails, if they're not implemented properly. But second, what if there is another model? What if there is a malicious agent?
What if there is just an outsider which is accessing your data? And of course, they will be ignoring those guardrails completely.
So again, this, I think, this reinforces our statement from earlier. You have to put your security closer to data. Because agents just and other, again, in this context, the agent doesn't mean AI. It means anybody with agency, anybody who is after your data. They can skip the application layer. They can definitely skip the guardrails. They can go directly to your data, and you have to prevent that.
And Hari, is it something you hear as well? I do. In order to make these things a lot more challenging, there are these ephemeral agents with the advent of MCP service, like you said, Alexei. So those give birth to agents which are short-lived in time. They are designed to create and complete a certain task, and they go away. But in the meantime, they would have accessed your data, done some changes to it, which did not happen before. If there was a human, there was trackability, and you can have them accountable.
But now even the accountability is held challenging, because that's where these agents need to be profiled. And you've got to know who is doing exactly what on your data.
Right, right. Again, something of this, I think, puts even more stress on this whole notion.
Yes, you have to be able to deal with non-human identities, and a lot of solutions we actually analyzed claim to be able to do that. But it's not enough to just monitor them. You actually have to be able to stop them somehow, ideally at every request. So your data should be able to say no to everybody, everyone, everything. And we already mentioned cryptography, so I will not stay on the slide longer. There is one aspect I wanted to mention specifically, in that we asked the same questions to every vendor. How is your solution getting prepared for the quantum deadline?
And quite a few just said, well, we only deal with AES-256. We only deal with symmetrical encryption. So all those new post-quantum standards developed by NIST do not apply. We are safe. We are quantum secure by design, if you will. And this is a very dangerous assumption.
Because, again, kind of a quantum computer being able to break your existing encryption doesn't just mean that they can read your secret Coca-Cola recipe or your financial records. Primarily, they can completely disrupt your existing trust architectures. Because they can break the algorithms which are used for identification and everything, starting with users logging into your applications, to the actual application data being delivered over the wire, TLS encrypted. All those channels will be compromised if we actually get a working quantum computer accessing that data.
So whenever some vendor is claiming, is making this claim, we are quantum safe by design, well, don't take that with a huge grain of salt and ask for more details. I think this is actually very relevant for sky-high security. Because you are basically mostly dealing with data on the wire, right? Yes. So we deal with quite a lot of data on the wire. And it's almost becoming the way vendors are expected to deal with quantum cryptography and understanding these asymmetrical protocols is no more an innovation, right? So this is becoming table stakes.
It's almost like in the olden days, you had vendors do anonymization of your data so that data is minimized to respect GDPR regulations and laws. Like that, cryptography will become one of the most requisite entities of every security that actually monitors your data going in and out, especially at a gateway level as well. So it becomes absolute most requirement. And we are accelerating, making every investment, and we can already understand quite a lot of these protocols. All right.
Okay, moving on. So the obvious question, anybody would say, okay, you had me at data security platform. Tell me how it's actually supposed to work. Where do I put all those controls close to my data? And we usually see three different approaches among those solutions.
The one, the very popular one, is agentless. They'll just say, we use cloud-native APIs. We will connect through your single, let's say, AWS or your Snowflake or something like credential, analyze your entire cloud-native data presence. We will never stand in an access path. We will get the entire picture within minutes, and you will have a working security deployment within hours or like in a single day. Sounds awesome, but you have to clearly understand the limitations. There is a lot of work there for assessment and detection and, again, security posture management.
But without any inline real-time enforcement, the actual protection options are very limited. Some of those solutions can go an extra mile, which is represented in a separate architecture, the native orchestration, where a solution will just say, okay, we know how the internals of this particular data platform operate. We know that they already have their native security controls, but we can do better. We can orchestrate and automate mass deployments of those security controls for you. We will do it centrally. If you have several platforms, we will translate the policies automatically.
And it would work without any runtime overhead, and which is true, absolutely, and very popular for certain cloud-native data analytics platforms especially. The only limitation is that it would only work for that limited number of platforms.
Again, it might be some of the AWS-provided engines. It might be Databricks or Snowflake or something else, but it will not work with, say, Oracle or MySQL or anything else. And obviously, it will not work with non-database engines. So there are some clear limitations here as well.
Finally, there is the inline enforcement approach. It would have to be in the line, in the access path, so it has to be a proxy, a gateway, a connector or agent or whatever they call it. But it will be able to intercept any data requests in real time and make a decision in real time and definitely block it if something is not looking right. This is the only solution that actually works all the time.
But again, it would have a lot of downsides with regards to latency overhead, integration costs, and of course, you have to design for high availability. Fortunately, you don't have to choose one because most products actually combine all three approaches. And the probably most interesting question to ask for the vendor is, tell me where exactly each of those approaches apply and what are the gaps between them? Or is it something your company is doing as well? Can they combine?
Yes, I think we believe in not just one mode of protection because as you rightly said, the enterprises believe in protecting top tier SaaS applications which are required for day-to-day productivity work and that is best monitored and applied protection via the API because you do not need to monitor them in real time. Whereas where data happens to move more freely into the open world of internet must have inline enforcement, especially with data being accessed by conversational AI these days.
You must have a mix of both those strategies and having a centralized location in which vendors can help consumers manage these types of enforcement policies in one place so that the core aspect of how the data is enforced, your data protection rule for an API-based cloud service must not be different to how you want to apply it for an inline. So how can a platform simplify all of that from within the same console so that we believe in not forcing our customers to go through four or five different consoles and managing these things differently?
And that's how we can make life easier for consumers. Okay, moving on. So I would probably be the first one to tell you take these specific numbers with a grain of salt as well because I cannot claim we analyzed every solution out there. But as you will see in a moment, I think we have a pretty good representation of different vendors, large and small, specialized and broad. But one thing we can definitely highlight is that discovery and classification of sensitive data is mostly a solved problem. There is no need to look for the best data classification tool because they are all pretty good.
Also, the vast majority of those solutions we've covered does offer fine-grained policy-based access control. Unfortunately, these numbers are not fully overlapping. So there are solutions which can only do one of those things, which means that you probably have to combine more than one. Or at least you would have to look much deeper to understand where those capabilities end and how they overlap and how they cover all the existing gaps, if you will. More sophisticated things are still living a lot to be desired, if you will.
So as I mentioned earlier, just about half of the solutions can actually remediate findings automatically. And it can be a separate, very heated and almost religious discussion whether customers actually want to have their findings remediated automatically or should they be afraid of something to break or whether everything has to be involving a human in the loop, if you will.
But again, this is the future. This is the age of AI. You will probably have to at least start thinking about it. And of course, there is a lot of technical peculiarities which you have to keep in mind if you actually want to design a properly secure solution. For example, step-up authentication. If something changes within an established session, can that security solution identify that and force a repeated authentication, for example? What if somebody has stolen your session cookies or tokens or whatnot? This is no longer a technical question.
This is a very popular attack vector for a lot of data breaches, if you will. And of course, only half or even less than half of the solutions can actually cover that crypto inventory question.
And again, the crypto inventory is the foundation for your entire post-quantum strategy, regardless whether you only need it for compliance reasons or you're actually seriously thinking about data protection from those risks. You have to have it, and you have to specifically be looking for those solutions to support it. And so on and so forth.
Again, everybody is now talking about quote-unquote AI security, but that's a very different understanding of what it actually means. Are we securing data from AI? Are we securing AI from malicious data? Are we covering the inference stage or the model trading stage or both? And so on and so forth. I guess my biggest takeaway from this slide would be, again, do not trust anything. Do not trust any label, any statistic, any number. Look further than that. Ask questions and assess your own risks.
Harry, how do you see those numbers? They are, I think, very accurate, too, in terms of where these trends are coming up and how we look at the industries and multiple different places we hear these from. Most researchers are matching what you're showing, and it exists with them, so I agree with them.
Okay, great. So, again, just to reiterate, stop looking at labels, because the same label can mean very different product capabilities.
Partially, this comes from very different lineage of those solutions, because, again, what we observed is that the market is growing. A lot of vendors are getting in to get their own slice of the AI security pie, if you will. They use the same labels, but the actual capabilities below them are very different.
So, you have to map products to your own requirements. You have to look for specific capabilities to address your risks. This is the only way. This is exactly what our leadership compasses are supposed to help you with.
So, what is a leadership compass? This is a document Kupinger Co-Analyst publishes on a number of topics within the identity and cybersecurity markets, and I just recently published another version, another update of the data security platforms one. I've been doing this for almost a decade now. The scope has evolved. It used to be called database security. Now it's data security platforms, but the foundation is the same.
Looking for the vendors who claim to be able to secure your sensitive data, understanding what they're actually doing through vendor briefings, technical demonstrations, and questionnaires, which we send to every participant. We evaluate all the information using our proprietary methodology.
We let the vendors review our write-ups, and finally, we publish our findings, and it usually takes around six months of work for the analyst, so I am pretty proud of the results I have, and this specific leadership compass covers eight different capabilities, which we were specifically looking for in a solution which we believe has to implement if they want to call themselves a data security platform. It doesn't have to implement all eight, but the more they can, the better.
We do understand that there are also very specialized solutions out there, and not every use case, not every potential scenario actually requires you to look for all eight of those capabilities for your RFP, but still, this is a useful foundation for your process, if you will. So, we're looking for data discovery classification. This is a must-have.
We're looking for, regardless of how I describe the term, solutions that do data security posture management, data protection, and crypto agility, because those are very closely related, dynamic access governance, meaning who can access which data and why, threat detection and prevention, including, of course, real-time protection, for malicious access, life cycle and runtime protection for the data, regardless where it actually resides and where it will be residing later in its life cycle, if you will, auditability and compliance, for obvious reasons, and finally, again, because everybody's asking for it, data security for AI systems.
It does not matter, it doesn't mean that we only limit AI systems for this one specific category, because obviously, it applies to every one of those. We still need classification and data masking and threat detection for security data in AI systems, but we were looking for vendor claims around those as well.
We have 21 companies rated in the specific report, meaning that we actually have a lot of numbers, hard numbers to put behind the ratings, and we covered 23 additional companies which, for different reasons, could not participate, but at least we believe that they were important enough for us to mention, and you will find all those, of course, in the actual Leadership Compass. The link will be included in the later slide.
So, yes, this is a crowded market. It's a rapidly growing market, and it has not yet settled into a stable shape. You can see that there are so many different heritages to observe from this list. The company ages and sizes and their historical developments from traditional database vendors to traditional, if you may call them so, cloud native vendors, and so on, and, of course, Sky High Security was also participating. This is why we have Harry today to maybe kind of give his view on this whole process as well.
Harry, I don't know if you're personally involved in the process of the Leadership Compass, but at least what's your feeling about being placed into such an interesting crowd? I think, first of all, the Compass in itself, how you designed it, Alexia, must credit the way you have designed it because it covers a totality of what a data security solution must look like, right? So what can enterprises evaluate when choosing a proper security vendor?
And that kind of also helps vendors like us to shape and tune our platforms to meet with the expectation of what industry is actually expecting to solve this problem, because there are many different ways in which data security can be approached to, and this kind of is a solid validation for what we do and how we solve the problems for the customers, and your rating of what Sky High has been able to achieve and how we can show that value to our customers is through edification of our investment into security and our promise to our customers to secure data, both at rest and motion, and especially for the age of AI, we really value your feedback.
So thanks for that. Right.
So yeah, I guess this is probably the most interesting slide of them all. We have identified eight overall leaders, and in our definition, overall leader is the company which achieved three different successful evaluations being a product leader, meaning delivering on those capabilities we identified, an innovation leader, meaning doing something in a unique and groundbreaking way so that they stand out of the crowd, if you will, from the crowd, and finally, having a strong market presence, meaning a lot of customers and strong revenue numbers.
As you can see, the overall leaders are in the top right corner, and we have Thales, Oracle, Sky High Security, BigID, Secupy, Mage Data, Capital One Software, and Centra. Again, my point here would be they are coming from very different heritage, serving not necessarily very different, but kind of definitely not the same potential customer personas, if you will.
We have a mixture of long-established veterans next to companies founded really within the last decade, and unfortunately, or fortunately, you cannot just say, okay, I will take the top rightmost vendor, and this will be the best fit for my scenarios. No, unfortunately, that's not how it works. It will give you interesting candidates, but I would want to emphasize that you have to be looking deeper. You have to actually read the whole report.
I highly recommend it, and see which capabilities actually map to your specific requirements and which vendor can address those capabilities, again, for your requirements, for your technology footprints, for your industry, and so on and so forth. So, Harry, how does it feel to be in the top right corner? I think we are very privileged, and that is an honor to be on this campus at that level, and thanks for the assessment.
All right, and again, I have to emphasize that every vendor gets a detailed write-up with all the strengths and challenges and capabilities highlighted. They also measure individual scores, and we show them on the SPIDER chart, so you can instantly see where a specific solution is doing well and where it still has growth opportunities.
Again, I'm using SkyHigh Security for obvious reasons as a sample, but the other 20 vendors we have covered are represented in the same way. So, without further ado, the leadership compass next, what can you actually use those findings for, if you want to start next Monday? Before you actually start talking to a specific vendor, what do you have to prepare for? You have to obviously know where your data resides, not just your production data, but the copies, the archives, the test environment, and so on. You have to ask the vendor what the tool can actually fix, not just what it can find.
So, remember the holes in the fire. You have to ensure that the solution can actually enforce those access policies per request and reside as close to data as possible, and you have to make sure that it does know when an agent is acting on the user's behalf and whose users it actually was, and what access rights it delegated to the agent. You should not ignore the cryptographic deadlines, and of course, again, kind of read past the leaders, read past the labels, look for specific capabilities, look for specific limitations, and do not ignore the roadmap items as well.
Harry, how would you respond to a potential customer asking these questions? So, we look at each one of their requirements from an RFP standpoint as to, you know, how an enterprise is looking to move forward. Is it a regulatory compliance requirement or securing an AI?
So, we look at every customer from that lens, and then we try and help provide a curated solution that, you know, works for them rather than take our solution and make it work in a particular way. So, it absolutely depends from what an organization's overall goal is. Right. Okay.
So, we kind of promised a panel discussion, but we technically already started one. I think we had a really interesting discussion so far, but let's continue until the questions we have identified in advance. One thing which everybody who knows high security, you do focus on data in motion, if you will. This is your lineage, your history, and this is actually a really nice and really important question. Yeah.
So, why is everybody talking about data discovery at rest and data classification and protection at rest if it's actually mostly in motion when the data is being attacked? And, as we know, the data only actually generates value when it's in motion and transformation, because when it's just sitting on a disk, it doesn't do anything useful. Does it mean that we have to completely redefine the whole notion of data or are we still doing something useful?
No, I think the short answer is no, because that's the wrong place to stop, right? So, detecting everything that is actually sitting, residing in your environment is a good place to start, but that's not the place to stop, right?
So, the way we look at it is an inventory tells you what you have. It doesn't necessarily tell you what is actually leaving, and in the field, the exposures that actually hurt most enterprises is actually where there is data movement, and that's the most important part, which is why we believe in looking at both the data in movement and at rest.
So, to give you a real example, for example, when someone runs a discovery, let's say in a financial institution, even looking at 100 terabytes of data and we find usual things like the ID numbers, card data, payment information, the usual stuff, right? So, usually these things are not alarming when you find these things in data at rest, but what's alarming was in one of our example financial institutions that I was talking to, some of their employees were uploading all of this payment information into a conversational AI service, and that's the real risk, right?
So, how you take that and map it to a risk is only possible if you monitor data actually which is in motion and not just discovering data which is at rest. So, we believe in tying the two together to provide the greater of the value.
So, the sequence I would flip is, you know, by asking those two questions, what would it hurt for us to hurt the most, right? When the data actually is residing or when is it actually shared with someone outside, or you look at all the egress points that you already have, be it web, SaaS, email, or now AI, and that's what makes it really, really important to understand how the data actually moves.
Okay, great. To this, I can only add maybe a few words. Some of the most critical decisions about your risk, your data risk, cannot be made by just knowing where the data is now. You have to at least know where it was before, even better if you can predict where it will be in the future.
And, yes, this can only be identified when you combine both data in motion and data at rest into some kind of a data lineage graph, if you will, for the lack of a better term. Great.
Of course, again, kind of everybody is now talking about adopting as much AI as possible, much faster than any sensible approval process can actually review them. So, how do we, how should we all deal with shadow AI services? Do we block them? Do we try to put as many guardrails around them?
Or, is there any way to actually secure the content, the data itself? I strongly feel that blocking still has a place, Alexey, but it's become more like a backstop and not a strategy, right?
So, you began blocking every single application. So, two things that usually break that model of app approval, right?
So, first is the volume of the applications that are growing. There are thousands of AI services and there is every week, there is a new service that is being created in terms of new AI application that takes birth.
So, no review board can keep up with that kind of a pace in an app approval model to say that I'm going to block this application and this application moving forward, right? And second one is that people mostly underestimate AI is now inside some of the applications that you have pre-approved in the past, like your CRM applications or your meeting tools, Zoom, for example, or Google Chat. All of these things have an AI embedded in them.
So, application blocking doesn't really matter a lot, which is why looking at the data in context and understanding what is personal data and how is it actually being handled by some of these applications is the right way in which we should be looking at it, which is why it's really important to understand where is all your sensitive data, who is using that and how is it being uploaded to AI services most importantly, and then bring in that aspect of blocking which will make a lot more sense than actually second-guessing what should I block. That's how I would look at it.
Again, myself, I've always been a huge fan of that whitelist-only model for security. The only problem is that it usually doesn't work. If you would only say, okay, we will block everything, it would only whitelist very few specific services, it will break your entire business really soon. You will kind of dig your own grave, if you will, for your entire business. Unfortunately, we have to find compromises, we have to build defense in depth, we have to combine all those existing and new security controls too. Next question.
How should the buyer decide where enforcement must be in line and what performance cost is reasonable for it? We talked about those three alternatives and I said, basically, the enforcement is the only one that actually works all the time, but is it actually feasible to implement it all the time? To answer this question, I would encourage anyone who is asking a question first. If a certain event happens and I find out 10 minutes later, can I still undo that? Ask yourself this question. If the answer to that question is yes, you probably don't need inline.
For example, you are uploading a sensitive file into a sanctioned cloud storage and share it with a public link with someone outside. An API-based control can find it and your near real-time protection or scanning can suffice and work with that with no impact on anybody browsing and there is no impact on latency as well.
However, if the answer to that question that initially we asked is a no, it has to be inline. For example, anybody pasting sensitive data into an external AI service and uploading the file into a personal storage account. So these are places where data once gone, it's gone. There is no API that can bring you that data back. So in those situations, inline protection is a must. And on the performance, again, the rule is simple. The right cost is one where the user doesn't notice any difference. For example, anybody who is browsing the internet, the bar is very, very low.
For a very large file being uploaded, people can tolerate a very short pause. But whereas when they are interacting with AI prompts, it's even more trickier because people iterate very quickly with prompts. You ask a question, you then go and iterate over.
So again, the latency expectation there gets noticed really, really fast. So the practical advice that I would give is probably to say that test it with your own traffic from your own locations and then look at the slowest 5%, not the average speed at which your people are looking at. So identify the slowest basis where the latency happens to be the most and prioritize those places and inspect them and decide who needs to do what. Where can this be applied first?
And especially, to quote even more, in places like Europe, there are agreements where you would have to agree with your own teams to know what exemptions that you must be applying to respect those rules. And these deployments that skip step and end that exception list becomes longer and longer. So it has to be a judgment call and do it with the right approach. Like I said, depending on the traffic that is being looked at, you would have to apply controls either that in near real time or real time.
Okay, great. So, oh, that's an interesting one. It's probably a tricky one. So how does a CISO or anybody else responsible for actually deploying those data security platforms prove that they actually reduce risk, that they actually generate return on investment and not just make them work more to analyze those findings?
Okay, so this is a great question, because I often hear the same thing from quite a lot of CISOs consulting saying, how should I present this to the board? So if I were to be one, first of all, I would stop leading with the findings report, right? So you can find quite a lot of things from your security tool from a discovery point of view. But what a board actually sees, you know, counts on the things that which read as, you know, things are getting worse, or this happened here and here.
What actually really, really works is so board really cares about the coverage of your security tool, the exposure and the speed, right? So when it comes to coverage, you will have to answer the question on how much of our important data can we actually see in terms of my coverage, and in terms of exposure, how much sensitive data is reachable by people who shouldn't have access to it? Or is it publicly shared versus externally shared, sitting in an AI tool, where somebody can build an AI model? So these are the kind of exposure questions that you would have to answer.
And when it comes to speed, when we find something, how long does it take to fix that? So those are the kind of metrics that a CISO must be concerned about, where if the coverage goes up, for example, while the exposure comes down, that is your risk reduction story, right? And that's very, very hard to argue with, then you know, something like a stronger version that I would present to the board will be, we rolled out an enterprise AI assistant to the whole company, and we did it safely, because we fixed oversharing first, right? And that's not actually a cost, that's enabling the business.
And that's how I would present my, you know, findings of security, or how I use security to enable something, rather than just presenting a bunch of metrics to the board. Right. And there's a great point about that coverage thing. That's actually one of the questions I asked every vendor in every briefing. And unfortunately, quite a lot, they cannot even give you a dashboard for that, to say nothing about something more sophisticated.
So yeah, this is definitely something which a potential customer has to be really pushing the vendor about, because this would be probably one of the key metrics they could show back to their own board. Absolutely.
Okay, final question for today, at least for the panel part. So what can we predict for the next 18 months? And I've shown you that number, because this is when our next Leadership Compass will be published. So what should we be looking for, for stuff that will be new in the upcoming Leadership Compass? From my perspective, a real big one to think about is the notion of what a user meant in the data security space, because the concept of a user in data security now stops being a human or an actual physical person, right? Almost every security control built so far in the history assumes a human.
Someone clicks, upload on a website, someone shares a file, and then we can coach them. From a security principle or a policy perspective, we can go and enforce a policy, and we can coach that user from not doing something right. Over the next 18 months, I would say a growing share of data movement will be done by AI agents acting on someone else's behalf. It could be something like reading your inbox, pulling data from your CRM solutions, or calling some of the servers, writing a summary somewhere else, right?
And these things can happen machine to machine, all at machine speed, with nobody even able to see or send a coaching message to these agents, because they don't care about them. And that raises the questions for most teams haven't even started on, right? What identity is the agent actually using? Is it assuming the identity of the owner who created the agent, right? If it can read everything that the owner had access to, the agent is the most effective exfiltration tool that you have ever built and deployed. You must be looking at it like that, right?
Any agent that you have created is of highest risk if it is not guarded, and there are no guardrails to go with it. And lastly, beyond the agent AI, one that I would think is the derivative data protection. It's not just that data loss in the days that we have looked at until now, primarily because derivative data means something like your AI produces summaries, transcripts, embeddings, and vector stores and things like that. They carry a lot more sensitive content, and there is not a label or a permission that can be attributed to these things.
So understanding the context of what the data comes from, and the intent behind what comes out of these AI applications is also important. And lastly, I would say having a proof is the most important thing, because the regulations are growing. Auditors and, you know, GRCs are becoming a lot more important in every enterprise. So customers will increasingly ask, you know, what data went into this particular AI service, and can you show me? And these things are really important in Europe, especially with the AI Act that's coming up, and your DORA has already kicked in.
So everything is pushing your necessity in those directions. So practical advice would be to inventory your agents, get to know where is your data, and how your agents are interacting and using that data, and fix anything that is oversharing right now. Prioritize it before an AI attacks and, you know, takes control of your data. That's my perspective, Alexey. I think that was a great summary, and to this I could probably add one last point I am really trying to push in my recent publications. We used to enjoy this kind of dubious safety of security by obscurity for a lot of time.
People said we were too small, we were too insignificant to be hacked. Well, now this is no longer the case. Everybody who can be hacked will be hacked. However small or even however benign or insignificant they think of themselves, simply because there are so many agents out there who just kind of work like a bull in a china shop, they will break everything just by touching everything. No matter how obscure or small you have a vulnerability or unprotected interface or an insecure password or cryptography, it will be exploited. And this is unfortunately the new reality we have to adjust for.
And this is why data security has become so much more important almost overnight. And this is why it's a topic you have to at least follow up much more closely and, of course, consider the right solutions for your company. Thank you very much for being with us. Our webinar is over. Let's hope we will see you in some of the Coupling Calls future webinars or events. If you have any open questions, feel free to contact me or Harry. We can definitely forward your questions to him if you need.
Thank you, Harry, for being with me today. Thank you. Thank you for having me, Alexei. Thank you. It was a pleasure. Thank you.
See All Locations
See All Locations