Privileged access management was built on a simple assumption: a human logs in, does something sensitive, and logs out. That assumption is breaking down. Service accounts, automation scripts, and now AI agents hold standing privilege around the clock, request access on their own schedule, and rarely get the scrutiny a human administrator would. AI is only raising the bar: attackers can now discover credentials, map identity relationships, and escalate privileges autonomously, turning any one of these overlooked accounts into a path to critical systems faster than traditional workflows can respond.
Non-human identities already outnumber people in most environments, and privileged access is where that gap matters most. It is the layer attackers and, increasingly, autonomous agents actually try to reach. Closing it is not a separate initiative from your existing PAM program. It's the next stage of it.
Alejandro Leal, Senior Analyst at KuppingerCole, opens with the independent research view: why privileged access programs built for human behavior create blind spots once machines and agents are doing the requesting, and what a privileged access model built for both actually needs to account for.
Rob Ainscough, Chief Identity Security Advisor – EMEA at Silverfort, follows with the practitioner's view: how organizations are discovering and governing privileged machine access today, where vaulting and just-in-time access hold up under agentic AI, and where they need to be rethought.
Who Should Attend
Ideal for CISOs, security leaders, IAM and PAM professionals, identity architects, and IT security teams looking to secure privileged access and stay ahead of risks introduced by AI agents and machine identities.
Hello, everyone. Welcome to the webinar, Rethinking Privileged Access for the Age of AI-Powered Attacks. My name is Alejandro Leal. I'm a Lead Analyst at KuppingerCole, and today with me, Rob. How are you, Rob?
Very good, very good. Great to be here. Thanks for having us on. Awesome. So Rob is the Chief Identity Security Advisor at Silverfort, and he will be joining me today to discuss this quite relevant topic. So at the very beginning, I'll be talking for the first 10 to 15 minutes, and then Rob will jump in, and then we'll continue the conversation from there. But before we begin the presentation, just a few things to keep in mind for the audience. Some housekeeping. So all of you are muted centrally. There's no need to worry about your microphone.
We may run a few polls along the way and discuss the results during the Q&A section in the last 10 to 15 minutes of the webinar. So please send your questions at any time using the Livestorm control panel. We'll go to them at the end. And the session is being recorded. The recording and the slides will be made available for download in the coming days. So with that, let's dive in. But before that, the first poll question.
And some of you have probably seen it on the news with Black October, and this week Google Threat Intelligence Research reported a threshold being crossed in use of agents by threat actors. Just take a look at this poll. It will be visible on your screen, and take your time to answer this, but I will continue for the sake of time. So with my section, I wanted to begin first with one idea that, from my perspective, frames everything we'll talk about today. We've spent years securing human users, and now we need to secure what acts on their behalf. And that's the shift that we see.
Service accounts, scripts, workloads, and now AI agents are all doing work on our behalf. And very often with privileges, we've never give a person without, you know, asking a lot of questions. So how did we get here? I think it's important to look at the history a little bit, just to understand what led us to this moment. So we went from human to non-human privilege.
From, you know, you could say that privilege access management, PAM, has gone through roughly four different stages. First, it started in a human-centric era. So PAM was about a relatively small group of admins and share accounts. To some extent, credentials were static, privileges were long-lived, and the main controls were bolting and session recording.
In a way, lock the password away and watch what people do with it. Then came the cloud and automation era. APIs and scripts expanded the privilege access surface, and service accounts multiplied. But governance stayed largely human-centric. So we were still thinking about people while more and more of the access belonged to machines. And that brings us to the DevOps and machine identity era, where we've seen an explosion of non-human identities. Secrets management became central to PAM, and we started moving toward just-in-time access and ephemeral credentials.
But visibility and lifecycle governance for these identities is still limited in most organizations. And now we are entering what I call the autonomous era. AI agents act with delegated privileges. They make decisions, and they do it continuously. That calls for continuous behavioral-based access control, automated detection and response, and full lifecycle governance for identities that, in a sense, act on their own. The key point is that each era didn't replace the previous one, but it added to it. And that's a topic that Rob and I will continue talking about in this webinar.
You could say that most organizations today are living in all four at the same time. So if we look at capabilities today, what does that mean for what we expect from PAM? In the past, as I hinted in the previous slide, PAM meant password vaulting, rotation, share account management, and session recording and monitoring. These are solid capabilities, and they still matter, especially for small and medium enterprises.
But today, the expectations have grown. Just-in-time access, secret management, cloud entitlement management, KIEM, identity discovery and visibility, and machine identity security are features that most vendors today are trying to provide and most organizations expect to get. And looking ahead, we expect a zero-standing privilege, dynamic privilege enforcement, governance of AI agent privileges, intent-aware authorization, which means understanding not just who is asking, but why, and continuous risk-adaptive access. And that's why I like the line at the top of this slide.
PAM was never about vaults. It was always about protecting what matters. So the vault was just the best tool we had at the time. And from most organizations today, in the research that I cover and when I talk to PAM vendors or end-user organizations, standing privilege is still a problem. Most privilege access is still always on, right? It exists whether anyone is using it or not. And that matters for three reasons.
First, attackers log in. They don't break in. We hear this a lot. But with valid credentials and standing rights, an attacker can move through your environment without tripping a single alarm. And from the system's point of view, nothing unusual is happening.
Second, access outlives its purpose. Roles, tokens, and service accounts keep their rights long, and the task of the project never really ends. So privilege accumulates, and it very rarely gets cleaned up.
And third, machines need it for uptime. For non-human identities, standing access is often justified as an operational necessity. Some could make the argument that's true, but it's also exactly why this kind of access keeps growing and why the problem still remains there. So if you belong to an organization that belongs to the first era or you're still depending on your vault, you don't have to rip it.
Now, I know many of you listening have invested heavily in vault and session management. So you need to start to look beyond it. So what does that mean in practice? It means that PAM becomes a capability of the identity fabric and not a separate silo. So the way I frame it is keep at UNIFI. Keep vaulting and session management where they fit. For high-risk human admin access and tier zero systems, they still do a very good job. But you must add what the vault cannot cover on its own.
So JIT access, discovery of NHIs, and runtime protection, meaning controls that apply at the moment access is used, not just when it's granted. And third, UNIFI. Unify it all through an identity fabric with shared policy, context, and signals across your tools and across all identities, whether human, machine, or agent. So the idea is that PAM becomes a capability of the fabric, not a silo. You need to choose between keep everything as it is and rip it all out and start over. There's a middle ground. And for most organizations, that's the most realistic path.
Here's a picture of the coping and call identity fabric. I'm not going to walk you through every box, but I want to point out two things.
First, look at the left side right here. Alongside human identities, like workforce partners and customers, we now have NHIs. So devices, operational technology, workloads, cloud services. And at the bottom, autonomous agents and bots. They are part of the same model, not an afterthought. Second thing to look at is, in the middle, privilege access sits alongside everything else. Authentication, authorization, including just in time and privilege evaluation.
Analytics, ITDR, et cetera. So PAM, it's a tool. It's a tool, but it should serve as a shared capability. And that's why coping and call, we often talk about shared signals, about orchestration. So all of your different tools need to communicate with each other. That's what unify means in practice. Privilege access is not a separate world anymore. It's part of how you deliver identity services across the whole organization. So the next thing to talk about is, the story is changing quite fast. The next privileged user in your environment won't ask for access. You will already have it.
So when we look at AI agents as privileged actors, AI agents are not just another kind of service account. They're a different kind of privileged actor for four reasons. Machine speed, agents chain actions together in seconds. They review and detect processes that move at human speed. And human speed cannot keep up with that. Second thing is autonomous and steerable. They choose their next step, which is the point of them. It means that they can be misled through, let's say, prompt injection or poison data. An agent can be talking to doing something it shouldn't.
Another aspect is overprivileged by default. So to make agents useful quickly, they often get far more access than they actually need. And there's also unclear ownership. Is it security? The developer's job? Is it IT? The business or the business unit that deployed it? If nobody owns the agent, then nobody owns his access or is accountable for the risk. Which brings me to the next slide on how attacks now chain across identities.
This is one of the things that, if we go back to the slide of keep at unify, if you have a strong foundation in your PAM, then you are going to be more ready for the threats that we see today. Of course, things are moving very fast. Every week we see new attacks, we see new breaches, it seems like every week there's something new going on. And as you belong to an organization, you're a CISO or an identity security leader, it's hard to keep up. And I know Rob is going to talk more about this later on. But if we look at this slide, we see a new kind of attack path.
It might start with a human, let's say a phish or stolen credential. From there, the attacker moves to a machine identity, such as a service account or an API key or a token. And that leads to an AI agent with broad delegated access that it uses at speed. And from there to the crown jewels, your data, your tier zero systems, your business processes. Every hop uses valid access. None of it looks like a break-in or like a crisis. And that's exactly why traditional controls struggle to see it.
Now, on the next slide, I wanted to offer some practical recommendations. So what should you put in place before agent scale? There are four things.
One, treat every agent as a privileged identity. Two, use short-lived task-bound credentials.
Three, enforce at runtime. And that's crucial, not at review time.
And four, keep a human in the loop. Not for everything, but where the blast radius is large. Because humans, again, cannot really keep up with the machine speed. So it's important that human on loop is only in those transactions that are quite important for your organization.
So now, before giving the floor to Rob, where to focus now? To wrap it up, here's where I would focus. So from my research on PAM, we published a one of the main findings is that privilege access is expanding beyond PAM. As you can tell from my part of this webinar, the notion of privilege has expanded. It's not only about who holds a specific account, but it's more about what an identity, any identity, is capable of doing within an environment.
So first, adopt an identity fabric to reduce silos and fragmentation. Shift from static to dynamic privilege. You have to start treating agentic AI and NHIs as first class citizens, not side projects. Enforce least privilege continuously, not periodically, and design for hybrid cloud and machine-driven environments. If you take away just one thing from my part, let it be this. Organizations need to treat PAM as a strategic business control layer, not simply as an admin or technical security tool. It's about protecting what matters to the business, whoever or whatever is doing the accessing.
And with that, I'll hand over to Rob, who will show us how these attacks actually work, and he will share more insights. So Rob, over to you.
Perfect, thanks Alejandro. I could not agree more with that key takeaway and final statement.
You know, just before I share my slides, I think it's so important that we think of PAM not as a discipline of vaults. Often PAM gets kind of interchangeable with vaults as a technology solution from a long while ago, but it's really about protecting what matters, protecting the business, making sure that we've got what we need to protect the business and make sure we're resilient and we're ready.
And what AI threats are doing, and we'll talk about them just as I load my slides, really AI threats are bringing home what's important and bringing new requirements into it, into what we need from privileged access. So the scope is for sure expanding, the requirements are changing significantly. And a really interesting time for us to be talking about this.
So, you know, as we said in the poll, we've seen so much going on in the last few weeks around agentic threat. So this is something as a company that we've been talking about for a number of months now, you know, I think everyone knows about the development of this, but, you know, frameworks like Artex that we're seeing being talked about a lot at the moment, if you don't know anything about Artex, we'll talk a little bit about that in a while, but they're really expanding the scope of attacks, the sophistication of attacks, and they're changing what we need from PAM.
And that's where we'll start, because really what we should expect to see, regardless of how this plays out, and how the use of, you know, autonomous attacks by threat actors plays out, there's two things that really stand out for me.
So the first thing is, we should really expect to see more footholds more often in environments, you know, for months, since, you know, maybe since Mythos was released back earlier this year, there's been a lot of discussion, and almost all of the discussion has been about software vulnerabilities, about exploits, about zero days, about end days, that kind of stuff. But the reality is, from all of that, we're likely to see more footholds more often, right?
Any assumptions that we might have built in about the perimeter, when it comes to then our privilege access within that perimeter, I think those assumptions are being quickly eroded. And that's where it's really raising the bar and requirements for privilege access. The second is, and I've touched on it already, there was often a trade-off in the past between sophisticated attacks that could not go broad. So they'd be very focused, things like nation state attackers, where it's high value targets and very focused.
But then you've got kind of less sophisticated, scalable attacks that we've seen, more the kind of e-crime type attacks in the past. And that was really based on the complexity and the skill of the attacker. Now that's coming away, and now it's even being commoditized with frameworks like Artex, and there's others out there as well, where attacks can be both sophisticated and scalable.
So for example, I think it was Wiz, we're looking at Artex as a particular one, and they saw that it conducted something like 200 attacks, different organizations in two weeks span, and was quite successful in that. So all of this is putting pressure on privilege access management. So we're more exposed than we have been in the past, and there are more threats out there. And that's really magnifying what we call, what I call our identity debt, what our current posture is. It really magnifies it and brings it to life. And that's what we're going to talk about today.
We're going to talk about how we respond to this as identity leaders, as PAM leaders, because none of this is to kind of put us in a pessimistic place. It's not to kind of, you know, scare people or anything like that. But it's really to then focus on what we can do, because I was at a great event a couple of weeks back called the Identity Underground Summit. So it's a community that I'm part of. If you haven't heard of it, do check it out. It's a fantastic group of people, fantastic discussions there.
And really the feeling was, this is identity's moment as a whole to step up to being a real key security control, really coming close to that core of what the CISO cares about and what the business cares about, and really step into that role. So it's a real opportunity for us in identity, in PAM particularly, to do that. So the question is, you know, the requirements are raising, we're going to see more footholds, likely a higher volume of attacks and successful attacks at that. But is PAM ready for this new world? So as Alejandro said, you know, we've seen different eras of PAM.
Now, as I talk to companies around the world, a lot of them are stuck towards the left-hand side of those four eras that Alejandro discussed. And this is what we commonly see as we walk into client organisations is, you know, PAM overwhelmingly focused on tier zero and also regulatory scope systems. It's largely about human administrators rather than non-humans or agentic identities. And it's really, really focused on protecting accounts. It's quite binary, right? You either vaulted that account and you have control or you haven't done it yet and you're exposed.
And that's frequently what we see. We've got the kind of tiering pyramid here on screen.
And often, you know, beyond tier zero and into tier one, which, you know, I always say this is where the business runs, right? This is where the money's made. This is where things are made, things are sold, things are bought. That's where it happens.
You know, often these PAM programs mainly focused on vaulting, you know, that's where they really struggle to get that adoption and really struggle to get that comprehensive coverage, human, non-human, agentic identities as well. And what that means is those accounts are generally vulnerable. And what that then means is a knock-on effect of that because they're reliant just on usernames and passwords for protection. Then we're coming back to two key things that we're really relying on for our posture. So the first is permissions management. What are the scope of permissions? Is it appropriate?
Is it least privileged? And the second is how good is our credential hygiene, right? How good is our credential storage? Do we have them written down in scripts? Do we have them in SMB shares? Do we have them in SharePoint links? Do we have them all over the place? What's the strength of those credentials? So we fall back on these kind of weaker methods of protecting. But this is a really, really common place that I see companies get into. And that's really something that's not really ready for this new world. And I'm going to show you next why that is.
Because we've been working against kind of machine speed attacks and attackers and red teamers and things like that for a number of months now, probably about six, eight months since our first kind of engagement on that. And we've learned a lot from that about what those attacks look like, what the characteristics are. And I'm going to share some of them with you because following our work with our customers, we built something called Cypher. So this is an autonomous red team agent. It's effectively two things. So one is Cypher, the core orchestration agent. So this owns the plan of the attack.
So its job is to go and do red teaming. And this uses interchangeable models under the hood. So it uses commercial AI under the hood. We swap them out. It's got a number of tools it can use. If it needs another tool, it can go and get that and use that. So we don't constrain it very much at all. When we were developing Cypher, actually, we tried two different approaches, one where we were more prescriptive about how attacks should work in terms of kind of phases and approaches and methods that might work. And one where we just gave it a goal and a set of tools and said, go do it.
And we actually found that it was more effective when it was able to be more creative. It wasn't constrained to a plan. We weren't really giving any specific input about how to do it. We're just saying kind of find your way. So that's Cypher. That's the brains of the operation. And what it does, it spins up an agent fleet. So to do the jobs, to do the tasks that need doing, it then goes and executes those with these specialized sub agents. So they might be specialized in recon or credential discovery or lateral movement or privilege escalation. And that's very much how our tech works, right?
It's very, very similar to that. We built this a few months ago for us. Why did we build it?
Well, we built it to improve our product, to improve our understanding of how these things operate, what they do, the tactics they use and what the dynamics are. And we do that ultimately to improve, as I say, our product and also the usage of the product capabilities. So understanding what's effective in key controls. And we use this against a lab environment. It's a small corporate environment intended to be representative of what most companies look like on a small scale. And as I say, we built a UI over the top of Cypher to show you what this looks like.
So I'm going to go through a little bit of the learnings from using Cypher because I think these are really important. So we've got a video on the right hand side here. I'm not going to focus on that today. That's not the kind of main thing we'll talk about. But it is worth saying, if you do want to see kind of a full agentic attack in action, then you can go to softwalk.com. And there's a link off the homepage to book a meeting and see that in action just to see how these things work and understand them a bit better. I think that's really important.
But generally, what we see is these things operate very fast. So we give Cypher a goal and we give it a foothold in the network, but no credentials. And it operates really, really quickly. So the fastest run to domain dominance from a position of having no credential was three minutes, 27 seconds.
So very, very fast. And to do that and why it's so quick, it works in parallel. So it spins up generally on average over 60 agents in any given run, and it will run them in parallel as well. So on the right hand side in this UI, you'll see kind of lines representing the work and the sub-agents doing the work in the environment. And that's often in parallel.
Now, when we see this happen in the wild, so I think it was hugging face where it was over 1000 concurrent agents working in the environment. So that parallel nature is really, really important to this and makes it really, really hard to defend against because these things, as they operate, they tend to go fast, but not quietly. So they're generally generating a lot of telemetry, a lot of events, almost an overwhelming number of events to try and correlate, understand, and then respond to. So they're quick, they work in parallel, which underpins that speed.
But ultimately, what they're doing is they're exploiting or discovering, exploiting and also chaining, as Alejandro said, well understood identity vulnerability. So they're generally not doing anything too inventive, they're not doing anything new, it's not special. But the dynamics of how it does it is totally different.
So just in terms of our general findings from running this hundreds of times to test our product and understand how they work, some of the top methods used along quite far ahead was ASRET roasting, Kerber roasting of non-human accounts, service accounts was a big factor as well. There really does seem to be like a built in bias of these models towards NHIs and NHI compromise. So that's a really important one. But they also use kind of less visible, maybe things that we've got. So things like exploits for ADCS, and also SMB looting.
So if they can authenticate to an SMB share, and grab credentials out of it, they will try that as well. So it's really exposing a lot of what we've got in the environment, this identity debt, with, you know, old methods, but there's just new dynamics to it, it's much, much faster. And therefore, from a detection response perspective, very, very much more challenging. So what does all of this mean for privileged access management and where we need to go with it?
Well, I think, really, we're talking about a move from protecting accounts to protecting the business, right? Protecting what matters, I love that phrase, Alejandro, protecting what matters to the business. And it's really ultimately then comes down to resilience, because what I commonly find with privileged access programs is, you know, they're kind of brittle, they might rely on the assumption of, you know, a perimeter having integrity, they might rely on an assumption of credentials not being discovered and used.
And that is something that we really have to change, we have to build resilience, so we can operate despite disruption, and assuming failure along the way. And that's a really important part of this that we'll talk about more. So there's three important points that I think about here. And these are just foundational points, right? Not particularly new or novel, but they are really, really important. So the first of them, one of my colleagues in the US, a gentleman called Rob Larson, he had a phrase, he used to work at General Motors as a practitioner in their security team.
And he said, every account must have a plan. And I think that's so true, right? We can't distinguish between the risk of a human user, a non-human user, an agentic user anymore. And we can't differentiate between high risk accounts, like tier zero and regulatory accounts, and lower privilege accounts, because what we see from Cypher as we test it, is attacks kind of propagate through tier one, they propagate at the lower levels, and then get themselves to a position where they've compromised a credential to be able to escalate privilege to goal.
So really, every privilege account is a risk, and we don't want to leave any of them behind. The second one, as I said, NHIs are a real primary focus for agents. We've seen this in Cypher, we've seen it in the wild, we've seen it in our customer base. So we work with a Fortune 50 customer, and they were red teaming with Mythos, they had access to Mythos. And there was a point in the attack chain, in this red team exercise, where the attack could have chosen either to issue a certificate for a domain admin account, or an NHI with equivalent permissions.
It chose to do that for the NHI and not the domain admin that it could have issued a certificate for from ADCS. So it's really seemed to have an inbuilt bias, we think that's probably because of the training data on the internet saying that they're high privilege, under monitored, under protected, and credentials are generally everywhere. So real focus on NHIs here, they're really, really important to bring into scope of your PAM programme. And then blast radius control, this is the real evolution, because as I say, it's not just about protecting accounts, it's about protecting the business.
So making sure that we have protection in place or controls in place for accessing tier zeros, or critical business systems, or, you know, segregating one environment, one AD from another, or cloud from on-prem, or whatever it is, or pre-prod from prod, that's always a good one as well. So really putting in place guardrails for blast radius control. And that's really, I think, the next phase for privilege access management, because it's not about protecting accounts, it's about protecting the things that matter to the business, using identity as the vehicle to do that.
So resilience is really, really important. And the way we do that, as PAM professionals, as identity professionals, we're really often thought about kind of proactive protection, right? We're trying to reduce exposure, right? We're trying to take the number of unprotected accounts down and the number of protected accounts up. And that's really important. It continues to be really important. So in this space, you know, things like segmentation, blast radius control, geofencing of NHIs.
So, you know, despite their permission, which might be broad, we only allow them to do what they normally do on a regular basis. We take a behavioral baseline and we say, hey, we're going to fence it to that. MFA and just-in-time, obviously, and also closing down vulnerabilities that might make them susceptible to compromise like Kerberosity. So that proactive side is really important, but I know a lot of practitioners out there feel like they're under time pressure, they're on the clock right now. So we have to also be reactively ready, right? That's important for resilience.
We have to assume controls failure. We really have to be ready to help from an identity side and a PAM side in particular, because if we are able to put in the right technologies and make the right choices in PAM, what we can provide our security teams with is response levers, right? So we've seen an evolution in the industry from something we call big red buttons, which is things like denying all authentications in an environment. So that's obviously the worst possible case to small red buttons now. So it might say things like deny all third parties from authenticating.
Or it might say, hey, you can't make the transition from cloud environments to on-prem environments unless certain conditions are met. So really using those response levers to be responsive to a very, very rapid threat that's very creative in how it runs attacks. So having possibilities for response is really, really important for me coming into that PAM scope. So it's not just about having those static controls that say, hey, accounts are protected. That's fantastic.
But we also want to be able to respond when the situation changes to things we maybe didn't plan for, expect or understand, which may be an agentic attack of will, because they're able to enumerate your whole environment at once and have that in that context window, understood in terms of vulnerabilities and ready to exploit. So on the right hand side, we want to be ready. And we're really seeing identity and security come together here.
So in the last four or five months, I can't tell you how many security teams I've talked to about identity, many more than I did before that period of time, because it really is forcing the two together. It really is a big opportunity for PAM professionals or identity professionals to bring these two worlds together and get the right investment to push forward in terms of protecting the business. In the middle, we've got this signal generation.
So if we're able to do this in runtime, if we're able to do this with full visibility and full context, and we'll talk a little bit about so forth works in the next couple of slides, then we can build signal. So for example, every proactive protection measure you put in place, if it's failed, if someone is denied access, because that condition is not met, well, that's a really good signal that something's happening that shouldn't.
If a service account is being used in an anomalous way and is denied access because it's anomalous, that's a great signal for the SOC to be able to go and investigate. And that's what they need right now. They need that signal to understand what's going on. So we call that contain to detect rather than detect and contain. We also think about things like tripwires and honey tokens.
So if we can, you know, if an account externally when enumerated has very high privileges and looks unprotected, but we know under the hood, we have policies in place to deny its usage, we can create a differential between how it looks and how it's protected, that then is effectively a tripwire, and again, can provide that signal for the SOC to go and respond.
And finally, Alejandro talked about this already, but really, we want to make this into a fabric, we want to share risk signals across identity, but also not just about the identity fabric, I think of it more as the security fabric that we're part of here, right? How can we share risk signals from identity to endpoint and network security products. So we get this dynamic, joined up defense. But overall, what we want to do is we want to be proactively ready.
But given the urgency of this, it might be more valuable for some people to think about how we can be reactively ready, how can we assume failure and be ready to respond and have the levers and possibilities to respond when we need to. So really, it's about both achieving PAM goals, and also security goals at the same time, really anchoring in what we care about protecting the business and building resilience. And so to do that, we need to think a little bit differently about some of the technology. And this is a bit of a left field comparison.
But I was talking to someone the other day about how Silverport works. And they said, Oh, well, it's somewhat like my car's collision avoidance system.
I said, Go on, tell me. And what he said was he got a new car. And he was driving down the road and someone in two cars in front of him suddenly braked. And his car automatically braked and told him that it was a pre-collision assistant.
Now, that's somewhat like how we work. So the first off to do that, what do you need, right? To achieve that level of control to prevent harm, what do you need? So the first is you need sensors. You need sensors to understand context, right? In the car terms, it's assessing the road. In identity terms, we need to understand what is going on in identity. We need context of who you are, where you're coming from, why you're doing that. Is that normal? Is it showing an IOC? We need those sensors. Then we need a risk engine. We need to understand the danger, the threat.
And that's done in the car, but it's also done by Silverthorpe. So Silverthorpe takes in this context of authentication in real time. It intercepts authentication data. Then we process it and we understand the risk, whether it's behavioral or indicators of compromise. We assess the risk of the authentication. And the final step is then having the right levers to be able to pull. And those happen in the moment, in real time. They don't happen after the fact. They're proactive, right? This is to prevent it happening. So in Silverthorpe terms, we're talking about risk-based levers, right?
We might want to alert the SOC, as we've said already. We might want to challenge someone with step-up MFA. We might want to restrict access because it's anomalous or block it entirely because it doesn't meet our standards. So we're really talking about inline, real-time protection that's really proactive. That's what we're trying to bring to this. So it's not static defenses, but they're dynamic defenses. So it's a very different technology choice, right? We talk about identity security. It works really well alongside those vault investments you've already made, as Alejandro said, right?
It's about keeping, extending, and unifying. That's really what this is about. So if you've already got a vault investment, it's great to team Silverthorpe with that investor. We'll talk about it a little bit later, to bring this together. So we've got our static defenses, our session management, and credential control, and rotation, and things like that. But we're layering on dynamic defenses from there to make sure that it's effective. What we often talk about is making sure that you're generating the outcomes you want from PAM.
Because PAM isn't about how you do it, but it's about the outcomes you generate in terms of protection and risk for what matters to your business. And this, for me, is the acid test. It's the key test of whether we're going to be ready for agentic threats and all they mean. Because they are fast, they operate in parallel. It's going to be really difficult for SOC teams to decide to take containment actions because they move so fast, right? And they generate so many events. We need to understand and then take appropriate containment action. It's really difficult.
So what we need from identity really is to think of it as a drawbridge. So I've talked about it like this for a few years now. I think it's the direction of travel that we need to go, particularly in PAM. But it just means that we're more dynamic. We're taking the dynamic context. We're applying policy, so we don't need to worry about onboarding and we can scale this.
But also, we have this risk signal sharing with the rest of the identity stack and also the rest of the security stack as well so that we can start to pull up the drawbridge when we see more threat, right? So maybe it's okay to authenticate as a third party when, you know, the threat level is low.
Maybe, as that threat level raises, maybe that's not okay anymore and maybe we want to restrict that. Or maybe we see specific kind of risk indicators and risk coming from a particular account at the point of authentication.
Well, we might want to then apply more stringent measures or even deny them if it's really important and we're not sure. So it's really about these dynamic defenses because these are genetic attacks. They're obviously fast, but they're also dynamic, right? They do unexpected things. They go broad in your environment very early on as opposed to a human attack, which is very narrow until late on.
So we really want our protection to be able to be dynamic, to be able to raise that drawbridge, to have the possibilities that we need to kind of push the button and take proactive containment action despite the fact this thing's moving really quick and we probably don't understand the full impact yet. So I think that's going to be really key and that's really driving identity and security teams together, right? With identity being a key lever for containment and control of agentic threats. And I think that's squarely in the PAM space in terms of what we need to do.
So how does Silverthorpe help? Well, what we commonly do, coming back to that pyramid we talked about in the second slide I showed, we commonly find that what we do is extend those controls, as Alejandro said, right? We extend the protection across all tiers, that real-time protection in line across all tiers, whether it's human, non-human, agentic, we don't care, we apply protection. The second is we want to control that movement and blast radius. So we want to segment the environment. We want to stop, for example, movement from tier two to tier one or tier one to tier zero.
We want to stop movement between environments or into critical systems. We want to set rule sets around that. We act as a firewall for authentication and so we're able to do that. And that is going to be really important as we go forward, because I don't think it's about being perfect. I don't think it's about trying to get to protect every account immediately, because it will take some time, but we need to be ready to respond and we need to control the blast radius of any possible issue, assuming failure. And ultimately, that means we build resilience.
We have the levers to respond if it goes wrong. And ultimately, as I say, it's about resilience. It's about not having a brittle position that if one control fails, then we're in a bad spot and then we've got chain vulnerabilities to go. But it's really about building resilience and robustness throughout identity. That's what we're trying to do here. So in terms of what you could do tomorrow morning, this is my final slide. So every account has a plan. I think that's such a clear, simple premise and really, really important, right? No account is low risk anymore is how I see it.
I've thought about that, thought that way for a long time, since I was a practitioner, but it really, really is true. The second one is NHIs can't be left behind. We can't forget about service accounts, AI agents, we can't think about all of those things, because they can't sit on the risk register, they can't sit on the backlog, they are a key target for these things. We want to segment that risk, we want to take control and be proactive in setting the rules of the game for how you can use identity in the privileged tier, how you can move, what are the conditions for that to happen?
And also, we need to assume compromise, right? We need to think things will fail, we want to plan for that, we want to have big red buttons, small red buttons, increasingly granular red buttons, so that we can balance the need to contain bad things happening with our ability to do that without impacting the business more widely than we need to.
So, having those response leaders that are appropriate and flexible and can be applied really, really quickly, that drawbridge can be raised, is going to be extremely important. So, I think that's my last slide.
So, with that, I'm going to pass back to Alejandro. Thank you, Rob. I really like the last point, assume compromise. It's very honest, it's very true, it's also something I tell vendors and in-use organisations, and I also really liked when you talked about, it's not only about the identity security fabric, but it's also about the security fabric.
You know, the SOC needs the identity data, the identity teams need the threat hunting, the threat intelligence from the SOC, that's when ITDR comes in. Maybe that's a conversation for another time. But I also really like the Cypher, what you guys are doing there. A little scary, but I'm looking forward to it. I'm really trying, Alejandro, not to scare people with it. We've been talking about this for a number of months with a lot of companies, and it's not about scaring, but it's about understanding, right?
It's about understanding that requirements are changing and trying to use that to understand what we need to do. What do we need to change? Where's the bar raising? What do I need to do differently?
So, I think it hopefully led us to a positive conversation about what we can do. Of course.
Thank you, Rob, for sharing all your insights. We have around maybe 10 minutes for Q&A, but before that, I will quickly share my screen just for the last poll question, and then I already see a couple questions in the chat.
So, again, you should see the poll question on your screen. What is your organization's biggest challenge in managing privilege access today? I will just quickly go forward. You can find more information on PAM, on the SOC, on ITDR, on our website, and I met Rob the first time back in May at EIC in Berlin.
Hopefully, I'll see you there again next year, but for you, those that are interested in learning more or attending the sessions and talking to us, feel free to check out our website, and yeah, I think it's time now for the Q&A. I will stop sharing here quickly. There's a question for me and one for you. The question for me is from Andrew. He's saying, Alejandro suggested a human should be looked in when the blast radius is large. How can you identify those instances in advance and quickly?
Well, that's a good question. Recently, my colleague, Martin Kupinger, he made a distinction. He said human on the loop, not human in the loop because his argument is that AI agents operate too quickly and at too large a scale for humans to approve every action. So human on the loop means humans provide oversight rather than being involved in every action an AI agent takes.
Of course, I don't want to have a discussion on terminology or etymology, but I think the point manually, I think the ideal scenario would be to build intelligence into PAM that can map privileged relationships and understand dependencies between systems and even predict the potential impact of privileged actions. That, from my perspective, allows us to identify high blast radius scenarios in advance and then bring humans into the decision-making process only when their judgment adds real value. But what do you think, Rob? Absolutely agree. There's a couple of bits to this.
So from our perspective, you really have to understand the identity. You have to understand its effective permissions, not just look at AD groups and guess whether they're high privilege or not. You've really got to understand that context of the user and the potential access. So I think that's the first thing that we really focus on. How do we understand that? How do we understand a high-risk user by access? Whether it's direct access, whether it's delegated privileges from things like Active Directory and high-level rights that you might have in there.
So really understanding that user is the first step to then apply that at runtime, apply that at the point of authentication. I think the other thing to say is in future, and particularly when we think about agentic, as you say, getting in the way of that is going to be very hard from a human perspective, given the pace and volume. But really what we're working towards is integrating AI decisioning into the product, so that we can really try and really separate the confident good from the confident bad, and also this middle ground.
But I do think we've got a good opportunity to then create a feedback loop into that data model to then really understand, as a person says, no, that's okay, that's fine. As the resource owner, that agent can access that thing. Then we can feed that back into the model, make better and better decisions over time. I think that's something we're investing in. I think it'll take a little while for us to mature into that kind of thinking, but I think it's probably the only way in the long term as agents roll out and things just get quicker.
Yeah, thank you, Rob. Two more questions. One is for you. How do you handle, I'm guessing Silverford, how do you handle agentic identities? I'm worried about potential compromise of agents using inherited permissions.
Yeah, it's a really good question. I think of this at a few different levels, really, that we think about. The first is thinking about the agentic infrastructure that these agents are developed in and run in. How do we make sure that if you're using something like Copilot Studio, we've got guardrails in place so that we can control the tool calls that that thing can make? We can maybe restrict it from writing email versus reading email, or maybe read only versus write, versus delete writes, so that we can start to put those guardrails in at the tool call level.
We can also do that via an MCP gateway, but then also I think what's really important, I think there is a bit of an over-rotation here. We focus a lot on how we strengthen the substrate, your infrastructure, that might be interacted with by agents. So thinking about Active Directory, for example, how do we make that ready for something that is fast and deterministic in its usage and likely using static inherited permissions?
And that's where some of the guardrails that I talked about today to kind of fence and restrict and deny conditionally access to that is really, really important because it's not just about setting guardrails at the agentic layer, write tool calls, MCP, those kind of things. That's important, but also we want to have really, really hard guardrails at the bottom end when it matters, when it's going to authenticate to that key system or data store, what conditions do we have in place at that point in time?
Because I was reading something yesterday and they're like, agents will do what they want to do, we have to accept that, and we've got to set really deterministic guardrails into the infrastructure, which is what we do, to prevent that happening and make sure we're sure about who's accessing what matters most. And again, comes back to that theme, what matters most and what guardrails do we have in place, whether it's human, non-human, or agentic, it doesn't matter, the same rules apply in my book.
Speaking of guardrails, you pretty much answered this next question, but I want to make sure that audience member Devika also gets an answer from you. The question is, when actions are done independently by AI with extremely high privileges, what are the guardrails that are in place?
Yeah, so I think just before we get to guardrails for a second, I think the first thing is understanding what permissions agents have and what agents have been built. I think that's the first step. That's something that a lot of folks are still struggling with. I see a lot of companies using multiple agent providers and agent boundaries to build these things. So bringing that all together, so we understand the permissioning relationship to the agent, right? How is this configured and what's the blast radius of that thing? That's kind of step one for us.
Step two from that is those guardrails that we've talked about. So it might be what tools can this thing call? So that could be based on the user calling the agent. So if we say someone in finance calling an agent might get a different answer to someone in HR and have different tool calls that are possible because of the permissions and what the agent's for. So we try and set that at the kind of tool call level and MCP level as well, but then ultimately it comes back to protecting the things agents already interact with, right? The systems, the data, right?
What guardrails are we putting in place there? What deterministic rules are we putting in place to prevent harm? You might want to say, hey, these are all finance system things and no agent is touching finance systems. So I'm going to say, hey, you must meet these conditions to access this resource. And just putting those really, really cool straight guardrails is a great start in dealing with what is a new and big evolving problem that's a big deal for many practitioners. Awesome.
Thank you, Rob. I know we're almost out of time. So I just want to take a quick look at the poll results just to maybe get, I think there's another question, but I don't know if we have time. If you have more questions, you can reach out to Rob after this webinar or reach out me if you have any more questions. But based on the poll question, what is your organization's, oh no, sorry, the first poll question was, are autonomous attacks affecting your PAM strategy and investments? The vast majority, 86% said no, not yet, but it will next year.
And the rest of the responses, no, it's not affecting our PAM plans and roadmap. I'm sure that will change next year. The other poll question, what is your organization's biggest challenge in managing privilege access today? Seems like it's kind of split, 38% lack of visibility into privilege accounts, 25% say that it's about legacy PAM tools that do not fit with cloud and DevOps, the other 25% securing non-human and agentic privilege, and then the rest integration with broader identity and security systems.
Well, thank you everyone for participating in these polls. Two more minutes, Rob, any last thoughts, any takeaway?
Yeah, I mean, it's an important topic, so timely, it's interesting to see that it's changing investment priorities for next year. And I think there are new requirements out there. It's important that we think really about what PAM's there for, what we need it for, and what's important in this new world. So as I say, we do have that attack lab that you can sign up for a session to see that live and see the whole attack. I only showed you a small amount of it.
So that's a really good session just to bring it to life and really understand it and have a great discussion around it and what can be done. So I recommend people doing that if they're interested. I will do that for sure.
Thank you, Rob, for your time. It was really nice talking to you again. And thank you everyone for your attention to this topic.
And again, if you have any questions left, reach out to us. Thank you everyone.
See All Locations
See All Locations