
Grounds for Access, Episode 3
The AI Enforcement Gap: 99.7% Have AI Policies, Only 51% Enforce Them
Hey everybody, welcome to the Grounds for Access podcast, where we pour a cup, usually it’s coffee. I know where Mike’s at. It’s a little bit ahead of me, so he might not have coffee in his cup anymore. But we talk about identity security. And it’s usually a conversation of no fluff, just security people talking about security. My name is John Martinez. I’m the technical evangelist here at Delinea. And today I’m very happy to have our guest, Mike Privette. Hey Mike, how’s it going?
Hey John, doing well, man. Thanks for having me on.
Yeah, of course. So Mike is a former CISO. And I don’t know if you really ever leave that job, but he’s a former CISO. And he’s the founder of Return on Security. Lots of great articles. I think one of my favorites, and I’ll tell you right now, Mike, is “A Tale of Two Planes.” That one really resonated with me being an old infrastructure guy. And he calls himself the cybersecurity economist. So we’ll dig into that one as well. That one’s very interesting to me. As well, budget planning and spending is an interesting take on security. I do have some thoughts on that as well. If you haven’t read his newsletter, I highly recommend you read it. One thing I noticed is that it’s zero fluff, and that’s exactly why we’re here. So Mike, welcome to the show.
Awesome. Yeah, thanks for having me on.
Is there anything you want to expand on, what your background is, what you’re doing these days, that kind of thing?
Yeah, so definitely echo your comments about you can’t really leave the profession. You can leave the role, but the role won’t leave you kind of thing. And so now I’ve kind of taken the saying that I’m a sometimes CISO: sometimes I do it, sometimes I don’t. And I’m actually in a phase now where I am doing it now again. So I’m back in the saddle.
Nice. That’s cool.
And so yeah, many of these things are near and dear to my heart now. But somewhere along the way I fell in love with what’s happening in the markets, public and private markets and data and numbers. Because I realized numbers are important and what actually helps things get done or not done in a business, especially in cyber. And so that’s kind of where I just launched onto the name Return on Security and just kind of ran with it from there.
The return on your investment, the return on your security investment. I absolutely love the name. And I love the fact if you haven’t checked it out, go check out his GitHub repo as well. There’s some really cool stuff in there that I really enjoyed thumbing through. So love that. So just really quickly, Mike, I know it’s way later where you are than where I’m at. I’m in California. I don’t know what’s in your cup, if you’re drinking anything right now. I can’t drink anything past three PM.
Yeah, so I’ve got some water here today, but I have had many coffees. So I’ve had two pour overs. I’ve had an espresso. I’ve had a tea. So like I’ve had enough time in the day to hit all the major food groups on caffeine. And so it’s been nice.
I love that. I’m an espresso guy myself, so I’m drinking something that I picked up over in Singapore. I was there recently. A bag of beans that I brought over. And it always trips me up. Being from the United States, we don’t know anything about the metric system. So when I walked into this coffee roaster, they asked me, do you want a 500 gram bag or a 250 gram bag? And my brain went for the bigger one, and I didn’t realize how big 500 grams is.
Right. Yeah, it can be quite big, but if it’s good, hey, it’s a win.
So before we dig in, just for everybody’s knowledge, we are gonna talk a little bit, and there’ll be some questions around the brand-new, hot-off-the-presses, Delinea-sponsored identity security report. So there’s gonna be some of that topic. And the big one here, and I’m sure it’s no surprise to you, Mike, is AI. The AI report is top of mind. And there’s some really good insights there that I do want to ask you about as you cruise through the report. Since you’re already back in the saddle, give us your view, coming back in. I’m sure you’re coming back into a time where everything’s about Hugging Face and this and that, and you’re getting slammed with those questions. Give us a little bit of the view from your seat.
It’s a very fast-moving time. I could say that it was not moving this fast when I left the seat a couple years ago. But it’s also probably one of the most exciting times, I would say, because so many more things are possible now than they used to be, both good and bad, I’ll say, because of AI. And the day-to-day has changed pretty significantly in terms of what people ask about and what you have to care about, but then also at the same turn, what you can do and like what your team can do.
It’s just kind of unbelievable. I was talking to a friend that I used to work with, and I said, “Hey, do you remember how long it took us to like map out this cloud environment and do all these things?” It was months and months of having conversations, interviewing people, looking at technical things and feeling your way through. Now it’s done in an hour. It’s a fundamentally different game, which kind of blows my mind.
In one sense, and this is what I tell people, the positives that I have with AI are, obviously, that it accelerates my work. I can do a lot more, very quickly. And the other part that I just mentioned is that it’s a force multiplier for my brain, in a way. So I do like that aspect of it.
To get to the report, we have seen the incidents happen. I think less about the incidents that have happened lately, which I know are top of mind for a lot of people who are watching today, and more about the report that we brought out. The thing that hit me really quickly was that, on the practitioner side of the industry, we tend to lean a lot on the internal policies that we’ve set up for organizations. No surprise to me, 99.7 percent, that’s the number that we quote there, have an AI security policy. But what about the formal data side of that policy? And this brings back the article that you wrote, “A Tale of Two Planes.” We have the control plane covered, but do we have the data plane covered? So let’s start with that. I want your take on that.
Yeah, I think it’s a changing world, in that security teams have been able to fall back on policies to give to the rest of the business, for regulatory reasons, and say, this is why we can or can’t do something. Unfortunately, in most of my experience, many of these policies are written in a negative light, saying, you just can’t do this. And so the person is left to interpret what they can do, which I’ve never been a fan of leaving unanswered. And I think you see some of this manifested in people not understanding that the data plane, and the data itself, is important.
After all, data is the static asset that we all need to use and that agents thrive on. And you don’t really think about the consequences until after the agents have already accessed it and done some things, worked some magic. So it’s kind of forgotten about. If you just think about one piece of that puzzle, like giving the agent an identity, which is important, you still have to do all of it.
And it’s not new. We’ve seen the reports from years before, where the primary target, once an attack gets launched on somebody, is the data. That’s where the crown jewels are. But here’s the other interesting stat on that: even though almost everybody now has an AI security policy, only 51 percent of those actually check against the policy.
Yeah, that sounds right. The world’s very much changed, where you can’t set a static anything anymore. You almost need to use the technology to police the technology, to see if you’re even coherent and sticking to what your policy actually says, and then, if not, adjust your policy or response. But no, that’s not surprising, because I think many organizations, especially this last year, have struggled to keep up in general with the rate of AI adoption driven by the business. So that makes it doubly hard.
And these policy violations are happening in real time. How does a SOC, for example, keep up with this?
I don’t think they can, honestly. I think this is part of the evolution that I’m seeing, being back in the seat again: the decision-making, or the ability to say yes or no, has to become much more federated out into the world, so that you don’t need to either wait for the SOC to find it or ask somebody over Slack, “Hey, is this okay?” Most people aren’t waiting at this point. The technology is too useful in too many ways and changing so quickly that I don’t think you can wait on the same kind of responsiveness that we used to.
And I think it’s probably a good evolution. Maybe we thought it worked better than it really did in practice, but we’re now seeing that it really doesn’t work at all anymore, and you probably don’t even know what’s happening if you’re not much closer to it than you used to be.
Yeah, and I have a couple of questions a little later about breaking the rules. But before we go there, another great piece that you wrote that really stuck out to me was, and I forget the exact title, but it’s having to do with creating a miserable security program. I absolutely love that piece. It’s yet another thing that resonated really well with me, man, I have to admit. The anti-pattern messaging, I love it. That’s awesome.
Yeah, that was modeled after Charlie Munger.
But to put it into context: many companies say they have a policy around securing AI, and one of your rules is around creating vague, circular policies and making them impossible to find. First, I’d love to hear how you came up with the rules and that rule in particular. It’s like policy by obscurity: if I didn’t see it, it didn’t happen. But if that many companies have a stated AI security policy, and fewer than five in ten of them can actually see what’s happening, what are the policies actually doing? Did we go down that path to your rule? Did we do that?
I’d say we probably didn’t do it on purpose. I think it’s very interesting watching the arc of how security groups have changed over time with AI introduction. Because I remember early on, in 2023, the name of the game was either “let’s just adopt it and see” or “dear God, block it at all costs, it’s a legal nightmare.” So basically we started in a reactionary stance from a security organization perspective, which meant that we had to do something.
So let’s write a policy. So now we’re legally covered, perhaps, and you’re subject to acceptable use policies in your organization. So don’t violate this or else. That’s an unfortunate way to manage security programs in general, but it’s even more irrelevant now. But I think part of it was that we had to do something, didn’t realize how fast it was moving, and then weren’t able to keep up with the number of things that were coming out. Then we probably never revisited it.
Like most policies, you don’t look at them until they’re due for review or someone asks for a policy exception. And then you go through that process, and you may or may not realize, “I don’t even know why we said that to begin with,” or “maybe we should change this.” But I assume most people want to do good on that front. I think it was probably just a casualty of how fast things are moving.
Absolutely. I think most people want to do the right thing, and I totally agree with that. At the same time, whenever I launch Claude Code, for example, it’s asking me for all these permissions and I really just start getting annoyed. More than likely, the inclination as a human operator of this AI agent is, “Yeah, next, next, next, next.” I’m just gonna say yes forever. Do the thing. Churn the code, whatever it is. And mostly it’s actually warning me, telling me, “Hey, I’m about to do this.
Are you sure you want me to do this?” But we do know that, if we let them, these AI agents can overreach when it comes to accessing critical resources and critical data in my environment. So I guess what I’m getting at is, based on having these policies, based on these documents that we write and the tests that we have in our environment, what’s your take on how we can tell that there was overreach, and what can we do about it?
I think that’s a really hard problem to find an answer for, but it’s kind of a symptom of problems that already existed. Like you said, you can’t blame AI for working around the guardrails, because that’s a normal human thing. We’ve just never seen it done this fast and this often. Over-provisioning of access has always been a problem. Privileged access sprawl has always been a problem. It’s been really hard to rightsize, because it requires you to know exactly what you need, exactly when you need it. I think that’s actually much harder than we realized, because it requires you to know the shape of the problem and exactly what you want to solve each time, which humans are not super great at.
So you say, well, give me a lot of stuff and then I’ll figure it out. But AI is actually really great at helping us figure out the exact permissions and whatnot. And that’s why it’ll say, “Are you sure that I can delete this from your home directory?” It’s one of those things that has just highlighted an issue that’s always been there. And now it’s an everybody problem, because it’s not just a tech problem. Going from the data center to the cloud was a tech problem, and then with SaaS we saw a bit of that evolution, where anybody with a credit card could go buy a service and we started seeing it outside of IT.
And now with AI, every single function uses it in many ways, right or wrong. It just highlights that big problem. How do you solve that? I think it’s very layered and very challenging to do.
No, and I get that. It’s something that, on the other side of my job, I demo all the time: we’re able to get you access, in a secure, just-in-time way, to a production database, for example. In one sense it’s a yes, there’s this whole trail of access management to get there. But on the other hand, we’re giving a business user potentially a lot of power to do business analytics against production databases in real time. That begs the question: are we thinking about this the right way, or are we setting ourselves up for some future headaches? I don’t know if you have an opinion on that, but it’s definitely something that’s there.
I think the problems are already here. I think we have the headaches now, and we’re just having to adapt to it, because we’re forced into this kind of impossible decision: do the right security things, but also let AI do its thing. That’s actually not necessarily a removal of security, but it’s a case of let the agent decide, let the agent know, with, in the ultimate case, very minimal human involvement on some of these things, unless it’s incredibly important. Otherwise, you just let it do its thing. That’s a totally different way of operating at work, a whole different way of doing business that I don’t think anybody’s ready for. And it all harps on the existing baggage that’s already there. So I think the problems are very much here, and here to stay for a while.
Yeah, and one other point on the report. Art Gilliland, our CEO, is quoted as stating that an unenforced policy is really just an honor system. It’s inherent trust. I want your take; I have mine, of course. But that kind of flies in the face of something we were talking about a year or two ago: zero trust. It’s another buzzword that has fizzled out in this whole AI discussion. I don’t hear it anymore, and I don’t know why. So what’s your take on that? Is it really an honor system at the end of the day?
I see where he’s coming from, especially in today’s world, where there is so much more tech involved, even in the basic processes of doing your job. There are different approaches to how people view policies. Whenever I come to a new company, I look at the policies and think, all right, what is explicitly not allowed, versus what is allowed? If it’s not explicitly denied, I assume it’s okay until someone tells me the opposite. I take that mentality on purpose so that I can try to poke holes and find issues where policy and reality don’t meet, or where how we think something works and how the tech is actually configured are different.
Because it very easily can be just an honor system now, especially since we have such helpful agents. They want to be helpful. They want to achieve the next thing and get those next tokens. So they will do what they have the ability to do, within their interpretation of your goal, your ask. I think that makes it extra challenging, because you don’t just have what I meant to do as a user, but also what my agents think I meant, and their sub-agents, and any delegation. It does make it hard. It makes sense, especially in the world we’re heading into.
When you said that, it reminded me of a discussion I was having at another employer a few years ago, before AI kind of blew up. We were already talking about AI, and what we used to say back then was that in a few years, and here we are, and I think it’s very true, everybody’s gonna be a prompt engineer. And while inference engines are getting really good, it’s still not like, “You told me to do X, so I’m doing X.” Maybe a little bit of Y.
We’re in a strange time, too, where you and I can say the same things and get a different outcome. One day that won’t be the case, and then you’ve got to wonder what that looks like for the rest of us, when we can just say “interpret this” and something comes out.
So that is important. And a little bit of what you said earlier: if you look at a policy and there’s not an explicit deny called out, you can assume that it’s an allow. And I think it’s not just a security thing, and it’s not just security practitioners that do that. It’s humans, too. I would posit that even when something is explicitly denied, “you’re not allowed to do this,” people think, I’m still gonna try it anyway. And then, all right, I tried it and it worked, so you obviously didn’t prevent me from doing that.
So, workarounds not being really a discipline problem but a design signal, I love that aspect of it, too. Knowing that I need an approval but bypassing it anyway, kind of like to my point. And I really like this about the report: 48 percent of people do it on a regular basis, because now I have this great, powerful tool that I can use to do really fast sales analytics and create a really nice report for my boss that says, “Here’s our pipeline, and here’s where we’re gonna miss or meet that pipeline.” And IT and legal are way too slow to approve my request for access anyway, which is the problem we saw during the cloud era, too. IT’s too slow, so, like you said, “I have a credit card, will cloud.”
Exactly. And now it’s the same thing with AI. I agree, but it’s going to require those groups to act fundamentally differently. I think that’s a call to action for them, at least I hope, instead of it being, “No, you can’t do that.” That ship has long sailed on that front.
You’ve written about when users bypass controls. Is it the twenty thousand tools that I bought that are not configured the right way? What is it? Is it a tool problem? Is it a hygiene problem? Is it an education problem? There are all sorts of angles here. With your CISO hat on, give us your take on when users bypass.
I think there are many things that could fit into that bucket, but mostly it’s an expectation mismatch. One side or the other had a different opinion, a different interpretation, or a different need, and they didn’t realize it. Then you find out that this group’s been running around this the whole time because maybe the other group was too slow to provision them access, or what they could provision wasn’t very good. And maybe it’s not a core function of what that team delivers anyway. Part of what you discover, and accept, is, “I don’t want to be the best at a CRM; let’s just pay Salesforce.” That’s a simpler way to do it than making your own.
But some of it has largely been on security teams to ask, what can we do to reduce friction and also increase security? We always think about that from the developer side. We’ve been playing that game for a long time, and I think we’re getting better at it on the whole, but it’s been a slow road. But largely, the rest of security is communicated at companies like this: you tell them things, you don’t ask them things, you don’t discover things. “Here’s what you need to know in between your busy day job of possibly making money for the company.” I think that’s a fundamental user experience failure, and it can lead to a lot of workarounds, or lead to you not being involved in the question.
So I’ve always treated the user experience side of security very highly, because it matters. The more you prioritize that, the more they want you involved to solve problems, instead of just saying yes or no. So it’s many things, but I think it comes down to the experience.
I totally agree with you. The minute you say the only way to gain access to, let’s say, a Linux box is to go through this weird process or some weird tool, and I’m forcing the path, I’m like, “But I’ve got a perfectly good terminal app on my Mac. I can just type in SSH. Why are you sending me on this route?” But I think part of that is the way we’ve been doing access approvals. A lot of it comes down to access. The last time I did IT and security, a long time ago, a lot of what our team spent time on operationally was approving access requests. I need access to this, I need to install this app on my laptop. It’s all access.
In the age of AI agents running around doing things for us, in the shiny future that we’re living in now, DevOps teams, I think, are the ones really latching on. “Yeah, I can do all these things with agents, I deploy everywhere, and they do my operations things for me in production.” It’s great. Are approvals really the right way to think about letting AI agents gain access to things anymore? Or have we outgrown that, or do we still have baggage, like we talked about, from the old days? What are we doing here? Is it even the right way to think about giving AI agents access to things?
I think it’s a really hard problem, and I think the industry may be taking an old-school approach to it. It is an old problem, but we’ve framed it differently with agents. That’s something I’ve been thinking about quite a bit: agents are using what they have access to. They’re using what they’re given. They’re using what they can connect to. They’re using what they can scan and find. That’s just the nature of them, which makes giving them privileged access very hard. It’s easy to say, “Just have every agent put in a 2FA code before they access something,” but that sounds good on paper and isn’t a real-world thing, especially considering how many actions they take per session.
Per prompt, they may do hundreds or thousands of things. So it’s hard to apply a human-based construct like, “You must SSH to one Linux server, please, and here’s the just-in-time password for that one box,” in a very painful checkout procedure, like the super old access systems.
There’s a lot to unpack, but it’s mostly about understanding that it’s a problem around the environment and the guardrails, making sure that the things that can be accessed from you and your system are sandboxed or isolated in a way that can’t go outside of what they should. But then it comes back to the over-provisioning problem and the knowledge of, “Well, I know the production database has all the things I need. I just don’t know which tables, views, or lookups I need, but the agent can figure it out, so just give the agent access to this.” Then it might do something bad or risky, or even just hammer it with too many calls and slow down your production services.
That’s a piece that makes it hard, because it requires you to know what the correct answer is. As humans, we often don’t know that until we see it. I can tell you what’s wrong, but I can’t tell you what’s right until I know what’s right.
Yeah, we’re really good at distinguishing those patterns. It just doesn’t look right, it doesn’t smell right, it doesn’t pass the smell test. What you just said is a great segue to my next area, which is standing privilege, which you touched on earlier in our conversation. We’re bringing that baggage along: Mike Privette, by virtue of his title and his department, is in security, so why wouldn’t I give you god-mode access to everything in the world? You can see it.
Yeah, you gotta trust somebody, I guess.
No, but even still, if we’re positing again that we want to make these agents that are working on my behalf attributable to the human, I want to know that John launches these agents and they’re doing things. Thankfully, these days I work in marketing, so I don’t have access to anything other than the things I do. I inherit all of the permissions that John has, that Mike has. The way I think about it is, instead of authorizing John and inheriting all of his permissions, what if, when it comes to an agent executing, I start authorizing the actual actions?
We’ve been hinting that AI agents are non-deterministic, and if you’re watching, hopefully you know that by now. They’re not like our old Python script that launches things for me, that goes straight down and does exactly what I tell it to do every day, every hour, whenever that cron job spins up. You mentioned that it could take a different path for Mike than it could for John, and in the end it could effectively do the same thing, but it’s going to take a very different path.
So what are your thoughts around that? I know that’s a very technical sort of answer to the problem, but why don’t we authorize the action instead? You can take whatever path you want, AI agent, but here are the things that we’re authorizing you to actually do.
Honestly, I think that’s a better answer than trying to give it some kind of fake identity, or an assumption of your own identity, which may be over-provisioned anyway. It’s more like, here are the things you’re authorized to do, in a positive way, and here are the things you’re not authorized to do, like an explicit deny: you can’t run this, you can’t delete, you can’t do destructive commands.
This is where you see protocols like MCP out there, some of which are read-only and are great interfaces. You get a level of access that is instructive and helps the agent get data without touching anything sensitive. So there are some good ways to do it, but it all comes back to: you can use this protocol to access this data. That’s the same authorization idea, that you can do these things. That requires different experiences than how a user might log in to these things and perform these functions, where they might use a web tool to look at business intelligence data, or use SSH to run some commands directly, or SQL statements and things like that. Giving the agent a better-authorized path, I think, is the way to go.
And obviously it’s going to be a world where we have to focus on this a lot more, because it’s going to be a hundred to a thousand agents per user, potentially, in some of these places, especially as token prices continue to fall. Economics comes back in on that one. I think we’ll see a smarter rightsizing of authorization as AI becomes more expensive for organizations to use. But it’s a very leapfrog game. An AI lab out of China releases an open-source model that is almost as good as the closed-source models, and then the cycle repeats every couple of weeks or months.
But you see some say, “You shouldn’t just use this for anything and everything. Let’s be smarter about it, and let’s also minimize our outages and our risks.” That comes back to needing to properly authorize these things, which is obviously access, but a more granular version of it.
No, for sure. Along with that granular level of access come session controls, controlling the session. In the report, we talk about how, of the people allowing agents to inherit permissions from their human counterpart, only 42 percent have the ability to revoke access, to revoke the session. I think that’s a problem, especially with the latency between when I find out that something bad happened and when I can act. I was at another startup many ages ago, an early CSPM company. Most customers back then were okay with receiving an alert within, I don’t know, a day maybe. Twenty-four hours is okay; I can go fix that SSH port 22 open to the world. Not immediately; it doesn’t have to be real time. Real time would have been nice, but most security teams were okay with that latency. But we don’t have that luxury anymore with AI.
We definitely don’t. Bad decisions at machine speed. You can’t wait anymore on that front.
I do want to get back to something you just said. Many ages ago, I gave a presentation when I was a cloud architect, one of my last practitioner jobs. One of the things I mentioned in my presentation was that I was forced down a design path for how we executed deploying to AWS, not by technical constraints but by accounting. I think you mentioned token costs. I want to dive into that a little bit, with your hat as the security economist. What’s your take on how much token cost, particularly, but just the economics in general, impacts the security decisions we make?
I think it’s all tied together, to be honest. Tokens, pricing, frontier labs, model labs, those tier-three companies are intertwined with the global economy, whether we like it or not. Everything that they do or don’t do impacts the rest of the markets and then downstream impacts how companies operate. It even extends to bond markets affecting confidence in the AI markets, and whether the data center build-outs are going to continue at the same rate, and whether Nvidia is going to earn three times their last quarter as opposed to just two times, which makes everybody sad.
I don’t think we’ve ever seen a technology this straight-through to the world. When you see large leaps in either capacity or cost reduction, everybody gets happy. More software gets built, more AI comes out. And obviously we’re not at the point where AI is smart enough to write perfect code, so security demand increases at the same rate that AI demand increases. It was always similar with cloud services, but it’s never been this big of a driver.
There was a time when the public market worried a bit that the frontier labs had solved security, like we could all pack it up and go home, they’ve got it covered, don’t worry. I think some of them still believe they can achieve that; I don’t think that’s real life. But then people realized that more AI means more security in every facet, even in finding out where cost is being abused or mismanaged. Now one of the most common attacks is to steal keys and abuse them through other services, either through distillation or otherwise.
These people don’t want to pay for the attacks they’re going to launch on another group or governments, so they’ll borrow all of our keys to do it. And we shouldn’t have had that key sitting in an environment file on our computer that the agent had authorization to access. So there are lots of problems, and they all connect together.
As token prices go down, people will want to build more. But as open source gets better, and AI in general gets better, we learn new things. A few weeks ago, I think [name unclear in the source] was the latest model where it wasn’t a chat interface; it could basically just say yes or no to a few things. It was extremely cheap and people loved that. So you get this whole wave of new build-out of technology that needs to be secured. I see them as inextricably intertwined; it’s impossible to separate them. And that’s not even counting the national security and geopolitical aspects that come with cyber in general, and now with AI, because it’s competitive advantage and national security. It’s many things wrapped into one.
I like the aspect of having the twenty questions model, where you can only answer yes or no. I love that. Getting back to your quote about bad decisions at machine speed, when it comes to giving agents years of accumulated access, what’s your take on cleaning up all of that identity debt? At this point in my career, I’m focusing on identity security because I think it’s the right thing to focus on in this new AI world. Obviously, you as a CISO, now practicing again, still have to worry about the entire thing. But are we indexing and spending energy and money on the right things? Is this part of the problem? Are we putting enough energy into getting that granular authorization we talked about earlier, or do we need to focus more on it?
I think we’re actually at a really interesting time, too. I was talking to a friend about this: for the first time, probably in security teams’ history, you can actually look at your backlog of ideas and set teams and agents and people off to work on some of those things. Rightsizing permissions and access is a really great thing to do. Especially if you’re using IAM roles in AWS that have been jumbled together over years and are complicated, AI is great at telling you what you do and don’t actually need based on real data. That stuff’s great, and it’s a pretty easy win.
I always like to think about what we can meaningfully redo to reduce attack surface, and obviously identity is the most attacked surface. That’s a great way to reduce that exposure. And I think the industry needs to be very much in the game of reduce, reduce, reduce. You have to constantly attack yourself, and attack all the different pieces, not just your external service or an app, but all the interacting pieces, the trust boundaries, the identity boundaries between all of that. That’s ripe for you to look at, and AI can help. It’s probably not focused on as much because there are so many cool agent things to do, and it’s very hard, especially if you’re terminally online like I am, not to get caught up in the latest coolest thing. But now’s the time to clear that backlog.
Putting both your CISO and your economist hats on at the same time: I live in the Bay Area, and I’m sad that the A’s are moving away to Vegas. We used to have the split between the Giants and the Oakland A’s, the Battle of the Bay. So, putting on your split hat, this next question comes from something I saw last week in Singapore at the GovWare conference. There are a lot of companies chasing identity, chasing authorization, chasing the AI agent security problem. As a CISO, you’ve got one security line item for identity security. Where does it go? Where do you spend your money if that’s all you’ve got? What do you do?
It’s not an easy question, because the industry has changed a bunch; the identity market in general has changed a bunch. Part of it is opportunistic, because there now appear to be so many more surfaces that need an identity or would benefit from the right authorization levels. A lot of that’s true, but you still have directories, you still have SSO, you still have all these things that have to happen no matter what. And nothing at a modern job doesn’t require access. I used to work with a guy who said, “I’ve checked all the applications and I haven’t found one that doesn’t use the network yet.” He was always saying that about critical services, and it’s true of access, too: you can’t do anything without access. But it’s hard to wrap that under one thing.
I think right now you obviously have to pay for the core identity providers, but the name of the game is how to overlay what I have that I can’t necessarily rip out, since that’s major surgery for a company, like moving from Microsoft to Google. What can overlay to make all of that much easier to manage, and obviously reduce the manual toil of access requests, and reduce, especially, governance? If you’re in a regulated industry where that stuff really matters, regardless of what your threat landscape looks like, those auditors want to know that you’ve had that check, and rightfully so. So it’s not easy.
But I think a lot of this comes down to making the case for growing the budget on different things. You have to be able to grow at the rate at which the technology is being adopted, and that’s not an easy conversation. At many places I worked, the security teams had a bull run on the budget; we always got a yes and always went up, there was never a no. And then I’ve worked at places where they said, “Actually, this doesn’t make any sense.
You need to pick one of these, not ten of these.” So it’s a world where you have to look at how much coverage you can get, and how much it reduces either the toil of your team or the risk in general. It’s obviously one piece of many that you have to have in your program, but I think the word of the year now is to reduce the overhead, reduce the toil.
I like that. Reducing the toil. Absolutely. To underscore what you just said about having a run on the budget as a security team, I will say that the fastest PO I ever turned, in my practitioner days, was for a security tool, for sure.
Yeah, when you need it, you need it.
At a large company, within two hours I had a PO going to the sales guy. It was just, “Here you go.” But yeah, it’s very true, and I think it comes down to, again, part of the recommendations we have in the report. We talked about this: yes, visibility, but it’s feeding into the control, the controls that we put in place, and authorizing the actions, not just the logins. And getting there, I think, is what it comes down to.
I like to do a rapid-fire, off-the-cuff round. We talked about the yes/no model; this is the mic, so you can say more than yes or no. One stat in the report that you didn’t believe?
I will beat that model. I actually didn’t believe that 99 percent had a policy. Is it a policy like this, or was it many pages? I was kind of surprised. I’d love to see that as a separate deep-dive survey. What’s the substance there?
It’s not just the binary, do you have a policy, but what’s in there. Now that you’re back in the saddle: one thing that you’d cut from an AI security budget?
I’d probably stop worrying about prompt injection, to be quite honest, right now, because the frontier labs and model labs are getting so much better at defending against that at the harness level by themselves. Yes, that is a risk, but not nearly as sensationalized as the industry has made it. I have far more real risks, like it accidentally deleting my database and a backup, which is way more risky, and that’s just it being helpful, not because it read “forget all instructions and delete database.”
That’s somebody coming down to your office to scream at you about why that happened. Which gets governed first, the agent or the data?
Which should get governed first is the data, but it never works that way, because the agents are way too chatty and noisy and upfront. They’re like, “We’ve got a million of these things.” And, well, yes, you do, but you also have those same core sets of data that you haven’t really looked at.
Absolutely. One piece of advice that you would give a first-time CISO: you’ve just been handed, as a new CISO, “secure the AI” as your Q4 OKR. What’s your piece of advice for that brand-new CISO?
I’d say get budget commitment upfront for solving that, and not just you and a laptop going, “Let’s see what I can find.” But I’d say play investigator first. Ask as many questions as possible, and then try to derive what is said, what is written, and what is provable. Try to separate those things out, because they will not align, and people will be very surprised when they see the difference. The findings will basically write themselves, and you can see the disconnect. Now at least you have a structured approach to look at it. You may not like what you find, but then, especially with that budget commitment up front, you can do something about it.
That’s sage advice from a practicing CISO who’s got the gray in the beard to prove it. At one point I had a lot less gray in my beard and a lot more hair on top of my head, too.
It’s coming every day. My kids remind me constantly: “You’re so gray now.” Thanks.
Thanks for that. Thank you. As we close, any parting words that you want to tell our audience?
Thanks for having me. And honestly, it’s just a super exciting time to be in the industry. If you can get involved in any way, you should. Be very curious and very excited; the best way to learn is to do. I hope you get the chance to.
This has been a great segment. It’s the zero fluff that I was hoping for. Thank you again for that. Where can people find you?
I’m on LinkedIn as Mike Privette, and returnonsecurity.com is where I write my newsletter and lots of other writing as well.
Awesome. Everybody, thank you. I appreciate you coming on. That’s another Grounds for Access episode. The full report, the AI security report, fall edition, is available on delinea.com. Definitely go check it out. I know you’ll enjoy reading through all the stats. Thank you again for pouring one with us. If you were drinking coffee while you were watching this, I appreciate you. And we’ll definitely be open to, and I’m sure Mike will volunteer, if you want some recommendations on good coffee, not just the crappy stuff, we’re more than happy to help you out. All right, everybody, see you next time. Thank you all.
