Episode Transcript
Karim Toubba [00:00:00]:
And this is, I think, bar none, this is the biggest risk that I've seen in my 29 years in cyber.
VO:
From KBI Media, I'm Karissa Breen, and this is KBKast.
KB [00:00:10]:
My guest today is Karim Toubba, CEO of LastPass, with 29 years in cyber and now sits across the AI governance problem at the strategic level. We talk about shadow AI as the biggest blind spot in security, why agents break every identity model we've built, and why security can no longer be the team that just says no.
VO:
Before we get into it, do me a little favor and hit follow wherever you're listening. It genuinely helps the show reach more people who need to hear these conversations.
KB [00:00:46]:
Alrighty, let's get into it. Okay, so Karim, AI is obviously changing the way people work. As we know, we've heard it, we see it in the news every day, but it's also changing the way people access tools and then move data, which creates risk, for example.
KB [00:01:01]:
So I really wanna know, employees are adopting AI faster than governance can keep up, as from what I'm hearing from folks like yourself. So I wanna start with this. Is shadow AI becoming the biggest identity blind spot inside organization, given your role and your pedigree in the space?
Karim Toubba [00:01:20]:
It's actually a great place to start off. Shadow AI is becoming the biggest blind spot, not just in identity, but in security in general. And I think so much of that is kind of driven by the pent-up demand and the promise of what AI holds more broadly. And what that effectively means is that security and IT organizations who are under an inordinate amount of pressure to actually enable, not stop AI, because of that promise that it holds in terms of the ability to unlock potential for productivity, efficiency gains, and growth for organizations of all sizes are being asked to support fast-paced adoption of AI. And that has really broad-based implications. That has implications, as you mentioned, KB, to identity, clearly, especially as you get into not just the human-AI usage world, but the agentic AI world, because identities of agents is a very different paradigm model than identities of users. It has implications to data. The promise of AI is to be able to process vast amounts of data very quickly and give you incredible insights and perspectives that humans can't.
Karim Toubba [00:02:34]:
And therefore, you have to have a much stronger data governance model. And part of that is around understanding what data people are sharing. And thirdly, it has real implications in how people are using AI. And the how is kind of rooted in 2 areas. The first area is what AI models are people using? Right? So it could be Claude from Anthropic. It could be Codex if you're an engineer from OpenAI. It could be open-source models. It can be open-weight models.
Karim Toubba [00:03:04]:
But then the second implication is form factor. How am I using AI? Am I using it through a browser? Am I using it through an extension? Am I using it through a desktop form factor like Cowork? All of which have to be understood to provide the right level of visibility and ultimately governance associated with that. So It's a pretty multifaceted perspective that I think people are really attempting to get their hands around quickly.
KB [00:03:30]:
You mentioned something before around human identities versus agent identities. Would you say people/companies classify those identities as the same versus treating them differently?
Karim Toubba [00:03:41]:
Yeah, I think it depends on the organization, how much they've adopted, their sophistication of thinking, the people they have on board. And so my general view is that I think people are not classifying agents and users the same. However, they believe that one of the areas that we believe is inconsistent thinking is what we see companies believe, which is that the model that was used to understand how to put guardrails around users is Being adopted to put guardrails around the agents. And the problem is the intent of users is drastically different than intent of agents. The scope of users is drastically different than the scope of agents. And there's a certain amount of judgment and training that can go into users that oftentimes is not the case with agents because oftentimes you're asking agents to do whatever they need to accomplish a particular task. So people see it as different. The problem is they're trying to adopt a similar model to protect their users.
Karim Toubba [00:04:48]:
as well as protect their agents. And I think in that lies the opportunity to both educate and deliver real value to companies.
KB [00:04:54]:
And sorry, just to clarify, when you say similar model, do you mean like similar approach on how they're—
Karim Toubba [00:04:59]:
Similar approach, right? Like, you know, do you like take PAM as an example, right? The entire PAM use case or privileged access management use case is pretty well defined. It's been a market that's been around for a long, long time. It's really platforms that are deployed for the, people that really kind of access and hold keys to the crown jewels of an organization. And it includes multiple controls like MFA, just-in-time access, session recording, all of which are repurposed to really enable to understand what a privileged user is doing and put guardrails around it. Well, agentic workflows, I think, kind of break that model in many ways, and trying to adopt a traditional PAM model to agents is typically not something that will scale in the fullness of time.
KB [00:05:46]:
So then just on that, looking at a traditional PAM model from my experience, we used to map it against like Karissa as an individual versus like Karissa's role now and to constantly revisit that. Now, however often companies revisit that is on them, but it used to be like once a year, once every couple of years, and then maybe someone, identity and access team in a big enterprise would go in and investigate it and then perhaps start to question me to say, hey, KB, you've got a lot of very privileged access here. We might need to revoke that or reduce it down. So, would you say there's a lot of admins still around looking at all these people? If you look at an enterprise with 50,000 people, it's hard to sort of map each role against it versus moving away from that traditional PAM sort of model.
Karim Toubba [00:06:29]:
Yeah. Yeah. I mean, look, the PAM model held true for many, many years and still holds true today for your true privileged users, your administrators, your security admins, your network admins, your system admins, your application admins, your database admins, and the list goes on. People that have access to some of the most critical infrastructure full stack for organizations. And there's a set of guardrails, including review policies, that really kind of give you visibility and guardrails around what those privileged users can do. And the model is structured such that you can enable those capabilities for a period of time and then revoke access. But it's really relegated to a small percentage of users. I don't know what the latest data shows, but A while back, I think it was a very small single-digit percentage of the total users inside of an organization are, in fact, true privileged users.
Karim Toubba [00:07:17]:
And so, in the world of agentic, that model completely breaks, right? Because, you know, the number of agents in relationship to users is significant. There are some early studies that show for every user that you'll have, you could have upwards of 100, 100+ agents that are doing work on behalf of the user. And then, of course, you have kind of like agentic sprawl or agentic harnesses that enable agents to subdivide the workloads into sub-agents. And so the problem is much broader in its scale. That's one issue that organizations have to kind of think through. The second one that they have to think through is the intent, right? As a privileged user, yeah, periodically you may have a rogue or a nefarious privileged user, but more often than not, you don't. They have very specific tasks. So for example, a database administrator will log into a database, look at particular tables with a specific intent in mind.
Karim Toubba [00:08:11]:
Let's say, make an update, solve a particular problem, troubleshoot a support-related issue as the database administrator. When you have an agent logging in, they have access to, first of all, typically they have access to a lot more. And their remit, the way agents work, is look at anything and everything you have access to, to drive the desired outcome that you are given as an instruction. That instruction could have been given by a human. That instruction could have been given by another agent. And so as a result, the scope at which the agent peruses through data, tries to do lateral movement, looks at different data structures is very, very different and is done at a breakneck speed. As a database administrator, I can only do so much. As an agent, I could do things a lot faster.
Karim Toubba [00:08:58]:
And so it's trying to think through what the implications of those things are systematically across each aspect of the system. to provide the necessary guardrails.
KB [00:09:08]:
So what do you think at the moment is really catching companies off guard when it comes to identities, whether it's around the human element, we've sort of evolved, but then also around the agent side of things, would you say, given that your sort of role talking to customers more at a strategic level?
Karim Toubba [00:09:26]:
Yeah, well, I mean, I think to kind of answer that question, KB, the first thing that's good to kind of discuss is to put ourselves in the place, in the shoes of a customer. Because I think it really kind of helps understand what's driving this. And the one thing that's been fascinating to watch, I've been around the industry, I've been in cyber 29 of 31 years, in tech 31 years. We've seen a lot of kind of like booms and busts and high pace of adoption, cloud being one of them, mobile being another. I've yet to see an adoption at this pace of scale and speed. Normally adoption cycles are such that you kind of have early adopters. AI is not the case, right? We see large-scale enterprises upwards down to small businesses or very small businesses that have 5 to 10 employees all the way up to large enterprises that are adopting AI. The pressure that boards, business owners, management teams, and executives are putting on the rest of the organization is widespread.
Karim Toubba [00:10:27]:
And it is intense in its variation across different parts of the organizations, meaning that business leaders all the way up to boards are basically saying, look, The promise and potential of AI is such that we are willing to take associated risks. It is why we're waking up today in August of '26, recognizing that the governance, visibility, and control of AI is lagging significantly behind the adoption curve of AI, which is why it's creating such a challenge for most organizations. And the best way I could capture that is a conversation we recently had with a customer. That basically said, the team of no will not survive, right? Typically, security teams have been in a position where they've been saying, look, we're not going to allow this because these are the risks to the organization. This is where it doesn't comply with policy. That's very different today with AI because basically they're being overridden by largely business people that say, look, I'm willing to take that risk because the upside, the return on investment is so significant in terms of growth, competitive advantage, profit margins. So you have that as kind of the background. So as a result, the real underlying question then is, how do you solve that problem? And really, in its heart, it really is one around visibility, control, and governance.
Karim Toubba [00:11:45]:
The control piece is a little interesting here because control in the security world historically was somewhat binary and contextual, meaning that you would say, hey, I'm going to block or allow. I think what we're seeing more and more is organizations less wanting to block, but more warn. So KB accessed a model that is not, quote unquote, sanctioned. Let's say you were an organization that had doubled down on Anthropic and was using Anthropic, but then you installed an application that was using another model. We don't want to stop you necessarily from using it, but we want to warn you and provide some context and said, hey, KB, we have an enterprise agreement with a bunch of guardrails and security measurements in place with Anthropic. Please go use the Claude app as an example, you have a week before we effectively shut down access to what you're doing with OpenAI. And this may be one of several examples we run into routinely. So, it's really about how do you use a mechanism to, after you gain visibility, to really better understand how you educate users and then ultimately put the same guardrails around agents.
KB [00:12:52]:
So, when you say educate users, I feel like even when I was on the practitioner side, we've had the same conundrum. For example, you know, there are people out there that at one stage in an enterprise marketing people were sending files through Dropbox. Well, obviously they had X amount of time, then we put controls and that sort of stuff. But don't you think the pattern of people just stayed the same though, especially when we start getting into shadow AI and stuff like that, that we've, that we've already seen the AI sprawl that I'm now hearing in the United States versus other countries. So you've probably seen that if you've been in the game for so long, you've seen the same pattern, but it just sort of shifts whether it's AI, whether it was cloud, whether it was whatever else going on. How can companies actually approach that effectively? Because you're right on the conundrum around executives saying we need to be doing AI, but they don't really understand the full fidelity of what that means. And then to your point, it presses down on the security people who are not, who are historically always said no, but then what do we have to do? We have to appease our executives who's been there for 40 years and so-and-so's ego. But then also what I'm hearing is people saying, Well, if we don't do it, we miss out.
KB [00:13:59]:
And back in the day, as you would know, people would sit around a table, we'd make decisions, it would take forever. Well, we don't have that privilege anymore to do that. We got to make decisions fast. So then when people are doing that, especially in a big bureaucratic company, do you envision that we're just going to see more mistakes happen? Perhaps they're not measuring twice before they're cutting.
Karim Toubba [00:14:20]:
Yeah. Well, we'll definitely see more mistakes happen. And we'll see more mistakes happen for 2 fundamental reasons. And by the way, KB, you're 100% right to compare it to SaaS, right? Just compare it to cloud and SaaS, right? That kind of came, right? Obviously, cloud infrastructure was a precursor to SaaS, but compare it to both of those kind of shifts, if you will, in the model, in the technology model world where people actually started using different kinds of form factors to do the work that they did. elements were an application in the cloud or a SaaS-based application accessed by a human or a series of humans, ultimately everybody within the org as the organization moved to SaaS-based applications instead of hosted applications. That was still relegated to X number of humans. So if you were an organization with 2,000 people, you had to worry about 2,000. You could put the guardrails about that.
Karim Toubba [00:15:11]:
So the 2 reasons this is gonna be different is, first of all, in the agentic world, as we mentioned earlier, you have a much higher ratio of agents than you do humans, which compounds the problem and makes it that much harder to manage purely from a scale perspective. The second reason is because if you think back of the early adoption days of SaaS-based applications and even cloud infrastructure, there was a certain amount of education that had to be done. Take Amazon EC2, all the sort of the cloud-based services. You had to have people that got kind of trained on what the infrastructure looked like, and you had a bunch of sub-industries that were built around creating toolsets around that. With AI, this is a natural language interface that you don't have to be that proficient in, right? And so, as a result, it's creating that access layer and making it available to the masses to reach that productivity. That is what's creating a lot of the risk initially is because you're making it available so quickly to pretty much everybody simultaneously. Whereas the cloud and the SaaS world was kind of very measured. in its adoption and growth.
KB [00:16:20]:
So, it's kind of like a double-edged sword.
Karim Toubba [00:16:22]:
That's right.
KB [00:16:24]:
So, then would the risk then be greater? Because like anyone, you know, Joe and Jeff who've got no clue about this can use it versus maybe a more skilled practitioner historically.
Karim Toubba [00:16:32]:
That's right. Yeah. I mean, look, I mean, this is, I think, bar none, this is the biggest risk that I've seen in my 29 years in cyber. I mean, there's just no doubt about it. And the early data kind of proves that. And, you know, we haven't even We haven't even started to talk about model intent, rogue models, as we recently saw with OpenAI and the Hugging Face incident and other things, right? I mean, I don't want to fearmonger because I hate when we do that as an industry. We scare people into buying our wares. That's not the intent, but I think people have to be very, very thoughtful and very serious about what's ahead of us.
Karim Toubba [00:17:08]:
I think the potential for AI and the potential for it to unlock increased profit, increased growth, efficiencies for business, competitive edge is real. There's no doubt about it. We're big users of it. Every organization I talk to is moving in that way, some more aggressive than others. But I think you really have to kind of take a moment, pause, take a deep breath, step back, and really ask yourself, okay, what kind of investments am I doing from a visibility, control, and governance perspective to really kind of keep up with the adoption that I'm seeing?
KB [00:17:39]:
The other thing that I'm hearing from people in the industry that I'm interviewing is saying like people aren't, they're doing governance theater. So it looks like governance in theory, but practically doesn't really do much. Or there's a lot of things called like drift risk, for example, where we start to see things shift. So how do people get a bit of, a bit more of a foothold on that? I mean, it's easy for you and I to talk about this at a high level on a podcast, but I mean, it's a lot harder to actually implement that. in everyday workflows, for example?
Karim Toubba [00:18:10]:
Yeah, it's a great question. Well, I will say, I won't spend too much time on this, but I will say there's a lot of AI theater happening, not just governance theater. You know, I talk to other executives who do a lot of what I call AI chest pounding, where they talk about how amazing AI adoption and implementation is. And I often ask, okay, so what's the benefit to the business, right? And so, you have to be very clear-minded of what the goal of adoption of AI is. Not just that it, at a high level, you know, has the potential to unlock the things that we talked about. But as it relates to governance, it's a really good question because I think this is the, this is kind of where the execution and adoption rubber meets the road because, you know, and I'll use us as an example. You know, we're ardent supporters of adoption of AI. We've adopted AI internally as an organization.
Karim Toubba [00:19:01]:
Now, we are a security company with just under 1,000 employees, so perhaps we're a little bit more security-minded than most, but we've been very particular and specific about what our employees can and cannot do with AI. We've also been very deliberate about what AI tools our employees can use. So we've been down to the point of the applications that people can use, the models that people can use. And we've put a fair bit of controls around the guardrails so that we can get visibility which is key into all applications, AI-related applications, both cloud AI applications through the browser, but also desktop applications. And so having that level of visibility and then the commensurate capabilities around it to either eventually block, but also warn users and guide them toward the right behavior is what allows you to adopt the governance model. So people sometimes will rush, and I think this is what you meant by theater. They'll rush to governance. But then they'll forget that the building blocks that allow for governance are really rooted in visibility and control.
Karim Toubba [00:20:09]:
And the old adage we've used, talked about in security for decades is true. You can't protect and you can't govern what you can't see. So rooted in that at the very core is you have to be able to inventory all AI usage in all form factors across all systems. And from there, you build the foundational policy for governance.
KB [00:20:29]:
So, just going back in your comment when you said before, when I'm asking people, what benefit to the business does AI really have? What do people really answer with?
Karim Toubba [00:20:39]:
It is all over the place. I'll tell you what our answer is. Like, we've been very clear since we started to adopt AI. And we said, look, you will use AI for one of 3 goals, period. Number one, you will use it to create better customer experiences because, you know, we're a vendor We provide technology for our customers, and leveraging AI to create a better onboarding and a better experience within technology usage of our product is first and foremost to us. Number 2 is use AI to unlock growth, right? I mean, that is a real potential because there are investments that you can make at a much faster pace that allow you to accelerate revenue. And we're not— we're a for-profit business, so that's really important to us. And number 3 is proven efficiency gains.
Karim Toubba [00:21:25]:
Right? And this is the part, candidly, where it gets a little bit more delicate, and you have to be very thoughtful about it. Because when the average person hears that, they think about jobs. But it's oftentimes not related to jobs. It's related to how can you use AI to actually displace technologies and tools that you had before? How can you use AI to actually streamline a process? This notion of process re-engineering, KB, is a really important one. Because I often see some customers make a mistake where they use AI to try to automate an existing process when they can actually get more efficiency if they use AI to automate the process, but simultaneously think about re-engineering that process to create more efficiency. Automation through AI of a poor process doesn't give you the right efficiency gains necessarily. So it's kind of that trial by fire. And really kind of understanding what your goals are.
Karim Toubba [00:22:21]:
And stating those goals upfront allows organizations to really be thoughtful about how they're gonna use AI because they've been given a clear mandate and direction of what the ultimate target is.
KB [00:22:33]:
That's an interesting comment you mentioned around the people thinking, well, we can displace X tools 'cause we can automate it. Or some people think, well, you know what? I can actually go and vibe code it myself. I was having conversations some of last week. big vendor that they want to push out because they believe they can do it themselves. So what's probably going to happen? They're going to call the vendor in 6 months and say, actually, we kind of need you guys. So it's always more like, I could go and potentially paint a house. Doesn't— I can go buy paint from this hardware store. Doesn't mean I'm going to be a very good painter.
KB [00:23:02]:
So we're going to start to see people going, oh, you guys are costing us way too much money. So therefore it gives an illusion that we're driving these costs down. then figured out pretty hard to do this by ourselves because someone still has to govern all this stuff at the end of the day versus we're gonna now call the vendor back and costs will, as a result, go back up.
Karim Toubba [00:23:21]:
Yeah, there's definitely the risk of that. I think that what we've been is very selective in areas where we decide, and remember, we're a technology company, obviously. So, this is different for other companies. I think we've been very deliberate and selective in areas where we've decided that there are things that we can do with AI that replace the capabilities of existing and technology tools that will allow us to reduce our costs and become more efficient simultaneously. And so, a lot of that has to do with understanding the economics of the business. A lot of that also has to do with kind of having a really direct conversation with various members of the teams about what is a core competency for us, what is not. The piece that's exciting, but also risky, about AI is that it can solve any and all problems. And so you do all of a sudden open up kind of Pandora's box where teams have historically said, wait, why am I building this? It's not my core competency.
Karim Toubba [00:24:21]:
Suddenly, you're coming back going, well, wait a minute. I can vibe code this now. Suddenly, it can be sort of a derivative of a core competency of mine. I think this is where leadership, again, has to be very deliberate about what the goals are for an organization. And I talked to some organizations that are kind of doing that very, very well, and they're being selective about the technologies that they replace.
KB [00:24:44]:
We'll come back to that in just a moment after a quick word from our sponsor. Enterprise tech leaders know that compliance isn't just about ticking boxes. It's about risk, reputation, and revenue. That's why companies trust Vanta to streamline their security and compliance workflows at scale. With deep integrations and automated evidence collection, Vanta takes the manual audit grunt work out of the frameworks like ISO 27001, SOC 2, and GDPR. Visit vanta.com/kbkast, V-A-N-T-A.com/kbkast, to learn more.
So I want to sort of shift gears now and talk about Shadow AI, we're hearing about it a lot now in the space. So where does it sort of begin? And do you think— and you've touched on it a little bit already in the interview— do you think it stems from employee impatience? Because now, as we know, people want everything now, faster.
KB [00:25:40]:
They'll go around things. Do you also think there's a little bit of leadership pressure? Or do you think it's just bad policy? Perhaps there's that governance theater bit in there, or a lack of unusable Yeah,
Karim Toubba [00:25:54]:
I think it's probably a combination of all of those things. We'll start kind of from the top down. As we said earlier, there's definitely pressure from the top down. Like, there's just no doubt about it, right? And it is impossible to be an organization of almost any size today without seeing the potential of AI. I was recently in Australia, and we had done a kind of a forum for small businesses. I was surprised 100% of the audience had already started to adopt AI. So this is not just an early adopter enterprise syndrome because I think most people see as one of the use cases that was put forward to me by small business that said, look, we're able to get 20% to 30% more efficient by the adoption of AI because we were able to put it in the hands of people. And so that is something that is undoubtedly driving pressure resulting and driving shadow AI.
Karim Toubba [00:26:48]:
I think the second one is kind of bottoms up. And I would put employees inside of companies into 2 buckets, KB. The first one is people that are clamoring to use AI because they kind of exhibit certain characteristics. They're technically curious. They also have always kind of followed the trend of new types of technologies. And they see it as an opportunity to themselves become more efficient. as employees. And then there's kind of a group of employees, typically smaller, that have concerns about AI for a variety of reasons.
Karim Toubba [00:27:26]:
Job security could be one of them, the fear of something new and different. You know, I mean, people come in all sizes, shapes, and colors, right? There are people that kind of adopt things that are very quick at a much, much slower pace. But undoubtedly, we're seeing the majority of organizations really kind of adopt AI very quickly. And I think both The pressure from the top down, which is giving people the remit and the latitude to adopt quickly, and the insatiable appetite for everything AI-related is really kind of driving shadow AI. You see that manifest itself in multiple form factors. The one last thing I would say is this is one of, unlike cloud and unlike the early days of SaaS, Those were things that happened and started in business. AI is being adopted everywhere, including in the home. And so there's this connection between the personal— the way I'm interacting with AI on a day-to-day basis with perhaps what started off as ChatGPT, which was much bigger in the consumer space before Claude came on the scene.
Karim Toubba [00:28:33]:
And then suddenly, I'm expecting the same level of access and capabilities, especially as models get more efficient. In the workplace. So I think that's driving it as well.
KB [00:28:42]:
And going back to the whole efficiency and all that sort of stuff, there's a site, I think it's called Artificial Analysis AI, and it talks about how much it costs for intelligent, like it's an intelligence index task, how much it costs. And so for example, just to give people a view, a task with Claude Fable 5 is about $3.14 US dollars. So what I'm hearing from people, and maybe you would have a little bit more insight on this, is people starting to discover, well, actually, we're spending a lot of time either telling and doing all of this stuff, which is costing us more money in terms of the tokenomics side of things. And we, that's a whole different conversation. It's costing us more money to do and implement these things than it is to potentially just do it. So I'm now seeing people sort of go the other way to be like, well, hang on, we thought that we were reducing you know, Joseph's $400,000 salary, but that's actually costing us more to set up the whole thing. So now I'm sort of, are people coming back in the middle? Like, have people over-indexed on it? And now there's going to be a blend of the two. They're still working it out.
KB [00:29:45]:
What does that look like? People are worried now to say, oh my gosh, look at my budget's been completely, as someone said, blown the doors off it. So like, I'm just hearing so many ends of the spectrum now that it's like, well, what is the actual truth?
Karim Toubba [00:30:00]:
Yeah, so it's a great point that you bring up, and it's kind of even a little bit more variant than what you described. But I'll start with what you described, which is very true in the pent-up appetite to move quickly. There were decisions made by organizations, especially early on, to make moves to drive efficiency. That could be back to the tooling conversation or people, where tools and/or people were managed out, if you will, only to come back and recognize and realize, well, wait a minute, there's actually some more work that really needs to be done because The balance of the cost basis is not kind of worth the return on investment that we've really seen. So, we have seen some examples of that. I would say when you move fast, that's kind of what you're going to get. And because this industry right now is moving at such a torrid speed, both in terms of capabilities and cost, that you really don't have a clear view of how that's going to evolve. So, organizations are kind of trying to figure out how to move fast without making changes that are regrettable.
Karim Toubba [00:31:02]:
I think one of the best conversations I had with a customer is just make sure we don't walk through a one-way door, meaning that a decision that is made can't be reversed. Because sometimes there are those types of situations that occur. And then the other thing, of course, from a culture perspective for organizations is be really thoughtful about things that impact humans and their livelihood. Because even if you choose to bring people back, just the the cultural impact of doing that to an organization and what that potentially means has a downstream effect. The other thing that's important as a subset of that we didn't bring up is that we believe deeply that organizations will move to a multi-model mode. And what that basically means is that, yeah, we'll use Haiku. We'll use DeepSeek. We'll also use Fable.
Karim Toubba [00:31:49]:
It depends on the task. And that ability to kind of understand what the task is, what the appropriate model Is and what the cost basis is. So that's, those are the 3 variables in the equation of driving efficient outcomes for leveraging of AI, understanding what the task is, understanding what the model is, what the right model is, and can it be efficient and effective to perform the task, and what the token or token economics look like are things that you have to take into consideration. This is a really interesting and important area. To get the economics of AI right.
KB [00:32:24]:
Do you think it mirrors itself against what we saw when cloud first started up? Meaning it was all about on-prem. Oh no, everyone's gonna do cloud. No, we're gonna do a hybrid now. Now obviously people do multi-cloud because do you think that companies are putting all their eggs in one basket? What I mean by that is when OpenAI was top dog, now it's like, well, actually Anthropic, now they're the top dog all of a sudden. Very volatile industry. So, are companies thinking we've got to spread across all these different frontier models or else it could become a risk, as we've clearly seen recently about certain companies out there not being able to perform. And if everything's hinged on that, that becomes a problem. That's why if I look back at the whole cloud stuff, very similar sort of approach.
Karim Toubba [00:33:10]:
Yeah. I mean, again, it's even more nuanced than that, right? The recent announcement a couple of weeks ago by Kimi K3 and some of the results that they've had on the efficiency of their model relative to frontier models at a fraction of the cost. You know, geopolitical implications, innovation against China, topics aside, which could be another podcast altogether. I mean, those are things that people have to actually seriously look at, right? I mean, you know, just to give you a sense, we're users of Bedrock because we're heavy users of AWS. AWS provides multiple models at different cost points when you use Bedrock and make calls. And so, I think most organizations will start to and have already started to build their appropriate infrastructure, understanding that the models will be substituted. You'll partly play an economic arbitrage game. You'll partly play an efficiency game.
Karim Toubba [00:34:01]:
I think those are things that have to actually be taken into consideration when you're leveraging AI as a whole.
KB [00:34:07]:
And then, do you also think it's going to be a bit of a race to the bottom? Because, again, if I can just get answers from something and I can get it for $0.10 versus $10, well, I'm probably going to go with 10 cents because it's cheaper. Because at the end of the day, that's what everyone— there's a reason why people buy on Amazon, because they can get a phone charger cheaper than if they go down to the store down the road. So do you think that that's going to come into play? But then as a result, that's obviously going to open other geopolitical tensions and all sorts of things.
Karim Toubba [00:34:32]:
Yeah, I don't know if it's a race to the bottom, but it's certainly a race— I, I don't want to say downward, but I think there will be a lot of pressure on the frontier models in particular. And I think the pressure is twofold. Number one, the frontier models are getting pressure from us as consumers in business to stabilize and ultimately reduce their prices. Because when you see companies like Anthropic, I don't know what the latest numbers were, but they're gargantuan compared to any type of growth numbers we've ever seen in technology. I think they're projected to go from $10 billion to $30 billion. Now, yes, they're signing up more customers, but a large part of that is a spike of increase of token utilization. And so, most organizations are looking at this saying, okay, I'm going to allocate X number of dollars to tokens, but I can't support a model that, to your point earlier, about a 10 or 20x increase every year. And so, as prices increase, the price per unit, per token has to come down.
Karim Toubba [00:35:33]:
That's one point of pressure they're facing. The other point of pressure they're facing, again, not to get too much into the geopolitical world, but You do see open-source, open-weight models, plus China, making models available at a much cheaper rate. And of course, they use a process called distillation to try to get to models that are built fundamentally off of other models as much as possible to maintain that competitive or to achieve competitive parity as quickly as possible. I think that the prevailing, at least our view, there will be I don't wanna say commoditization of the models, but you'll start to see more value accreted to either the infrastructure side of AI or the application side of AI. I think it's one of the biggest reasons why organizations right now are really starting to think about, okay, what do my applications look like? How do I leverage AI? And what does implementation— what do implementation cycles look like? Because you can have the most amazing models in the world, But if you don't have an efficient, scalable way to do implementation, it becomes a real kind of red herring for organizations, especially as they pile in and drive more dollars towards that investment.
KB [00:36:43]:
And as you were talking, what was going on in my mind is, do you think what you've just discussed is then going to be a correlation of what companies are gonna look at with their vendors that they use? So, for example, like even cloud hyperscalers, they now know they have to be a lot better on how that, like you said before, you can swap using different frontier models because it's gonna going to be a better bill at the end of the month versus what it used to be. People aren't as locked into a vendor. Now we're seeing random companies we've never heard of get all the way to the top very quickly. That never used to happen over the last 30 years. So do you think this puts a bit of pressure pushed down onto the vendor market to say, well, our customers are going to start asking us questions, or they're going to say, well, Karim, you know what, I'm just going to vibe code it. I'm going to get Johnny. He's really smart at it. He's going to do it.
KB [00:37:31]:
And therefore we're cutting it off for the year. But then, as we said, Companies may realize it's a lot harder to do it than thinking in theory, and then they may come back. It's just more curious on how that's gonna put impact on the vendors.
Karim Toubba [00:37:43]:
Yeah, I mean, look, it's gonna put impact on all the vendors, right? Both to implement AI within their product, keep providing value to the consumer, but also it's gonna put pressure on vendors from a cost perspective. I mean, there's no doubt. I think everybody is seeing it. The media has kind of run ahead of the curve of this whole SaaS-pocalypse thing where SaaS companies are dead. And, you know, sometimes I wish the media wouldn't run, you know, so far ahead of the reality. But I think there is some truth. We see it in the vendors that we choose. We're putting pressure on them.
Karim Toubba [00:38:14]:
They're seeing economic pressure. We're expecting better price per unit. We're expecting them to adopt AI within their technology that will drive more value and efficiency for our businesses. And I think that's gonna continue.
KB [00:38:27]:
And sort of just to maybe final question to conclude, what do you think sort of moving forward? I know we've discussed a lot of ground here today, but given where you're at and the conversations that you're having across the world. What can you sort of share with, what can we see for the rest of 2026 and beyond?
Karim Toubba [00:38:43]:
Yeah, well, so let me start super big picture. I think it's important to remind everybody that this is a major technology shift. There's no doubt about it. But like every major technology shift, we tend to overestimate what's gonna happen in the next 12 months and underestimate what's gonna happen in the next 10 years, right? That's a very tried and true thing. It happened in cloud, it happened in the internet, right? Internet was supposed to be amazing. And then there was the bust in 2000. And then look at everything by 2010 that had unlocked. It completely transformed how we bank, how we purchase, how we travel, how we engage online.
Karim Toubba [00:39:20]:
And so there will be incredibly deeply transformative results as a result of AI. They're not all going to happen overnight. So it is important to keep that in mind. This is an underlying technology that is going to create And ultimately recognize a lot of the promise that's associated with the current hype. I would say, as it relates to the topic we've been talking about, which is security, I think what I expect to see is a serious, serious move by organizations to attempt to close the gap between AI utilization and visibility and governance. I think the AI horse was let out of the barn for all the aforementioned reasons that we talked about. And I think that the industry is now kind of taking notice and attempting to catch up and close the gap and ask itself, okay, what are the controls that I have in place? What's the visibility look like? And what is the associated governance and policy? And I think through the balance of the year, and certainly in '27, you see a big push, certainly by vendors, because, you know, obviously there were, you know, there's a self-serving interest in educating the broader market. But I also think that as I talk to customers, there's a real pent-up concern, candidly.
Karim Toubba [00:40:32]:
They understand and fully appreciate the fact that they've had to take a very aggressive approach to adoption everywhere. But I think many of them are now looking around and saying, okay, there's a real risk associated with it. And while we're not— while we're being told we can't block and stop utilization, we are being told that we need to really provide the associated guardrails to minimize that risk.
KB [00:40:57]:
That was Karim Toubba. What stays with me is his warning about walking through a one-way door, making an AI decision you can't reverse. If you're a CEO listening to this, before you cut the tool or the team, ask whether you can actually get back to how you worked before, because a lot of companies are quietly walking those decisions back.
VO:
I read every reply. If you've got some thoughts on this one, send me a message on LinkedIn. KBKast - Cyber for the C-suite.
