Next in Tech | Ep. 259: The RSAC Conference – Agents on The Loose.

View Full Transcript

Eric Hanselman: Welcome to Next in Tech, an S&P Global Podcast where the world of emerging tech lives. I'm your host, Eric Hanselman, Chief Analyst for Industry Research at S&P Global. And today, we're going to be talking about the RSA Security Conference, a little preview in terms of what we're expecting and what our outlook is when we start rolling into the first of the big security events of the year. And joining me to talk about it are Brenon Daly, Scott Crawford, and Dan Kennedy. Welcome back to the podcast to you all. 
 

Scott Crawford: Thank you, Eric. Always a good time here. 
 

Eric Hanselman: We aim to provide entertainment for our listeners in various forums. So starting things off, what are the expectations we've been talking about, the various AI aspects that are rolling along. We can get that straight out on the table, all of the scary things that are happening with agents, maybe things like dabbling an open claw, community projects that are out there. Where shall we start heading into what is going to be, I think, a thrilling and certainly entertaining week at RSA. 
 

Scott Crawford: Gosh, where are you start? I think it was 2 or 3 years ago, our headline for covering the conference then was AI everywhere all at once. And that is a theme that has just not abated over that time. And now of course, it's agentic AI everywhere all at once and all its manifestations. And you alluded to one of the big stories that we've seen in the run-up to RSA, and that's the viral outbreak. Is that a good way to refer to it of open claw, the appearance of Mebook as a social network for agents and the ability -- the agents gone wild basically out in the wild. And people giving them carblage in a lot of cases to do any sort of functional thing that they wanted them to. And it really -- one of the key things it illustrated was that we're trying to take a measured control, relatively adult approach to adoption of agentic functionality within the enterprise, there's this demand out there and an awareness of what agents can do. That's one of the key things that took away from the Mebook episode and the viral rise of open clause is delta between what we're trying to manage in the enterprise, and the idea of people want so much more. And to the extent that it appears within the enterprise, a thumbnail definition of shadow AI, shadow agentic AI following that same trajectory. So that was one of the key themes we saw rolling up to it. 
 

Eric Hanselman: Well, and it had all of those characteristics that all security professionals want to opine about, which was it was not fully autonomously created, but the actual site for it got built by an agentic framework and had a whole bunch of security vulnerabilities. Trusting the agent with credentials. The author had an API key exposed because he trusted the agent with it. It seemed like one thing after another in terms of the cautionary tale cavalcade that became. But as you're pointing out, this is -- we're getting into territory in which there is such strong desire, the carry-through of this imperative that's been put on AI in an environment where, I don't know, depending upon how you count it, in security, we've been touting AI as the way to manage scale and scope for what, last 5, 6 years, I mean you look at talking and debate about those first entrants of putting some level of AI and/or machine learning into place. But now we're in a slightly different territory. We spent decades locking down human privileges, and now we're handing root to nondeterministic shadow agents. I mean -- so I think there's legitimate nervousness about that, and that's why we're rolling into RSA talking about identity expansion. 
 

Scott Crawford: Among other things, yes. I mean, we've been -- at the beginning of the generative phase here with the emergence of ChatGPT, we talked right away about 2 main domains where we would expect AI to have an impact. And one of them is AI for security and the other is security for AI. AI for security, of course, employing AI and security tactics, tools, technologies, not to be too iterative about that, but in defense and security for AI, defending AI itself against threats. But now we have a third category, which we recognize then with increasing sophistication, phishing attacks, leveraging generative AI. But now with agents, being able to emulate reasoning, I want to carefully use that word emulate reasoning. I don't want to imply consciousness to anything agentic yet. But the ability to emulate reasoning and make decisions based on the context, take actions, and use tools. So now this third category we think of now is security from AI. And we have to think about what impact that's going to have on security technologies and defense. And that's one of the key broad areas that I expect to see a lot talked about rolling on into RSA. 
 

Daniel Kennedy: It's going to be interesting to see where we apply it, where is Agentic really going to matter. We're spending a ton of time in the talk sphere on really complex use cases, finding zero days and bug bounties and wherever else. And even writing code, we're backing into these explanations like, well, it's very deterministic activity. AI is good at it. No one's considering for a second that the people best enabled to leverage AI are using it for the use cases they care about. And is that where we want to end up at security? Is this the best use case? So pointing AI at some very high-end use cases and arguing about them, but we're potentially ignoring other use cases that are less technical, more mundane tasks that just aren't being addressed. The CEO, we deal with regularly brought up this morning on LinkedIn Privilege Management. He was saying one of his early tasks was trying to model job roles and then enforce least privilege. This is something that enterprises spend an inordinate amount of time doing. No one enjoys it. Everyone agrees the process doesn't work, and we're just asking the CFO to validate all the access to financial systems, and he or she is just signing off on it because they really can't do that in a realistic way. Does AI give -- that's just one-use case. Does AI give us a better way to actually realistically do these things? And are we spending this myopic focus on just the most difficult out there use cases and ignoring some obvious ones. 
 

Eric Hanselman: We have to do the hard problems first, Dan, come on. 
 

Daniel Kennedy: It's fun. It's very shiny. I mean that's a good talk topic, and it will certainly fill the audience, but it doesn't move the needle much in enterprises. So it's like, all right, how do we resolve that difference? 
 

Eric Hanselman: The issue in many cases is where are the rewards buried and who's trying to dig them up. Right now, the rewards are for all of the hard problems. If you look at where investments are being made. This is the kind of thing that people are chasing. 
 

Daniel Kennedy: Brad, explain the stock market to us. 
 

Scott Crawford: Please, someone... 
 

Brenon Daly: So when Scott was talking about what Agentic can do or could do or will do, Wall Street has already priced in that it's done it and that it will never go back to the way we did things in the past. Agents will be tasked with everything of security and non, by the way. Everything will expand out to agents will be able to replace applications, every single app. There will be no work that flows through apps, they'll flow through agents. And it's as shortsighted as some of the use cases that Dan is highlighting, people struggling to come up with. And from my perspective, again, we're from the financial point of view, we're coming into a very different RSA. Scott's mentioned similar themes throughout the year since the introduction in November of '22 of ChatGPT, security has been kind of a distant thought afterthought around look at what shiny new parlor trick, AI can do. And now I think when there's starting to be business impact around AI, what it's generating, the threats that this new technology brings, whether it's the model, in the data, in the agent, wherever those threats are coming from, and we saw in 2025 more security M&A transactions than we've ever seen in history. higher spending rate than we've ever seen in history, higher share of overall tech M&A any time in history, right? We saw CrowdStrike do 2 deals. Zscaler do 2 deals. Palo Alto Networks do 2 deals. Check Point do 2 deals. Varonis do 2 deals. Everybody got into the shopping center because they understood that it's no longer wide open for AI. We're going to put some restrictions out of necessity. And that's a very different approach to the largest trade show for the industry. 
 

Daniel Kennedy: The scary part is it is Phase 1 of that. So all those deals are around security for AI, and they're all around -- I don't want to make them too simplistic, but we're talking about guardrails and implementation of guardrails in enterprise organizations. That is the fish emerging on land with 4 legs stage of this. 
 

Scott Crawford: I wondered how effective the existing tools that are being bought in those deals are for the future that is quickly emerging. You mentioned Palo Alto and the acquisition of CyberArk, primarily around privileged access management. But if you think about privilege management in the agentic era, agents are nondeterministic, meaning they can emulate reasoning and make decisions, take actions in real time. And we don't always know which way they're going to go, what decisions they're going to make, what tools they're going to use. Obviously, we wouldn't be using cloud cowork unless there was some predictability and limits to what we expect out of that, one of the adventures with open cloud, what they actually can. So now we're dealing with a situation where the use of privilege is going to come in line with actions that may be unexpected, nondeterministic. How do you determine a privilege set in any given moment for that set of actions fast enough to deal with that and not have it wreak havoc either by restricting access and privilege too much or leaving it far too open. If you're going to veer in the direction of allowing greater privileges to let agents do what they need to do, then you run some risk there as well, too. So the tools are going to have to evolve to this reality, which is really demanding in terms of real-time responsiveness, behavioral awareness, and the ability to deal with all this at speed and scale and a level of detail we have never really seriously contemplated in cyber or much of anywhere. 
 

Brenon Daly: And I would contrast what we're seeing in AI in general with what we saw in cloud. There was a transformation kicked off a whole new set of requirements, both in terms of technology and security, provided the largest exit in information security in history, provided the largest venture-backed sale of a company in history, and it's being wiz, of course. But once that -- there's nothing like the velocity that we've seen, to Scott's point, with what's happening in AI right now. And if you want to think about it in terms of volume, basically, right now, in terms of Gent AI security transactions, we're doing in a single year, the number that we did for an entire decade in terms of cloud security. because cloud security was a very sort of a long-duration transformation. scaled in, you put in workloads, and it evolved very slowly. Here, people are just spinning stuff up. Cloud was infrastructure. These were the building blocks and hey, you can build the blocks and then you do some other things, but we were fundamentally doing the same kind of things, just building them on top of different building blocks. So we already had some idea of how that was actually going on. 
 

Daniel Kennedy: To Scott's point, we are entering probably in 2026 and 2027, the golden age of blaming AI for things. Starting, we saw that if you brought up cloud, and I was thinking large cloud providers just had this debate with itself in the media that the AI was responsible for an outage. -- no, the responsibility of humans were supposed to be watching the AI. comparison to the show Silicon Valley and Son of Anton deleting production. So we are entering a stage whereas more is handed over to Agentic, we're going to see some interesting stories emerge. 
 

Eric Hanselman: Interesting in the sense of the Chinese curse. 
 

Scott Crawford: It's going to beg some issues as well. We were talking about privilege management earlier, but that links back to identity. Another aspect of identity there is authentication, how do you identify agents? Are they distinct from people? Do they have their own identity? Agents often operate within the privileges of an -- when they're operating as an assistant, generative assistants often operate that way. But the fact that they can take actions, use tools even when the human is not directly in attendance, begs the question, they need their own identity even if they operate within the privileges of another. What's the accountability for that when things do go south? What's the liability for it when things go south? These are presents that are going to encompass everything from standards bodies and at least consensus on how to implement and deploy to case law very likely and regulation at a national and international level. It's going to be good times in that field for some time to come, and we'll see a fair amount of that this year. 
 

Daniel Kennedy: I have sophisticated technology providers sitting across the table though saying, how do you think we should approach this? And I think we're -- in terms of the regulatory landscape catching up, we're a little ways away, but it is interesting. You bring up the point, Identity is going to move from something that's always been very important to something that's foundationally important. As with cloud because we talk about cloud-native security, going forward, it's going to be security architecture for AI native systems with identity and access privilege as a foundational element of that. 
 

Brenon Daly: Because at the heart of all business is trust, right? You can't transact if you don't have some degree of trust. And that's what makes markets, and I'm a markets guy. So that's why when we think about agentic commerce, Scott's question about liability is real. If I lose money on a counterparty trade, am I going to hold that agent responsible? -- am I going to hold the person who deployed that agent? I'm going to get paid back, but who am I going to get the check from? 
 

Scott Crawford: My mind wonders if that. 
 

Eric Hanselman: It's an important point. And that degree of trust is also going to translate into perceptions of risk. And so is that something where we start pricing in higher levels of risk where we don't have trust and who manages that? And by the way, how do you edge for that? In the markets, of course, maybe there are positions you can take that help with that. But the other big piece of this, the business side, is what's the typical way of dealing with a lot of risks is insurance. And now what's this going to do to insurers. We've already been through a couple of big swings of cyber insurance cycles with insurers charging in, realizing losses were substantially higher than they expected, then pulling way back, raising premiums. Now they seem to be back in market again. You can debate how many cycles this will get through before we reach equilibrium. But when you look at the broader technology risk, where are the actuaries that know exactly how dangerous agents are? 
 

Scott Crawford: Well, in the first macro cycle about this, first macro cycle, that's maybe not accurate enough. But during the one prior to the pandemic, we had seen that insurers had pretty much priced that overpriced. -- loss ratios were significantly lower than they were right after the pandemic about the time that ransomware became a lot more popular, people working from home didn't have the same approach to, I guess, you'd say, discipline as far as securing home-based systems as you could find in the enterprise, among other contributors to that. But once ransomware swept through the neighborhood, those loss ratios basically flipped from what they were before, going from roughly 30% to 70% in the first few years of the pandemic. Are we going to see something similar in the Agentic era? I don't -- I think you're right. I suspect that I don't know that anyone with an actuarial background, I haven't spoken to anyone specifically about this question, but how do you price that in at this early stage? And I'd be very surprised if there is a lot of consensus on that. 
 

Daniel Kennedy: I would guess it's going to go the way with ransomware, which the early answer was we're going to stop covering this. We're going to find a way to not cover this and then reentering the market later. But RSA is interesting. And we talked about identity -- in terms of the breakfast, there's probably a few other topics. AppSec, we're presently industrializing buggy code at scale. So we're AI slapping all of application security and then pretending that we have these new tools that will just clean it up automatically, which is a pretty crazy conclusion. And then we're part of energy now. So I have to note that there were a bunch of talks at RSA about IT/OT convergence. And Scott kind of had an interesting diagram, and I'm referring to a diagram on a podcast, which is cruel because no one can look at it. But the idea of it, it was interesting to me. It was -- we're talking a lot about convergence, but let's get real about the boundaries. Here's where they converge and here's where they don't. But the overall point is it's going from a niche topic to a pretty mainstream one. 
 

Scott Crawford: I'll diverge on that a little bit. You mentioned us being a part of the Energy division here at S&P. The dynamics of that are worth considering the fact that the AI investment is having such a dramatic impact on energy markets. Now politically, we see global situation having an adverse impact on energy markets and very likely some downstream impact on dependencies there as well, too. That's one of the reasons why we're lined up with this division is because those synergies are driving the global economy, global society and AI is becoming such a fundamental part of that. It's rapidly becoming on an impact that energy has had over the last century and will have for the century to come, probably not too expensive to say so given the trajectory of AI, but that's one of the reasons we're lined up with the division that we're in right now. Getting back to a point raised earlier, Dan, you mentioned the little adventure in automated agentic vulnerability discovery and management. That was one of the things that set off a bit of the AI scare trade in the cyber tech market. And Brendan, I was curious about your take on that because we've seen this sweep across other application markets and you were mentioning the pricing of application equities had been already pretty sky high. And so downward pressure was likely anyway. 
 

Brenon Daly: Yes. And it was -- so to separate out, first, we have the broad enterprise software freakout where no work is ever going to go through. We're never going to sign another license for an enterprise application as long as we live, just like all workloads are going to the cloud, and we'll never -- I mean, we're 20 years into this journey to. 
 

Daniel Kennedy: Build me a Salesforce, please. 
 

Brenon Daly: Like so -- and even now, I mean, I don't know, Eric, you might know, but like 15%, 20%, 30% of workloads are still on-prem. And that's multiple hundreds of billions of dollars every year invested in cloud infrastructure, every CIO directive go to the cloud and all that. And still, here we are 15, 20 years later. So when we think about the scare trade, right, first, it was enterprise software, the complete overreaction. And this idea that we're going to rip out ServiceNow -- and we're just going to have agents take care of my task. I have had agents pop up and offer to do everything in virtually every website in the lower right-hand corner of every website. And I have yet to see that could scale to an enterprise. But. 
 

Eric Hanselman: Come on, Clippy was wildly successful. 
 

Brenon Daly: You could handle everything. 
 

Daniel Kennedy: Agentic clippy yes... 
 

Eric Hanselman: It looks like you're trying to secure a website. 
 

Scott Crawford: What happened to Clippy? It looks like you're digging a grave. Is that a business grave or a personal grave? 
 

Brenon Daly: Yes. So Clippy got clipped and so did all the other software stocks. I mean we're talking about trillions of dollars of market value, this distant threat. And if we know anything, it's like technology doesn't work that way. And more importantly, technology budgets don't flow that way through technology budgets. It's not a case where you just cut off Salesforce. And you'd suddenly now agents are taking care of all of that, not just the Salesforce implementation, but the $6 or $7 for every dollar of license, there's $6 or $7 of services. Those services are much more vulnerable than the underlying license, right, the subscription, far more because agents can automate some of those services. So yes, there's vulnerability there. But you look at every pick of stock, ServiceNow is still growing north of 20% in the $1 billion run rate, right? Zscaler grown north of 20% in a $1 billion run rate. CrowdStrike growing 23%, a multibillion dollar run rate. This reaction that we saw was so out of sorts. And the idea that what Claude, thropix claw that the sort of the catalyst or the bomb that went off that destroyed all this valuation for -- in the information security was like Dan could speak more eloquently and knowledgeably about this than I, but it's a fairly limited code tool. 
 

Daniel Kennedy: I just don't -- it's looking for context of code vulnerabilities. And I always get caught because I read about this stuff before anyone reacts and I go, what sounds exciting. I'm looking forward to seeing that. And then I see a bomb go off and I try to explain it and it's going to replace all security tooling. How? Explain that to me clearly, like I'm from Gartner. And then this will replace SaaS. And it's not doing the same thing SaaS as you're doing. And by the way, I can run SaaS over and over again and not incur an API rate limit or a token charge or run up against anything. So are you sure it's going to do that? And the kind of discussion you just had, the nuanced discussion about enterprise software and SaaS is what needs to start happening because honestly, I made the joke, build me a CRM, build me a business analytics tool. It doesn't necessarily work that way. But there are a lot of people working very hard, enabling tools like Claud, to have the sort of layout of these databases and know how to programmatically ask for information and how to have persistence that it doesn't necessarily have natively. And if I disintermediate the UIs of these tools and the modular sales of these tools, what does that do to their positioning? That's a legitimate question. So sitting here in the market and saying it's going to knock this out completely, that's a silly discussion. But saying there are some incremental things going on here they're going to change the way SAS seats are sold. We're going to change the value of this tool or how you're using who is the front end of this tool. And does that cost different? Those are legitimate discussions, and that's where we need to move in terms of this entire thing. 
 

Brenon Daly: If you think about what's most vulnerable from Magenta threat, just look at where OpEx is the highest, right? So there's a lot of human involvement, right? Because remember, agents don't automate jobs. they automate tasks, right? Now a lot of our jobs are made up of tasks, some we enjoy doing, most we don't. So agents, as they automate tasks, you look for a high degree of OpEx. Where are you spending your money and not getting a return. And for years, SaaS has been basically a UI on a workflow. That's basically been a SaaS. And you could spin up an application, you could verticalize it, you can make it horizontal. -- the basic package that you were selling, and you were selling it very well, you got 80% gross margins, 30% EBITDA margins, 20% grower. Private equity could take you private if you made it to the public, they could recap you, you could find investment money no matter what, if you were a SaaS. And you basically had a front end and a back end. To Dan's point, yes, some of that does look a little more vulnerable than it once did. But there isn't a universal vulnerability. 
 

Daniel Kennedy: I'd say. And that's what makes it so difficult to discuss this topic because the immediate takes in the marketplace are so polarizing, you find yourself sitting there going, let's talk it back from there. But let's not pretend that nothing is happening either. Let's have a real discussion about this. 
 

Brenon Daly: Wall Street is not a place known for nuance. You talk about instead of buying stocks, you blow in and you blow out. That's how you approach when you trade, you got CrowdStrike, Palo Alto Networks down, Check Point for that matter, SentinelOne, you've got them down 15%, 20% year-to-date. Zscaler unfortunately delivered great numbers, down 30% year-to-date, right? So they're all coming into RSA here back to our show, and there's a little bit of a cloud hanging over that. And there's a definite cloud hanging over M&A., right? Imagine to go back to the deal Scott's talked about around CyberArk and Palo Alto Networks. Imagine you're the largest security vendor. You've gotten 15%, 20% whacked off you already this year. You're going to come to RSA. -- last year, RSA, Google had just done Wizz, right? $32 billion deal, right? That's $32 billion to put that into context, $32 billion in a single transaction eclipses some entire year spending in cybersecurity M&A, like a single deal topping an entire year. And then a couple of months after last year's RSA, CyberArk gets bought by Palo Alto $25 the 2 largest deals we've ever seen right around RSA because there was momentum. And now that momentum has swung the other way. And instead of blowing up to the top, we're heading to the bottom. And that's the way Wall Street kind of works. 
 

Daniel Kennedy: I've had a number of people reach out over the past 2 weeks, and they'll say to me, Dan, you're one of the rare people who has a Series 7 license, but also in application security. What do you think? And I'd respond that I'm invested in index funds, and I don't own any security stocks. Make of that what you will. 
 

Scott Crawford: Eric, I know you want to jump in here, but Dan, you mentioned the breakfast. The 451 Research Security breakfast does host a breakfast for anyone who's interested on Wednesday morning of RSA Week. And when this podcast appears on LinkedIn, I will -- one of us will gladly share a link to that breakfast. But I wanted to leave the audience knowing that the question we're going to close the breakfast with this year is, does AI pose a disintermediation threat to cybersecurity software and cybersecurity technology. So some of this conversation you'll hear continue and get more specific about the cyber space. at RSA Week. Come join us. 
 

Daniel Kennedy: It's changing so fast. You have different by that. 
 

Scott Crawford: That's why there's no slideware for this. It's a question. 
 

Eric Hanselman: I think we may anticipate more nuanced conversations about the various topics. But we are at time for today, and I don't want to leave or give away too much that you won't already hear at the breakfast. Scott, to your point, that's what I was about to drop in on our annual breakfast on Wednesday. And it will be great to hear what comes out during that and interact with the audience and see where this goes. But for this. 
 

Brenon Daly: We will miss you, Eric, you won't be there this year. 
 

Eric Hanselman: That is true. For the first time in many years, there is a conflict with CERAWeek with a date change with RSA this year. So I will be down in Houston, helping lead some nuanced conversations in Houston at CERAWeek. 
 

Brenon Daly: But that's true. 
 

Eric Hanselman: Different aspects on very different sides of the table, but also still talking about AI and cybersecurity. the conversation to be having that week. It has been great having you all on. I will encourage our listeners to check out the links in the notes and to see us in person on Wednesday at the RSA Conference. Thank you all. 
 

Scott Crawford: Thanks, Eric. 
 

Daniel Kennedy: It's a pleasure. 
 

Eric Hanselman: That is it for this episode of Next in Tech. Thanks to our audience for staying with us, and thanks to our production team, including Sophie Carr, Feranmi Adeoshun, and Kyra Smith on the marketing and events teams. If you like this episode, please subscribe or like us. I hope you'll join us for our next episode because there is always something, Next in Tech!