Insecure Agents — Episode 30: Mark Dorsi, CISO at Netlify
Podcast: https://insecureagents.com
Listen: https://podcasters.spotify.com/pod/show/insecure-agents/episodes/Ep--30-How-Security-Changes-When-Most-Product-Users-are-Agents-Mark-Dorsi--RSAC-e3hb7pb
Transcript (auto-generated, lightly edited for clarity).

Allie Howe: Hey, I'm Allie Howe. Welcome to another episode of the Insecure Agents Podcast. I am your host and today we are coming at you from RSA. We're here at Snyk Sandbox Space. And today joining me is Mark Dorsi, the CISO at Netlify. Mark, do you want to introduce yourself?

Mark Dorsi: Yeah, Mark Dorsi. I'm a startup specialist. I really enjoy helping out the community and chatting with all sorts of folks across the board. But this is at Netlify now for four years. I really enjoy it and you know, we're the place where builders build, so this is really exciting to be here in a moment where everyone's building everything it seems.

Allie Howe: Yes. And then we told this, like, close.

Mark Dorsi: Yeah. Thank you.

Allie Howe: No, it's awesome. Yeah, it's an amazing, like, crazy time. There's so many, like, new startups out there. I was walking the RSA floor and one thing that stopped me was just like how much the messaging kind of like sounds the same. And you were telling me that, like, this is much, very so much the Zero Trust era, wherever we kind of have the same messaging. Does this feel like that moment? To you.

Mark Dorsi: Yeah, exactly. Yeah. Everyone has, you know, something to do with agentic, which is fair. I don't blame them for doing that sort of thing. But it is very much the zero trust this, zero trust that, agentic this, agentic that. And that's great. And I think it goes back to a couple of my friends that have mentioned that while the agentic's interesting, I think it's sort of like this cuss jar I've been using for analogy for a bit. If you have to use the word agentic in order to make your product be amazing, you should probably rethink what you're doing and then figure out, OK, great, what is the actual product that we have?

And can it survive on its own in or out of the Zero trust slash agentic slash whatever era?

Allie Howe: Yes, absolutely. I think it's also worth noting that you are my very first guest on the Insecure Agents podcast. We did our very first podcast here at RSA a year ago with Mark. It was a live podcast. So what has changed in the last year that we first spoke?

Mark Dorsi: I just feel like a ton so much. Everything has changed. Yeah, I totally like I compared to where we were a year ago, which was a lot of unknowns, where we were all just sort of speaking in, well, we're not really sure what's going on. Like there's definite direction now, definite momentum around all sorts of different companies, all sorts of different places that are going on. And the main sort of thing is that folks are finally starting to find their ground on what will be important in the agentic era of security. And then we're starting to see companies shake out in that way. And a couple of notable things that are going on is that, of course, there are, you know, companies who are developing software that isn't necessarily for security teams, but they're still interested in developing some tools and some other sets.

And there's some other folks that are doing a great job around trying to really lay into the overall security angle and really try and find, you know, what everyone might want to call it, synergies within the industry, whatever that is, so they can actually provide true value to those teams that are going to use them the most. Before we didn't have any of that. Before it was just a lot of subject, subjective information that may or may not have been there, some anecdotal references. Now there's true signal, there's true capability, and everyone has become a builder just like myself. So I went from coding really just maybe an hour or so a day.

I'm fully coding 6 to 8 hours every day now, which is a complete swing and change just as the rest of the company has gone. So we went from a company of some doers to all doers overnight. I think we're seeing that. So the shift is incredible. That's the truth.

Allie Howe: Well, it is incredible that everyone's become a builder and you as a CISO coding like 16 hours a day, that's amazing.

Mark Dorsi: Yeah, exactly. Yeah, The number of PRs I open, you know, it's that thing, there's nothing all of them overnight, right. So, you know, just in trying to fix different issues with code, build out different tools, sets, whatever it might be, just to ensure that, you know, for Netlify, we're doing the right thing by the customer at all times. And so if everyone's building and we're being secure about it, it, it'll be amazing.

Allie Howe: Yes, absolutely. Yeah. I feel like the big change from a year ago, June touched on is the fact that like a year ago it was like all Gen. AI chat bots and now we've like slowly crept up this autonomy curve where now we've got agents actually doing like meaningful work for us.

Mark Dorsi: Oh yeah, running in the background like, you know. 100% of the time, yeah. So I was, I was out to lunch with a friend today and I was speaking about like all these little agents that I've created in the background and some of them are like custodial. So they go around sweeping up after me. So we haven't done it. Go out, create yourself an agent that sweeps up after all your little file systems. Just tell it where you want it to be, how you want it to do what you don't want it to touch and tell it to do it every day. It just did it.

It was fantastic, right? And so all that sweeping up all the way to really impactful things around, you know, what is the next version of bug bounty look for look like at a Netlify? What is the next version of customer support look like in a world where we have, you know, where we're headed to 100 million developers that want to use Netlify, right? The number of agents that are going to want to use Netlify and companies similar to that is at least two orders of magnitude bigger. So what does that look like in that world? And then that's what we're building for today. And so everyone is building in some agentic fashion in order to take those on.

Allie Howe: How are you hardening Netlify for a reality where I mean most of its users are agents, not just humans?

Mark Dorsi: Yeah, Yeah, exactly. So I mean, that's a real, that's a that is a reality that we're grappling with and that's where like we are bringing products and services in house that we might have never thought of before. So when you're thinking about agentic and you're building, there's always the the line from Jurassic Park, one of my favorite right, which is like, you know, you were so focused on whether or not you could as whether or not you should was actually what you should do. What we're finding is that many of the tools and the way that they're priced and the overall models really breakdown at the scales that we're Speaking of.

And when you bring in agentic, it's an entirely new way to think about how companies and services need to be able to interact, especially in the agentic era. And when you think about agents needing to be able to speak with agents, are they really going to want to open up a traditional ticket? Are they really going to want to communicate with you in the same ways that they always have before? The answer is no. The agents and what's interesting for like companies, I think like a Netlify is like it's all CSAT is based on speed. How fast can you can you respond? And in that way when a customer writes in, you know, there's some period of time that passes well as agents try and do the same thing.

That period of time that you take to respond is a very real set of friction for that agent in a very real way. And it will very soon decide that you respond too slow to to consider your service to be viable and it will switch to those services that can respond in a timely manner. Security issues, support issue, IT issue, whatever that is, it will begin to pick and choose at speed and at scale. So if you can't scale with it, it is a big problem. So figure out how you can speak quickly with agents and and do it overnight.

Allie Howe: Yeah, it's a great point. How do we get agents to actually do like meaningful work and be able to respond quickly? And I think that means giving them the capability and the autonomy to be able to do so, which is challenging in its own right.

Mark Dorsi: Yeah, I know it totally is. And so those capabilities, right, I think it comes down for most companies to whether or not you've, you have trained on enough data over time and you or you have enough trainable data over time. So I can give you an example of our support and ticketing queue meaningful tickets. We have somewhere north of 100,000 meaningful tickets over time. And what we're doing today is that we're actively training against those meaningful tickets over all time. And we're actively playing a game whereby inbound comes in a ticket. We evaluate that ticket against an existing set of skills. Those skills, if one doesn't exist, it attempts to create a new skill, right?

So it never happened before. It flags it as a novel case that is then used as a response back, sort of like a a pseudo response back. And then sort of kind of a judge waits for the human to respond and then it compares the two responses and then backup dates the skill itself in order to have continuous improvement in time. And when you look at it in that way, if you open up now the door to agents coming through and asking questions, a lot of those questions will be the same. Yes. The issue is that those novel cases where it's never been seen before will be so time consuming.

So in a world where you used to have maybe 100 tickets in your queue, let's say there's security issues, customer support issues, whatever they might be, that just went up by two orders of magnitude. I mean the number of identities that are machined, the human is 101 today. So we're talking about at least a 2 order of magnitude problem for companies. And if you're not looking to really automate that and think about that from a scale perspective, you should right away. Otherwise the number of humans you need in order to support the machines, there is no number that will you'll be able to do that.

You will not be able to keep up. It's just not a thing.

Allie Howe: That's so true. And that's a recurring theme that I might I've heard that we need to be building for agents, honestly, just for humans. I had a really interesting conversation with James Cowling, the CTO of Convex on this podcast, actually talking about how with open claw, the Peter Steinberg had him on his podcast and Peter was saying like, or they were telling Peter like your Convex bill could have been lower like the way you did Convex with like vibe coding. It's not set up as efficiently as possible. Like would you like us to like, you know, create a PR to like fix it and say, well, why would your agent create a PR when you just like take the feedback from like the looking at the code and just like have the coding agent fix convex for me when it's like doing the vibe coding, like why would I want to as a human go look at it?

I don't know convex as well as you do. So why don't you just, you know, make that, you know, loop continuous? Like what building blocks are you building at? Like Netlify, either for like developers when they're creating on Netlify or like when agents are using Netlify. Like what building blocks do you maybe making for agents that you weren't making for humans?

Mark Dorsi: Yeah. No, it's really interesting. So today, of course, we have like what we call start with agent runners. So anybody can come in and they can build today on Netlify, right? So before you come to Netlify, you'd bring in some code that you'd already developed and then you deploy it to us today. You can just start right there. And that's amazing. But those interfaces, right, will become not antiquated, but you need to be able to handle the sheer speed of what an agent might be able to communicate at, right? And that's where you need to have be at A2A whatever these protocols that will, that will sort of bear themselves out.

I think it's really just going to be another API call. I always challenge folks, I'm like, does anybody remember Restful APIs? They're all just APIs again, right? So A2A, MCP, blah, blah, blah. It's all just MCP, right? It's all just API. And so that ability for an agent to be able to prove its trustworthiness from like a communications perspective, have the ability to show that it has an identity and be able to then communicate at a proper level to be able to receive either support and or create things, right? That's always being proved out right now. So over the next six months or so, I think we'll really start to see what that looks like and then we'll really understand what's there.

We're already receiving support tickets from agents working on behalf of the human right. We're already getting that, which is where our our concerns around scale are. And everyone should really be frankly concerned, right? The sheer number of people who have adopted open claw overnight, right is, is really leading into that. So I would I would imagine that in the next 6 months, the number of humans opening tickets not will go to 0, but will approach zero at speed, and the number of agents opening tickets will actually balloon and go through the roof. As far as that's concerned. It is the exact same thing for folks who want to develop.

So the number of human developers is still in the background, still the master of Puppets, whatever you might want to call it, what'll be the agent that'll be there trying to develop. And those interfaces that are needed for that to happen have to be developed today, deployed today so that they can be used by the agent today and.

Allie Howe: That makes sense when you think about like fraud. I know because the humans can like, you know, use support systems fraudulently and submit like the wrong cases. Like are you seeing more of that with agent? Are you more concerned?

Mark Dorsi: Yeah, it's fantastic. So the number of phishing sites that are attempted to be deployed and any given day is fantastic. So people are like really exercising what's going on there. That said, you know, we have some fairly robust mechanisms in place that are able to pick that off. We try and be really careful about allowing anyone anything like human or machine to be able to access the platform. So we don't restrict in any way. We restrict based on behavior. So we take a look at overall behavior of an agent of a human. We don't really care what it is. And then we can then choose to allow that entity to continue to work on the platform or not.

But that's the way we designed them from the ground up. So we don't actually care that an individual or entity came in through a particular door. We care what they do once they're inside the house, you know, so they did they play well with others or not, right? And so that's how we take decisions and oh, by the way, most of the tool sets that are available to us at that type of scale, the issue is that they become what I would call cost prohibitive or margin prohibitive in the grand scheme of things. So when you're trying to have, you know, a company of your own, you're trying to do whatever you are, of course, margins are really important, right?

You made some money last month, you actually made money. It's not EBITDA, right? So not it's not BS dollars, it's actual dollars that came in the door. So that the point being is that most of those tools we found don't scale with us, not necessarily because of capability, but because of cost model. And in the world of Agentic, you have to be really careful with what those cost models look like. Sure, everything is trying to use up a token, but then you have to start thinking about the scale and what that looks like when you went from, you know, 100 customers trying to use your product to 10 million to then, you know, a billion customers, right?

But many of them are agents, and many of them are humans. Likely the vast majority will be agents in the near future. Very bizarre.

Allie Howe: And super bizarre and like if you're a CFO right now, like how is this changing like pricing models like seat? Based

Mark Dorsi: seats are dead, right? So seats are dead Long live seats. They will exist in some theoretical model somewhere in the background where everything is really there. I really think it's going to end up back in the days where it reminds me of course of CPU bandwidth compute at the end of all of this. And how that will be sort of put together in a number of tokens where you're able to then negotiate kind of enterprise contracts. And those enterprise contracts are along the lines of that, which is, look, we're going to give you so many quote, unquote tokens across everything you want to use and go for it.

In a world of agents, thinking about having seats just means that there are people within an organization and or machines within an organization that can't use your product because of a gate around a seat, right? So instead, everyone's looking at the sort of arcade model, which is, look, you want to play this game over here. It's a quarter, that one over there is four quarters. Pick whichever you want, but this is a bundle of quarters you brought when you came in the door. Enterprises don't really work that way. Enterprise needs to understand their cost. We need to be able to build those models out for them and then they're able to then sort of bring and use those dollars over time, right.

And when it comes to renewals, then we can true up. We can talk about what, you know, the next contract renewal looks like, and we can continue to keep them as a customer. We'll see what all this bears out because all of that is in flight across the entire industry. We're just not sure where it is today, right. So, you know, seats are dead, Long live seats.

Allie Howe: Yeah, we'll see what happens. I don't know. Yeah, we'll have to do another episode a year from now. Let's say we'll see what where we landed on the pricing models.

Mark Dorsi: Yeah, exactly. So much is changing.

Allie Howe: Yeah, I know. It's wild for sure. I know. One of the things that I really appreciate about you is you really care about making the world a better place. I mean, you think about, like, all the startups that you talk to all the time. You're very generous with your time, Help a lot with the startup community. Who's doing that right now, in your opinion? Like who's changing the world for the better?

Mark Dorsi: Exactly. So it is true. I, I'm here to try and make the world a better place. We all only have so many skills that are out there. And for anybody who's watching and you want a little bit of time, no worries. I normally post like a link out on LinkedIn. You can reach out directly through LinkedIn. I'll be happy to chat with you. I allocate 5 hours a week to have conversations just like this one, but with the latest founders, VCs, whoever it might be. We're trying to find product market fit. They have questions, whatever that might be. And what I see out there in the industry right now is truly a set of founders who are trying to make the world a better place by enabling companies to be able to use agentic safe and secure.

And that's really what I want to see out of the world, which is, look, we have the best intentions. We're trying to figure out the right things to do. Not all ideas will hit. A lot of them will be about timing. Of course, some of them are not going to be great, and that's OK. But at the end of the day, but we need those founders to give it a whirl so that we can open up those doors. And I'm I saw a lot of that this week. This week I had this north of 20 conversations. Today's Wednesday, So Monday, Tuesday, I don't know, approximately 10 conversations each.

I had several today. So across all of those time slots, you have amazing number of humans who just want to try and do something, you know, amazing for the world. And that's really what it's all about South for those founders who want to try and make the world a better place. I wouldn't say today across all founders, this is my cuss jar sort of soapbox, which is create yourself a soapbox, sorry. Create yourself a cuss jar and promise yourself that you'll speak with 300 builders and out of those 300 builders, every time you use the security term, put a dollar in the cuss jar. If you do that, you'll be able to hear today's builder because they're not yesterday's builder.

These are not traditional engineers. These are all the folks who just want to build something amazing to really change the planet in some artistic, creative way. And you should be able to go out and use any number of platforms to be able to build those things. Our request is simple. Help humans build things help machines build things safe and secure without them ever having to lift a finger, because I'll guarantee you right now, every prompt I've ever seen by ending one of our customers never ends in and make sure it's safe and secure. So we need they need your help.

Allie Howe: We need your help, that's for sure. We. All need help to ship securely and to build secure by design systems.

Mark Dorsi: Exactly.

Allie Howe: Yeah, it's like that convex example of like, you know, this agent should just know what to do because like told to do the right thing. Like in the background. Yeah, it's all happening in the background, like in the agent harness now, like security is just assumed.

Mark Dorsi: Yeah, exactly. But it's, but there's so many holes as we've seen throughout the industry and there's new ones every day. So we really need the help in order to understand like how do we secure systems right and and do it at speed because a machine can create an insecure thing just as fast as human can. It just does it at scale,

Allie Howe: yes. Absolutely, and does the same pattern over and over again.

Mark Dorsi: Exactly. And it will just repeat the same error. It's very interesting to even use like a platform like Netlify with Claude Code will continue to try and use a feature that has been deprecated for three years, a set of functions that have been, you know, dead for three years. And you have to be very explicit about it, like why are you trying to use this? Use something else. It will work if you just use this, use V2, use V5 like whatever it might be. So there's all sorts of problems that can come with agentic, right? It just assumes the best it.

Allie Howe: Sounds that it wants to please you.

Mark Dorsi: It wants to please you and it will do things like create an S ton of mock data that never had anything to do with your platform or anything else. Why? Because it couldn't figure out how to get the data, so it created a bunch of fake data and pretty soon you're looking at. You know, Sally and Bob in your, you know, support queue and you're like, who are these two folks? Turns out it was all made-up. You never actually had a real thing. So yes, please help us make things safe and secure. We'll all be amazing for it.

Allie Howe: Yes, absolutely. What are you trying to get out of RSA this year?

Mark Dorsi: Amazing conversations with folks. This is an amazing conversation. Thank you so much for having me. Again, this is wonderful. But mostly it's the conversations. You never know where a conversation will take you. We only have so many skills as individuals and one of them is just your gift of time. Please take the time to meet people, speak with them, and that sort of thing. And that's what I'm trying to get out of as well, is just have those moments with folks because you never know where a conversation will take you. The market and the world, frankly, right? Just think about, you know, think about Anthropic's February, right?

Who would have said it went from 10 million to 100 million last year to 6 billion in February and already eclipsed in March? Like, you know, like, what conversations did they have to have? Who had to open doors for them to make this thing happen, right? And if those folks were not generous with their time, something like, you know, their pocketbooks and what they had going on, it never would have happened. I love that story between Jensen and Elon where Jensen is like, oh, my gosh, you have somebody. And and of course, you know, Elon's like, yeah, I have somebody who could use one of your processors.

You know, turns out it's a nonprofit, right? He's like, right. Well, that was OpenAI, right? So you never know where that conversation will take you, right? And so here we all are, standing on the shoulders of those folks who are willing to have a conversation before us. Let's do more of that.

Allie Howe: Absolutely, yeah. I think everyone's got a story to tell that way.

Mark Dorsi: That's awesome. For sure.

Allie Howe: Well, thank you so much for coming on the podcast.

Mark Dorsi: Always a pleasure, love doing this sort of thing. Thank you so much for having me and builders build. Let's make the world a place you know better. Place one app at a time. Let's start there.

Allie Howe: Absolutely, yes. Let's do it. Thanks, Mark.

Mark Dorsi: Likewise, Thank you.