Insecure Agents — Episode 66: Ed Sim, Founder & General Partner at Boldstart Ventures Podcast: https://insecureagents.com Listen: https://open.spotify.com/episode/4jb1FQTzxgROYMnjKKv7ua Transcript (auto-generated, lightly edited for clarity). Allie Howe: Hey, welcome to another episode of the Insecure Agents podcast. I'm your host, Allie Howe. And today I'm super excited to be here at AI Security Summit in San Francisco. I'm joined here by Ed, Ed Sim. I'm super excited to have him here today. Ed, how was the day? I know it's kind of the end of the summit. How was it for you? Ed Sim: It was fantastic. You know, it's kinda cool 'cause I was here last year, and you know, it was kind of early in, in the days, and now when you pick up the paper today, everyone talks about insecure agents, so great name for the pod. Allie Howe: Thanks. Ed Sim: And, and what a time to be alive right now. Allie Howe: It is. Yeah, so when I started Insecure Agents, I started it 'cause I thought, okay, people need to be educated on the fact that agents are inherently insecure. Ed Sim: Yes. Allie Howe: And I think people that are starting to pick up their Muse agents and Instinct, they're, they're trying to figure this out if they haven't already. Ed Sim: Yeah, my Muse agent is bothering me right now already, I was telling you earlier, so. And then when my wife uses it too, she starts asking me about security, which for the first time ever, it's amazing that my wife, my father-in-law are asking me, "How do I secure these things, and should I go do it?" Because no one ever thought about asking that question. Allie Howe: Yes. No, there's so much to get into. I know, I know one couple, one of my friends, is a, a... The guy is a, like a very technical engineer, and then the, his wife is, like, non-technical, but he needed a solution for their personal agents to communicate. And he saw the hugging face and said, he was like, "Oh, I'll just create a message board," 'cause obviously that works. And it's- Ed Sim: I love it... Allie Howe: I know. So- Ed Sim: Brilliant. Allie Howe: Yeah, brilliant. Yeah. And he's like, "Yeah, and so they talk, and, like, sometimes I don't have to answer questions because the bots just figure out for us. Like, dinner's booked. I don't have to remember." And it's like it's, it's better. Like, life's better that way. Ed Sim: Of course. Allie Howe: And I was like, "That's amazing." 'Cause I feel like we were also talking about before this, that personal agents have become kind of this, like, test bed for figuring out, like, what works, what doesn't, and then you can start to figure out those patterns of what works. Like, maybe a message board or identity and access control for your agents. I know with the Instinct incident, there was one where I went and booked a resi- reservation because it's- Ed Sim: And it, and it kept hitting it every second, right? I mean, agents are very good. Yeah. They're like, they're like a dog or like a retriever. Yeah. They'll keep doing the same thing over and over again. It might, might be the brightest bulb in the room, but I'll tell you what, they're relentless as hell, right? Now, that's a lot to talk about, 'cause I think we at Keycard also announced some big architecture and platform and partner story right around agents. Allie Howe: Yes, absolutely. Yes. There's so much going on at Keycard right now to secure agents in that way 'cause I think- With Keycard, you can be able to revoke access with, from agents. So if you're able to spot, okay, hey, my agent is making 200 API requests an hour to Resy, and it's on its way to getting my account banned, because Keycard holds the access for that agent, you can spot that happening and then be able to prevent another credential from being issued to that agent to be able to perform that action. Or really, a better way to say this, prevent a access to be issued to, like, a Keycard gateway. So, like, in that way, the agent nor its deployment holds the secret whatsoever. Ed Sim: I remember when Ian was talking about this, and Jared and Creager, I think we funded... The company was started, what, two, two and a half years ago? And they were talking about how e- every agent should have its own identity. And I remember, like, walking them into some banks and things like that, and the CEOs were like, "That seems very complicated for what I need now." And, and people thought, like, maybe the humans, you know, give them access. Let the agent have the human access and, and, and run amok. And it's really amazing how far the world has come, because even the last three months ever since the Hugging Face incident came around saying these agents are live and they'll keep doing the same thing over and over to realize a goal, and then also because of these consumer bots and how easy they are, and because of issues like that, I think that right now it is literally at the top of everyone's mind right now. You really can't use a consumer product. You can't use an enterprise product. And in fact Jamie Dimon yesterday on his earnings call, or two days ago, said the, the CEO of JPMorgan said, "I'm gonna really control what agents come in and touch my back accounts and which don't." So that at the end of the day is a really massive turn and it's very exciting about kind of what the world's gonna look like in the next few months. Allie Howe: Yes, absolutely. I think there's a ton of interest in this topic now, especially with, like, the Hugging Face incident and everything. Consumers, but also businesses of course, are like, I really need a solution to this problem. And that really resonates because I had Andy McMahon who's an AI engineer director over at Barclays Bank was telling me, like, his biggest problem is this, what he calls the fishbowl problem, of like, okay, one vendor he brings on will have its own, like, identity solution or whatever, or own gateway. And then so will another vendor, so will, like, a legacy product that he has. Like, it's a 200-year-old bank, so they've got all of these, like, interoperability problems from, like, the vendors they have, the technologies they have already. And giving an agent a stable identity across all of those different interfaces is what he said is the optimal solution. That's the fix in his mind. Yeah. And that's something that Keycard is very, like, poised to deliver. So I'm super excited to, like... I feel like it's, like, really, like, our time. Like, I feel the energy. I feel like-... everyone really needs a solution. I'm super excited to, like, help bring that to market. Ed Sim: I love it. I love it. I think you make a really good point because there's a period of time where I b- I still think every SaaS vendor has its own agent, right? And then, and it's very hard to, to know kind of what control points you have from every single ServiceNow or Salesforce or Slack or whatever, right? And I think that, and every model will have their own kind of agent security because they have to have it. But ultimately, I think the last thing people want is this sprawled out everywhere. They want one place for a central policy, and I think you have to understand intent as well about how these things are working in other places so that when they're working across different applications or databases, you can get the full understanding and capture all of that in one place, right? And, and without that, I think it becomes very difficult to secure kind of your agentic workforce that you're offering to your employees or c- consumers. Allie Howe: Yes, and it's really nice to be able to have, like, fine-grained access and policy for, like, okay, maybe, you know, Allie's permissions via OAuth or, like, Notion say that she can read all of these, you know, databases or pages. But if, what if I only really want this agent to be able to read just one database within one page in Notion? Or in some circumstances without- Ed Sim: By the way, for one job. Allie Howe: For one job. Yeah. Yeah, yeah, yeah. Or, like, one task. There's no forever access. It's inherently dynamic. Ed Sim: Yeah, AI just changed that You know what? Yeah, you know what's interesting? So, I'm just kinda going back. Yeah. We launched our fund last year -... our Fund Seven Boldstart Fund Seven, and I didn't wanna use the word AI. And I kind of, as a joke, said, "The fund is focused on the autonomous enterprise." A- a- and so the idea was not that we, we would have the autonomous enterprise, but the aspiration of having a one-person company doing a billion of revenue with no one else and just a bunch of agents running around was kind of the goal. And what I was really trying to say, and my firm was trying to say, was that we have to rewire the whole stack. The whole infrastructure has to be rebuilt for a world of agents and for scale. And I'll give you an interesting stat. If you look at... I, I just didn't know how fast it would come because when you look at, like, OpenRouter and the amount of tokens consumed by humans versus agents, for the first time ever, it crossed in February, where machines outnumbered humans in terms of token consumption. And then I think by September it was 7.8 times. Allie Howe: Oh my gosh. Ed Sim: And so it, how long will it take before we hit 100 times, right? And so when you look at that scale, so imagine that scale and, like, the thing that, that Keycard's building is that first of all, you have to support that scale, then you've gotta do dynamic access at that scale, a millisecond scale with a thing that's growing 10x, 20x year over year. I mean, this stuff that was built for Okta and humans, where they have single access forever, it just was not built for that. And I think if you take that across every single part of the infrastructure, everything has to be rebuilt, which I think people are finally understanding that has to happen, which is why I think we're in a good spot. Allie Howe: Yes, absolutely. And that reminds me of the OAuth Islands problem that Karl McGuinness was telling me about. Ed Sim: Oh, yeah. Allie Howe: Where you just, like, you know, generate... So you log onto something to Okta, and then you, or not Okta, but with OAuth, and then you've got this, like, OAuth token laying around, and it's hard to know, like, where it exists, like, can I revoke it? And for your IT, and like, a large enterprise, like, they have no insight into that, like, OAuth Island, or where your, like, personal, like, OAuth token lives. Ed Sim: Oh, yeah. Allie Howe: So, like, new, like, protocols, like ID-JAG and stuff can help with that, but just, like- Key, the Keycard like solution or problem space is so much better fit, I think, to like be able to re- revoke access or issue access versus like having these like long, broad scoped credentials just lying around. With your fund, like what are you seeing right now in terms of like what's working and what's not? Like, I feel like it's such a hard, like problem space to be in, like to be... It's hard to be a founder I think these days. Yeah. Because models have gotten so good and people, there's so much customizable software to be built. Like what are the problems that you're seeing that are worth solving? Ed Sim: Yeah, it's really interesting. I, I mean, well, first of all, to frame it, we're inception investors or partner when founders have ideas and so like company formation, just like a Keycard. And, and so I have to invest in things that are a year to two years ahead. So if today someone walked in and said, "Hey, I wanna solve the agent security problem," I'd have a very hard time doing that because when we funded Keycard, it was two years ago when people kind of scratched their heads and said, "I don't think this is gonna work," or, "This is too much of a solution for the problem." So we need to find founders that are seeing the future. And right now I kind of have to say that under ranked software is really, really hard these days because of things happening so fast. And then on top of that, I feel like funds have gotten so big that there's so much money getting plowed in from the very beginning at inception in these security companies. I mean, I don't know when you see- if you see it, but there's multiple 50 to $100 million announcements or rounds from the very beginning. And I think part of it is a couple reasons. One is because king making of companies has moved earlier. So people feel like if I throw $50 to $100 million and announce it, we're just gonna win. Allie Howe: Oh, yeah. Ed Sim: Which is, which, which I don't think necessarily is the case, right? I mean, in fact, I think it could lead to tons of burn and there might be some cases where it works but I think secondly I think also people are used to the old playbook, and the old playbook right now is I talk to 50 CISOs-. They have a problem today. There's a, there's a hole I fill, and I go in and sell it and get to 10 million of ARR. But the problem I have is that that's not seeing the future. That's just taking something incrementally and making it better. And so once again, with Keycard is we're seeing the future. We're actually transformed something two years ago, now it happens to be right place, right time. But if Keycard was started today, you'd probably already be behind a number of companies, right? So what have we funded lately? I keep telling myself not to do any more security 'cause I feel like it's so crowded. But the last thing I funded was a sovereign kinda cyber Neo lab. So the idea is like build your own cybersecurity model have lots of small... a family of small models that you own. And you can actually access it through an API, you can deploy it on-prem, and you can actually run all your own contacts, all your own data through it, and constantly that model will get trained and get post-trained and get smarter and smarter for your enterprise. And it's not to say that it wants to replace the Anthropics or the OpenAIs, but it wants to partner with them. Why spend, you know, all of your tokens and security on the frontier labs, which are very expensive for 100% of the dollar? So why not just do it for 10 cents, spend the other 90 cents at a much reduced cost for a much more focus. Two other things we looked at is hardware-related. So, I kept asking myself and why were trillions of dollars being spent in data centers, but no one- there's no security company that owns that. And there's a lot to secure, by the way. And so we funded a company in that space called Lava doing AI data center security and geo cloud security. And then I also looked at something else which we just funded in stealth, which is solving the physical AI problem. So I'm thinking about chips and firmware. I believe that AI's gotten so good that more people are gonna build hardware, and it's gonna be easier and easier to build hardware, and there's gonna be intelligent hardware interconnected kind of online. And then what's gonna happen is you'll find all these vulnerabilities, but you can't constantly update a car or a drone or a machine 'cause things will break. Yeah. So how do you provide real time prevention at the, at that layer so that when things deviate, just like the way Keycard would, when things deviate from normal behavior, you can block it. And so those, those are the things that I'm looking at now that I feel are two years ahead of the curve, the same way that Keycard was two years ago. Allie Howe: Yes, I feel like there are so many opportunities in like the physical AI space. Yeah. Like I heard about one that was like something that helps like project if like grapes on a vine, like for wineries, like to be able to like use AI to dynamically like know the right time as to like tend to the grapes or whatever. Or like I guess like farming and w- I mean, I just never thought of that 'cause I'm always in the software, like tech world, but there's so many hardware applications. Ed Sim: There's tons. And like, you know, our firm itself, we probably have a third of our investments in cyber, and the other two thirds are in physical AI, AI infrastructure, and a category we call the weird, right? Yeah. Just kind of like... 'Cause there are a lot of things that just don't feel like, they feel like the Land of Misfit Toys, if you ever watched Rudolph the Red-Nosed Reindeer. And so we like misfit toys 'cause these misfit toys are meant to become like the next Cabbage Patch, you know, doll or like Labubu or Laboba, whatever, you know, those little things. Allie Howe: Oh, that thing. Ed Sim: Yeah. Yeah, yeah. But we wanna catch them now like non-consensus or non-obvious types of things that are very early before they, they, they get there. But, so we, we were in a foundational lab for robots called Generalist AI about two and a half years ago. Two guys from Google DeepMind, someone from Boston Dynamics, and now that thing is doing some amazing things. Allie Howe: Wow, amazing. Ed Sim: So it gives us a broader view of just not just security, but it's, it's helped us to look at things that combine both together. Allie Howe: Yes. I feel like also another struggle for founders in this space today is just, like, how much noise is out there. Oh. Like, who, what, what do I listen to? Who do I listen to? What do I act on? As we wrap up, like, what do you think people should be paying attention to now in terms of like how do they... When do they, so when do you think someone should decide, okay, I'm gonna let this influence my decision or not? Ed Sim: I think that is an awesome question, and there is so much noise, especially online. A lot of these videos are fake videos, by the way, where people have, like, fake kind of followers. You, you'd probably know that. I was just at a meeting with a founder. We just had a board meeting, and as I was walking back, the founder asked me about, he goes, "What are some an- anti-patterns to think about in terms of building startups today versus yesterday?" And I gave him a couple, and, and like I was just thinking about, one was that y- you used to actually have founders selling everything, you know, the first 10 deals the founders would sell, and then don't hire anyone, hire two reps and blah. But I think the world is moving so fast now that founders are forced to hire a head of sales much earlier before they close the first time. So, so I think speed to market becomes much more interesting. Two is I think technology. There used to be a moat of three to six months, but right now the only moat is speed, it's innovation-... and actually it's knowing how to sell. So I think that's an art, so that goes back to point number one. Point number three would be there's this idea that you always land with a wedge. Like, land with a wedge, get to 5 million of ARR, and then, then put product number two on. But you don't have the luxury of time to do that anymore. You al- almost have to have, be a multi-product company. Once product number one is out there, you already have to have two or three or four products out there. And it almost sounds counterintuitive about focus, but AI allows you to build faster, and every space has multiple competitors in it coming down the pike. You're gonna have 25 to 30 competitors. So I think you have to think about at what point in time do I, do I add more products to my family? And I know that we have this infrastructure platform where we have almost different apps with different ways to go, but that's a very similar model. So I think when you combine all those things together I think infrastructure's so hard to build, but you're expected, once you kind of get out of the market, is to move faster than ever before with more things than ever before. And I think that's really disrupting a lot of things right now and it, and it's actually the founders that can navigate that and multi-thread that I think are the ones that can really win, and the founders that just stress out or can't do it or, or are stuck in the old ways are, are not gonna make it. Allie Howe: Good advice. Thank you so much for coming on the podcast. I really appreciate it. Ed Sim: Thanks. Allie Howe: Thanks for your time. Great to meet you. Ed Sim: Thanks for having me. Likewise.