Name: Ed Sim, boldstart ventures
Uploaded: 2025-10-23T17:00:00.000Z
Duration: 19 min

Clips
News
More from The AI Security Summit 2025

Ed Sim

Founder & General Partner

boldstart ventures

Agentic orchestration emerges as a new layer in AI security integration

The rapid evolution of digital systems is forcing security architectures to think faster, act smarter and adapt on the fly. As organizations embed artificial intelligence deep into their software pipelines, the demand for seamless AI security integration has become impossible to ignore — a safeguard against data breaches, compliance gaps and model misuse that could derail innovation.Boldstart Ventures’s Ed Sim talks with theCUBE about why AI security integration is critical for safeguarding innovation and data

play_circle_outline Protect AI's successful investment story and acquisition by Palo Alto Networks.

play_circle_outline Securing AI in Enterprises: Navigating Vulnerabilities and Challenges in Model Protection

play_circle_outline Transition from traditional finance to machine learning and coding security risks.

play_circle_outline Encouragement for entrepreneurs to build AI-native solutions in the current tech landscape.

Info
Transcript

Ed Sim, boldstart ventures

Ed Sim

Founder & General Partner boldstart ventures

In this AI Security Summit interview from San Francisco, Boldstart Ventures founder & general partner Ed Sim joins theCUBE’s Jackie McGuire to break down why securing AI must start with the code and extend across models, agents and infrastructure. Sim revisits Boldstart’s early bet on Protect AI – funded pre-ChatGPT and later acquired by Palo Alto Networks for $700M+ – as an early proof point that AI security is real and enterprise-scale. He details the evolution from model registries and model scanning (e.g., checking downloaded models for vulnerabilities) t... Read more

explore Keep Exploring

What were the developments in AI and ML around 2017 and 2018, and what impact did they have on enterprises and subsequent acquisitions in the industry? add

What problem were they addressing related to the use of AI models and their security? add

What is the individual's experience and perspective on the transition from finance to machine learning, particularly regarding security practices in coding? add

What advice would you give to entrepreneurs regarding future startup opportunities, especially in light of recent platform shifts? add

bolt Powered by CUBE AI

Ed Sim, boldstart ventures

Jackie McGuire

>> Hello CUBE community. Welcome to theCUBE's coverage of the AI Security Summit. We are in San Francisco, California at the Westin St. Francis, and I am here joined by Ed Sim. Ed, welcome.

Ed Sim

>> Thanks. Thanks for having me.

Jackie McGuire

>> Ed, you are the founder and general partner at Boldstart Ventures.

Ed Sim

>> Yep.

Jackie McGuire

>> And we were just talking before we started recording about a whole bunch of investments that you've made. So you've become a little bit of an expert in the space and AI, and you were saying that you actually got into investing in different AI technologies quite a while ago, right?

Ed Sim

>> Yeah, I feel like the first wave of AI was more like ML back in 2017, 2018, and I think that set the foundation for every large enterprise now to get their data prep ready. We ended up funding a company called Protect AI a year before ChatGPT even came out. And the founder, Ian Swanson had come out of AWS kind of leading their data science practice. Fast-forward, they ended up buying three or four other small companies and ended up selling to Palo Alto Networks a few months ago for over 700 million. So I think that was the first signal that, hey, this AI security thing is not just an idea, it's going to be real.

Jackie McGuire

>> And what was the problem that they were solving that you saw a lot of value in?

Ed Sim

>> So the initial problem they were solving was trying to understand that anyone that was using AI models, people wanted to make sure how do you secure them? For example, what if I was downloading a model from Hugging Face and I wanted to make sure that there are no vulnerabilities in there? And then how do I make sure that the people in my organization can only use the models that have policies that have been scanned, et cetera, that fit my bill so we had a separate registry? Fast-forward, we added other products like automated red teaming, some agent security and other things like that because we found that the largest companies wanted one vendor to help them solve more problems than just one.

Jackie McGuire

>> Yeah, I got into machine learning. I kind of transitioned from finance into machine learning right around 2017, 2018 as well, and I was one of those people writing those really insecure models, because they don't necessarily teach data scientists about things like hard coded variables and maybe don't put your API keys into a Python script.

Ed Sim

>> By the way, you bring up such a great point, because there are so many problems right now with these keys being sitting there in code. And transitioning to the AI Security Summit, I'm also the first investor in Snyk on their board. And today, they announced Eva, which is their agent product. And our view is that, yes, we do a lot of the things that you want to do to secure models, but if you don't look at the code first, the code tells you what's happening. It tells you that there are keys in here. It tells you that there's MCP servers and endpoints. There's hundreds of them now in organizations people don't even know about. So you need to scan the code first in order to then start protecting it.

Jackie McGuire

>> Yeah, I heard an anecdote one time that if you're an entrepreneur, you know it because your ambition always outpaces your ability. And so I think we're kind of there with AI right now in that our ambition to use these technologies is kind of outpacing our ability to secure them, because there's a lot of things that haven't determined. What's the pricing model going to be for authenticating agents? How do you deal with agents that are ephemeral and may only exist for milliseconds? And so I think it's one of those things with hard-coding variables and keys and tokens. I think because we're trying to race to achieve ROI so quickly, a lot of the vendors really haven't had a chance to build the technology. So Eva was great to read about this morning, because it offers a lot of the kind of key abilities that we're missing, this kind of orchestration, the ability to tie things together in different ways.

Ed Sim

>> Well, you should be working for Snyk, and you just nailed it, because I think part of it also is the complexity. So think about Eva as a swarm of agents doing work and each agent has a specific role. And you just need to make it easy for a security engineer or even for the developer, so they don't even have to worry about it. I think that's the point that you made. But I think another good interesting aspect is that, look, I've been in this business investing in first rounds in these enterprise software startups for over 30 years. And I've never seen a technology take off with AI-assisted coding like it has. About a year and a half ago, no one was like, "Ah, we're trying out GitHub Copilot." All of a sudden, it's wall to wall every large Fortune 500 right now. And because of that, people are then saying, "Holy crap, I think there's a problem here. I have more core than I've ever seen before. How do I secure this stuff? And then, holy crap, I've got agents like cognition now doing work now. How do I make sure those agents are secure?" So I think it is your whole point. Let's try this first and then, "Oh my God, it works, let me try to secure it." And I think it's always that constant chasing for security.

Jackie McGuire

>> Yeah. I think one of the things that we've seen in the studies that we've done is that as vibe coding, AI coding increases, you see a lot more commits. You see the size of the commits going up like five to six times literally, but very few pull requests. And so we're dealing with this huge issue where we're now writing thousands of lines of code in a couple seconds, but some human still needs to right now be able to read through all of those things, figure out what's right and what's wrong. And so I think one of the things that we're a victim of right now is there's always the joke that if you ask five engineers how to solve a problem, you'll get seven answers. And so I think that's part of the problem is that for a lot of these coding issues, there isn't one way to accomplish them.

Ed Sim

>> Yeah. And that's the interesting thing about code is that some folks code as poetic and it's very short. Others take a much longer route to actually writing the same code. So I think there's a few things that you're going to have to look at when it comes to coding. I think first of all, that the idea of just coding and having more code is probably a bad idea, because I think it causes problems. And the maintenance nightmare is that people will come and eventually see are going to be absolutely astronomical. I'm a big fan of something called spec-driven development, which perhaps the guys at Kiro are talking about as well. And in my mind, spec-driven development provides guardrails. It provides the best of AI-assisted coding, meaning that it actually can generate things, but then also it can actually be deterministic. So you have this spec, and the spec is this is how my application should be built, these are the rules around it, and it's a living, breathing thing that can turn into code. That's a difference now versus what it was 10 years ago. And I think there, you can have some determinism and some stochastic kind of nature around it. And so there's so much big changing right now, and there's 18 different ways to do that as well. And so no one really knows.

Jackie McGuire

>> I was really happy when agentic AI, even though it feels like kind of a hype and people would argue it's a bubble, one of the things I thought was great about it is that initially because LLMs were so compelling and easy to interact with, we were trying to throw LLMs at everything. And developing code with an LLM is arguably the least efficient, most error-prone way to do it.

Ed Sim

>> Absolutely.

Jackie McGuire

>> So with agents, you really get the ability to have these deterministic things, probabilistic things, and then a layer between them to kind of melt them together. And you're not trying to do everything with a very large language model. So with that kind of foundation, we were talking earlier when you're investing, one of the things that we experienced this week or last week is that we've all become very reliant on cloud computing. We've all become very reliant on APIs. So when we had an outage that caused widespread disruption-

Ed Sim

>> Was there an outage?

Jackie McGuire

>> I saw someone call it an adult snow day, which made me really happy.

Ed Sim

>> Why is it always AWS East by the way?

Jackie McGuire

>> Because that's what it defaults to when you're coding in AWS, right? And so that's one of the things that sometimes we don't even intentionally create these focal points of risk. It's just by default that, that happens. I built a music festival in Minecraft on AWS, and that was one of the first things we did is like, "Okay, we need different instances. In every region, we need failovers." But if you're vibe coding and you don't have a background in cloud architecture security, and you don't understand our region's work, maybe that's not even something you think about.

Ed Sim

>> Not everyone is skipping that problem. They're just like, "Oh, I'll just code. I'll just throw more code at the wall, which means more security risks."

Jackie McGuire

>> So one of the theories I have is, and I think a lot of people share this, is that hybrid computing is going to be become a new focus because, one, maybe we don't need to be paying ingest and egress costs for petabytes of information. So how do you think about that as an investor in terms of, one, the risk associated with cloud services, and then two, kind of how that's going to push us back to the edge?

Ed Sim

>> Sure. So a couple things. First of all, let me just hit the point on agents earlier. I think that agents are going to take over the world. We're going to have 100 or 1,000 agents working for us all the time. And I think the biggest problem that agents have, the beauty is that they automate things at scale. The problem is that they can F up things at scale as well.

Jackie McGuire

>> At scale. And velocity.

Ed Sim

>> Yeah. And so I think that you're going to have to rebuild the whole fundamental identity architecture. Agents aren't humans. They act differently. They're project based, they're ephemeral. There's going to be thousands and thousands of them, you need scale. And they also have these workflows that span across applications. It's not just like a human and you don't want to give it forever access. For example, if I was buying a pair of jeans, do I want it have my credit card forever if it found the jeans buttoned? Hell no. Hell no. Do I want it having Snowflake access or S3 access forever? Hell no. So that's why I'm excited about Keycard. I think you're interviewing them later today. It's an amazing team as Snyk and Auth0 that are going to build the primitives around that. That's one. Two is when you talk about cloud risk, I think about less as cloud risk, but I think about what are the enterprises doing in buying right now? I was mentioning earlier in a panel that everyone thought like, "Hey, the cloud, no financial services institution is ever going to deploy the cloud." But then FINRA said, "Hey, it's okay." And then everyone deployed in the cloud and now all of a sudden everyone's going there. But all the banks I'm talking to now are like, "Holy shit." Anthropic and OpenAI, these bills are absolutely massive right now, because what happens is that as a cost per token comes down, when you have agents doing reasoning, the amount of work that they do increases exponentially. And so now, they're saying, "You know what? There's these really great open source models even from China, like Qwen." It used to be DeepSeek, but it's Qwen, and they're deploying that on-prem, and they're also fine-tuning their own models. And they're thinking to themselves like, "How do I actually deploy this in my own hybrid cloud? How do I deploy this on-prem? And how do I orchestrate all this?" And believe it or not, Kubernetes is the orchestration platform in my opinion, that won't go away.

Jackie McGuire

>> Yeah. I mean, it's open source. One, I think, I don't know, I've listened to a handful of podcasts in the last few weeks about, are we at the end of not open source? Are we at the end of walled gardens? And I think realistically, to your point, if we really think we're going to have agents running the world, we can't have walled gardens. The reason the internet works is because of standards and open sourcing. And I also think this is probably going to fundamentally change the business model for cloud compute. Because the hyperscalers, they've had no real vested reason to help you control the rate at which you create data. So it's in their best interest that compute is cheap. It's easy to spin up, you can store endlessly because they get paid that way, but that's also created problems when it comes to training data. And that the average enterprise has five and a half copies of their data. Do they need five and a half copies of their data? No, but why would Amazon or Google or Microsoft help them reduce that? That may shift. I mean, one, I think as these hyperscalers hit their end of ability to deliver capacity while they're trying to build data centers, they may have more incentive to optimize the data they're currently handling to free up capacity for new customers, right?

Ed Sim

>> Yeah, I think you're hitting upon a big point. When you go back to large enterprises again, the idea of who owns the data and what do they do with the data is still a big deal, which is another reason why I think you're going to move towards more of a hybrid computing environment, not just the cost, but also the data sovereignty, the data control issues around it. And the third part is accuracy. I think people are now going back and saying, "Maybe I should fine tune this a little bit more. Maybe I should train some of my own models to get really, really very specific answers." So I think that's kind of what's happening. If you tie that together with what agents do in a workflow, let's say it's 10-step process, you could have three steps that sit in the cloud, one step that sits on another SaaS vendor, another step for human in the loop, and another step that has some on-prem applications embedded there. How do you kind sit there and organize that and understand that and secure that? So there's a lot of new stuff that were at the tipping point of right now that people don't even understand yet. And those are the things that I get excited about is as an investor, because I need to be there a couple years before it becomes real.

Jackie McGuire

>> Yeah. So you had mentioned financial services before. One of the things that's occurred to me is I spent 15 years in investments, and the regulations there are really strict. And so for example, customer communication has to be retained for seven years in medicine. If you're dealing with anything that has to do with Medicare, it's 10 years from the date of service. So as these agents are interacting with customers and patients and things like that, everything they do has to be retained and it has to be retained long-term.

Ed Sim

>> And you know what's amazing by the way? Is that Keycard, since we give agents their own identity, that means they also have their own audit log. And that is a very important piece of it as well, because it can't access data through you or me. We're not going to know kind of who did it. Was it you, or was it the agent, right?

Jackie McGuire

>> Yeah. And so as these data volumes grow, though, I almost feel like we're about to have another golden age for hardware. One, we're bringing hybrid compute back, so we're going to need ... We have this whole wave of endpoints that are going to be refreshed in the next couple years and tape may come back. Because if I have to store 10 years of every single thing an agent does, if every time I train a model, that full-trained model has to be stored somewhere even if I retrain it. I don't think we've really addressed that yet. Are you seeing companies who've addressed the-

Ed Sim

>> It's very funny. Well, I feel like a cripple-esque aspect here in terms of, because I know that's where you came from. And you and I were just talking about one of my other companies called Grepr.ai. What they're doing is they're sitting in between Datadog or Splunk or any other system. They're using ML and AI to reduce the cost of the storage amounts from these folks by 90% or more.

Jackie McGuire

>> And so I don't spend 12 hours grepping through logs, is that the point?

Ed Sim

>> Exactly, yeah. And they're putting it in S3 buckets. I think S3 buckets are pretty cheap as much as you get. So I don't think that is the big problem, but I do think there's a problem around dominance by NVIDIA, as you can see from people betting on AMD from OpenAI doing deals with Broadcom. We as a world cannot be single-threaded on one kind of institution. So I think there's going to be a tremendous new wave of hardware startups, system startups, GPU memory startups, and we have a company that's in stealth right now that is literally solving the biggest problem that exists in AI today. That is the memory-GPU bottleneck. Because right now, they're tightly coupled together, and we are going to have a process that puts 15 terabytes of memory on a silicon onto a PCIe card, and we're going to use optical networking to actually speed up the interconnect between the two systems. And we're doing math on it using lasers.

Jackie McGuire

>> Interesting.

Ed Sim

>> Think about that. We're pre-processing data using lasers, going back and forth, and now we've expanded the memory pool, orders of magnitude, which then increases the performance, 5x to 10x. Those are the kinds of breakthroughs that people are developing right now.

Jackie McGuire

>> Yeah. And I think we're all just kind of hoping that those breakthroughs happen before we run out of concrete for data centers, right?

Ed Sim

>> I think we're already running out of that. And I don't know the energy too, but all of a sudden, I don't know what's the model.

Jackie McGuire

>> Water?

Ed Sim

>> Yeah.

Jackie McGuire

>> Water, yeah. I have a husband who is a chiller expert and chillers too, we can't get chillers, we can't get generators.

Ed Sim

>> Well, the only people with jobs would be the people building data centers. Every one of us are going to be wiped out at some point in time.

Jackie McGuire

>> Yeah. Well, it's a good time to bring in universal basic income, right?

Ed Sim

>> I agree with that, by the way. That's another topic I want to get at.

Jackie McGuire

>> No, absolutely. So then let's talk about physical infrastructure. We've spent the last 20 years trying desperately to get fiber in the ground in communities that really need it. So do you see the race to build this physical infrastructure? Do you think that's going to benefit us, or do you think the fiber now is going to be insufficient and we'll have to go to something like optical? Because I've actually heard about companies putting tapes in the back of a van and driving them from one office to the other because it takes so long to move petabytes of data.

Ed Sim

>> We actually had that problem. So the other area we didn't touch on is robotics. And you saw that Amazon, right? They want to actually ... So I have a company called Generalist AI, and they're building out their own transformer model for robotic hand dexterity. And we have these 3D printed devices that go in people's hands that watch what people do with sensors, and we provide the data and put it in the model. There's a ton of data. And in the early days, we had to build up a whole system of tape drives just to kind of store it all and then ship it because we couldn't get that much data back and forth, to your point. So I think it actually is an interesting concept that you bring up.

Jackie McGuire

>> Yeah. Well, and how we're going to ... Do we have to now go into all the fiber lines and replace those with high-speed optic?

Ed Sim

>> Well, the only good news is that fiber ... Back in the day, remember when the world blew up with a telecom boom and people laid down dark fiber?

Jackie McGuire

>> Yeah.

Ed Sim

>> They're still using that 25 years later. The problem with GPUs is that you overspend on GPUs, those things depreciate within 18 to 24 months. So there's another big question happening there. People are financing these things. It's not dark fiber. What happens if the enterprise revenue doesn't come? We could be in a world hurt.

Jackie McGuire

>> Yeah, absolutely. Well, and what are we going to do with all this scrap metal? I think Dell, I work a lot with Dell, and they're one of the companies leading the way in terms of electronics recycling, but that's going to be become a problem at some point too. All those GPUs are going to have to go somewhere, and it's hopefully not in the landfill.

Ed Sim

>> And Dell is an amazing company. I mean, look at that. I mean, they're benefiting from this whole wave.

Jackie McGuire

>> Yeah. I think Michael really early on saw the kind of hybrid shift, and they're going to be one of the beneficiaries of the hardware focus that I think we're going to have over the next couple of years.

Ed Sim

>> Oh, yeah.

Jackie McGuire

>> Well, Ed, I wish we had more time. I want to bring you back because I think getting an investor's point of view in this murky economy and murky kind of AI pond that we're in is always really helpful. Any last thoughts? Anything that you want entrepreneurs to be thinking about for the next hot startup you invest in?

Ed Sim

>> Yeah. No, look, I tell you, I've been doing this for, as I said, 30 years. This is the biggest platform shift that I think we will ever see and experience, frankly, and it's just getting started right now. And I just think that for founders, it's the best time to build. Start AI native from the beginning. Use as many agents as you can, try to be as lean as possible, and I think the opportunities are boundless, frankly.

Jackie McGuire

>> Amazing. Thank you so much for joining us, Ed.

Ed Sim

>> Thank you for having me.

Jackie McGuire

>> For theCUBE, I'm Jackie McGuire. Stay tuned to theCUBE. We are your source for tech news and analysis. We'll be back from San Francisco at the AI Security Summit.