Agentic AI Meets Cybersecurity + Solar Robots in the Desert | E2125

0:00Coming up on Twist today, we're talking to some of the best founders out there in the world. AI, solar, robotic, factories, and also what happens if you take agentic AI and bring it to the world of cybersecurity. We got a lot to talk about, friends. Let's go.

0:33dot com slash twist to join and skip the waitlist. And HubSpot for startups. Smart founders aren't piecing together random tools. HubSpot is the customer platform that thousands of startups use to scale efficiently. Get up to 75 % off plus three months of perplexity AI for free. Go to HubSpot dot com slash startups. Hey, everybody, welcome back to twist. This is Alex. I have two amazing interviews for you today. The first one is with a company called Zio Set. And it touches on two things that I care a lot about. One, cybersecurity, and two, agentic AI, a term that I'm sure you've heard by now.

1:08Zeosec wants to apply the latter to help with the former. Yes, agent-driven cybersecurity. I love the idea. I love talking to them. You're going to love it too. Then, after that, Terabase, a company from the Twist 500 that wants to use robots on-site to build simply enormous solar panel fields. If you've heard us talk about AI and the enormous power demands that come with it, well, this interview touches on a lot of those themes. And it's good fun. Let's get started. I'll see you on the flip. On the show, you have heard us bang on about AI week in, week out, day in, day out. I'm sure you're very sick of it.

1:42But one thing that I'm also very excited about is AI in a cybersecurity context, or perhaps I should say cybersecurity in an AI context. The world now has digested AI models and is getting used to AI agents, deploying them across the enterprise and the consumer landscape. But what that does is open up new security issues. And there's one startup, Zeosec, that is working on agentic AI security. So I wanted to bring him on, talk to him and learn more. So please join me in welcoming Aaron Walls and Andreas Ushetkas. Hi, guys. How's it going? Hey, thanks for having us. All right. So let's start absolutely up at the top, guys.

2:17We're talking about agentic AI security. Everyone's heard the phrase AI agents or agentic AI. Perhaps we should just start with some ground level here. How does your company define that concept? AI agents are effectively an LLM that sits within a framework, right? You have a web application that connects off to different tools, and they do different things. It can be as simple as going to open up an email program and sending an email on your behalf, all the way down to operating something more local and physical within an environment, security systems, locking doors, unlocking doors based on inputs.

2:54they have far, far reaching applications. So LLMs are powerful, but LLM agents are game changing. Okay. And the reason why I think cybersecurity or security in general in an agentic AI context matters is because the thing is doing its own activities. It's executing on its own frameworks per se, but like it's still out there on its own. And that to me, from the lay person's perspective, of, Aaron, really opens up a lot of worrying security issues. Am I driving down the right road here? You're exactly right. And really what's happening is we have guardrails that are being developed by the model producers, by the hyperscalers.

3:35But really the challenge when it comes to adopting these from an enterprise standpoint, from a corporate governance standpoint, is we currently don't have very many good ways of confirming that those guardrails are working properly i'm going to try some big words here is that because ai models are inherently probabilistic instead of deterministic and therefore old tests the old methods of testing to see if something is operating as it should don't exactly apply effectively yeah and i think audress can actually go into a little bit more detail on that. Absolutely. So if you look at the models these days, they're not just models.

4:14They're not just answering your questions anymore. They're integrating with all these tools. So they have all of these new protocols coming up, like model context protocol, like MCP, that allows you basically to call all of these tools in a standard way. Now that gives the model a lot of power, and there is not much validation and security for that power. So basically, they can interact with the databases, they can interact with APIs, and that opens not just new ways of hacking into the systems, it also opens new ways to hacking into the systems using all the vulnerabilities. So you can still, if you have a model, for example, that interacts with the database, you can still use like old school vulnerabilities like SQL injection, and basically tell the model to execute the SQL injection against the actual database and get the data that way out of it.

5:03Now, when would that come up? Because when I think about AI agents, I think about someone instead of a corporation, setting something up, setting some rules for it, and putting it to work. It's not something that I thought you could do from the outside. So when I think about someone who might want to do SQL injection, I'm thinking about an external person who wants to get in there and cause mayhem. But my view of AI agents was always that they're internally sourced. So Andrews, what am I missing in this picture to make this make sense? So not all agents are internally sourced. I mean, there are a lot of offerings out there, obviously, that can call APIs to send an email, for example, like an assistant agent or something like that.

5:40So that's an interaction that happens from the public perspective. Model context protocol allows for interaction with other data and other sources of information, say, and so from them, you could have, ah, okay, I understand. Absolutely. So just MCP is basically what the model calls in order to get the data out of the actual tool. So it's just a way to talk to the actual tool. But you mentioned the private models as well. I mean, obviously, they're not exposed, but there's still an issue there. But it's, you know, local user exploit availability. So basically, somebody could still do like horizontal privilege escalations, vertical privilege escalations, and get data that they're not allowed to access to, even if they're an internal employee.

6:19okay aaron as if i understand this correctly if i'm using an ai agent and i'm using mcp which is the anthropics new framework that everyone seems to really like and people are adopting to bring in information into my ai context is there no built-in security there whatsoever it just brings stuff in and then huzzah you can the agent can use it because that seems uh even from my perspective relatively insecure it's less that there's no built-in security and more that the fact that when you are bringing this information in, your attack surface is now expanded. And there are more things, there are more links in the chain that could have a vulnerability exploited in it.

6:59So really, it's just a matter of, well, we used to have something, you know, a gen AI that was all contained. And all we really had to worry about was, can we make it give us, you know, an answer that is clean or prevented from telling us how to make a bomb. Whereas now we're really expanding out the different areas of infiltration that can happen with the advent of NCP and A2A frameworks. A2A is agent to agent. It's the recent Google standard that allows agents to talk to one another. So NCP connects to applications externally. A2A is essentially an agent to handshake as far as I can tell. Okay, so this actually now makes perfect sense to me.

7:41I was a little confused why you guys were choosing agentic AI as your focus area, but based on the last two minutes of conversation, I now really get it. I guess my silly question, Aaron, is given how much the tech service is being broadened here, why aren't we hearing more people talk about the cybersecurity risks of using AI agents? Because based on what you guys just told me, it seems like a pretty glaring issue for a product that seems to be gaining real enterprise market share and share. Right. And I think you hit something that's endemic to the cybersecurity market in general, is that a lot of the cyber defenses tend to be chasing the problem.

8:21And the development of these secure services first has not really been a thing we've done as an industry. A lot of what's going on now is we have these enterprises, the development teams who are building brilliant AI solutions and transforming the way these companies are operating. And then you have governance committees within the organization. And then the security team saying like, let's pump the brakes here and talk about like, how do you know what you have built is secure? And a lot of what the development team will say is like, well, we are utilizing AWS Bedrock. We have the right guardrails in place, or we've trained the model, or it has the correct permission sets.

9:07It's like, great. That's a great step one. But what about the 67 other potential issues that are very well documented by the company, the likes of OWASP, as well as MITRE? They have gone forward and published an incredible amount of information on where all of these holes could be. So we have this... Aaron, I'm going to just pause you into some acronyms here.

9:35OWASP is the acronym. I missed the second one. I'm not familiar with it. Can you just explain that to the folks? If you're a startup founder, you've got a million things that you're worrying about at this moment. You know what you shouldn't worry about? Your bookkeeping. Your bookkeeping should be perfect. And you should have a partner who makes sure it is so. And that partner is Pilot. It's the industry standard. Pilot is the largest accounting firm out there built for startups. They know how high the stakes are. And that's why companies like OpenAI, ScaleAI, and Airtable trust them with their books and have done so since day one.

10:10When you use Pilot, you're going to get a dedicated team for everything you're doing from booking to taxes. And now, you know, listen, if you need that CFO level guidance, they're going to give you that. So you can stay focused on what matters, building your team, building your product, and delighting your customers. You should not be stressing over spreadsheets with your P &L and all this nonsense. You want accurate financials delivered on time every time, and you want to be compliant with your taxes. You don't want any last minute surprises. And when it's time to raise your next round and you're scaling up, Pilot's CFO services team is going to help you plan and grow with confidence.

10:43Startups that use Pilot tend to raise a bigger Series B and a bigger Series C round than the average startup. Why? Because when you're buttoned up from the start, everything gets easy. Focus on your product. Let Pilot handle your bookkeeping. This week in startups, listeners get$1 ,200 off their first year. Just go to pilot.com slash twist. That's P-I-L-O-T.com slash T-W-I-S-T. Yeah. So Mitre and Mitre Atlas in particular is this new framework for establishing where the specific holes are when it comes to these agent frameworks. So MITRE has a long history of providing intelligence and attack service management to the industry.

11:31And they have this great MITRE attack framework that has been used for a good part of the past decade. They've updated with the MITRE Atlas framework, which is now focused on AI deployments. Okay. Now, Andrews, I think I now understand why we need to secure AI agents more so than we are today. Tell me why you guys have picked pen testing, offensive security, continuous testing. Why is that the right approach to this issue as opposed to a different angle to start? Well, first of all, pen test is my background. I've been doing pen testing now for 25 years. Is that why you're on Linux today and you couldn't use QuickTime like everyone else who comes on the show that is why i'm on linux yes it's much easier to do testing from linux environment than it is you know from a lot of other environments why why offensive because again validation is needed and the best validation is basically acting as an adversary so basically you're you're the hacker in that case you get into that persona and you try to break into these systems llm is advancing so much that you cannot just hire a pentester once a year that's usually you know industry standard.

12:38So if you're running like e-commerce environment or something like that, you will hire and do a manual pen test once a year. But these environments advancing so fast, you need something that can do it on a weekly basis or sometimes even daily basis. That's why continuous pen testing. And again, you know, coming from this background, we're not just aware of these new attack vectors, like, you know, prompt injection and things like that. We're also looking at all the old school stuff that we can do just through the LLM. So we're using LLM not as, you know, all in one, so to say, we're using it as a way to get into the system and see just exactly how far we can go, just like an actual attacker would do.

13:16Okay. And in most cases, when you run a pen test on a company that hasn't really worked on its agentic AI security, what do you find? Are they wide open, like a barn door to exploitation? Or is their security okay, but just needs a little extra help? It depends on the environment, but these days, we're early in this game. So we see a lot of issues like, you know, getting access to things like financial reports and financial information and getting access to the user information that you're not supposed to have and things like that. Which could be a HIPAA violation, a GDPR violation, a breach. A HIPAA violation and so on and so forth.

13:55Yes. All the bad acronyms, essentially, will come and bite you. Okay. So, Aaron, I now understand the problem. Why are you going about it this way? I'm curious about the market itself because, to me, we've been talking about AI agents industry-wide for like 18 months or so. I presume that that's a little bit late. But what I'm not clear on yet is just how far they've been rolled out inside of actual enterprise-scale customers. Basically, how much out there today is hype and how much out there is real use, driving real need for what ZioSec is working on. Well, so I think that is such an incredible question because we are at this point where the interest from the executive team is there.

14:36They're pushing the organizations to deploy. Their investors are asking, how are we adding AI to our workflows? How are we adding AI to our products? So the momentum is going, but the challenge is the security teams don't have the tools to fully evaluate and fully mitigate that risk. So they're pumping the brakes. So you see a lot of talk about this, but if you talk to the organizations and you see how many agents have you onboarded, we're talking two or three, maybe, right? We're not talking a lot. But so just to put that number into context, how many agents do you think the average, I don't know, Fortune 50 company will have in say five years?

15:20Thousands. They're going to have - They have basically zero compared to what they're going to have down the road. You got it. And it's because of these systemic issues that we're facing. The technology is new, kind of like you're saying, probabilistic models, they can get things right most of the time, right? But it's not perfect. So a lot of what we're trying to do with our testing, we're offensively testing, but we can also do things like make sure that the model is giving the same answer, or at least within the same context of an answer. And being able to report out that degree of granularity to the security team, the compliance team, and the governance teams will then empower deployment of AI throughout the enterprise.

16:00Because right now, that's the major holdback is they don't know what they don't know. And they're pulling back because the risks are huge to having an agent you don't know how it works completely have access to all of your critical internal company data. Yeah, that's what interns are for, not agents. I mean, I joke, but it sounds like a real problem. But to me, it feels like everyone who's currently trying to sell AI agents, be it Microsoft or Sierra or whomever, should be piling capital into Zeosec because if you guys can solve this issue, it could unlock an enormous explosion of adoption of agents that right now are sitting somewhere between the C-suite and actual implementation.

16:45Is that fair? Uh, that is exactly how we feel. Yes. All right. Microsoft, come on. You have a VC firm. Cut the check. No, I, my next question is what's the curve going to look like here? Clearly there are still some issues that you guys are helping to sort out, but when do we reach the actual agentic era? If you will, cause I feel like we're not quite there yet, but we've been on the cusp of it for some time here. Yeah. And I think in talking to a lot of the leaders in the industry, um, you know, the leader of Anthropic. I mean, we've got, you know, Mark Betty off at Salesforce, I mean, it's saying there's going to be over a billion agents, you know, deployed within the next, what is it, six months, 12 months, you know, I think there's a lot of expectation about this.

17:30But until we truly have those business cases that work within the enterprise, and ones that are safe to deploy, then it's going to be a matter of like solving these problems in order to get to that point. I'm guessing six months is when we're really going to start seeing the floodgates open. Andrews, do you agree with that timeline? Six months until the floodgates open and the agents come and take over our lives and ruin our ability to have jobs? I think it's optimistic. But again, this industry is moving so fast that I can't really predict at this point. I'm not even going to venture, I guess.

18:06but yeah, six months to two years. Well, that's also Fusion and a whole bunch of other things. Okay, so we mentioned MCP from Anthropic. We've talked about A2A from Google. You guys wrote a post digging into kind of multi-agent systems. And to me, like there's gonna come a time in which the idea of an agent that does multiple things probably gets collapsed into chains of agents. Does that make the security issue simpler or more complicated? Yeah. way more complicated because now these agents basically interact with each other and you know they can talk to each other and extract information from each other so it's not just testing one agent at that point you're basically testing an entire environment so to say entire infrastructure right and so that's where the idea of pen testing in an agentic context to me andrews doesn't seem to be as effective because if you have so many agents constantly talking is penetration testing the the right way to go about it?

19:06Or maybe I'm asking, what's the next step in securing agentic AI past pen testing? Because it doesn't seem to cover every single problem that I could invent in my head that you might want to solve. Yeah, but usually pen testing is not limited to single application. Usually when you hire a pen tester, they will test an entire environment. They will even go to like LinkedIn and get the profiles and see what they can hack into, who they can send spam emails to or phishing emails and get information that way. it's the entire like holistic testing and that's how the agent testing is going to become as well so eventually depend testing is not going to be this simple single agentic ai it's going to be an entire environment that you're going to be testing at the same time okay founders it's all about efficiency and performance in your startup but what about you what about optimizing yourself when's the last time you got in there and said how could i be better stronger faster sleep better diet, exercise, mentally, all of that stuff gets skipped until now because there is a new product that I'm using called SuperPower.

20:11It's the ultimate health membership, specifically designed for founders, but anybody can use it. It's built to keep high performers at peak performance. You get full body testing across 100 plus biomarkers. This is your organ health, your hormones, inflammation, and it gives you actionable data and clear insights. You track all of your KPIs in your startup, why not track these for yourself with Superpower? Visit superpower.com slash twist to claim your spot and unlock peak health today. I can't say enough about self-directed healthcare, which I believe this is the foundation of. I've used a bunch of other services, and this is the best one I've ever used.

20:50Better health equals better founder. It's that simple, which equals better business. Again, superpower.com slash twist. How long until we have AI agents doing pen testing for other AI agents? And I'm not being facetious, I'm actually curious. Right now, so our solution is based on AI testing AI. It's not just one-off prompts, basically. It can interrogate the other AI. So we can basically tell our system, you have 10 turns and you have to extract this information from the other AI. And it's our model doing that extraction. So you guys are literally like the AI police? yeah i guess it's a really interesting paradigm that we're moving into because you know police are for bad actors and ai itself can be trained to be bad actor or just the negligence of an ai can have bad actor tendencies so we need to do the discovery of these tendencies and provide visibility into the org where it's needed the most and give them the ability to say like this you know exposure here like yeah maybe they didn't get the answer right all the time or whatnot but that's okay because it's customer service and we're only going to be getting 70 of those right anyways even with humans so right so it totally depended on the application whereas let's say you're now dealing with loan processing and applications, totally different set of requirements.

22:22So I think, you know, saying like, we're policing this is a really interesting way. But these agents are, I think the right way to think about agents is they should be treated much like humans in the organization. They are unpredictable, much the same way humans can be unpredictable. But they are given jobs, and they need to be evaluated on the work that they're doing. And I think that's where we can play a critical role. Talk to me about the future, though. I think I have a reasonable grip on pen testing in this context, but what's next for the company? Well, I think the most interesting piece of this is the company where we're going now, the people who understand this problem are on the security side of the equation.

23:03That's their job. That's what they've always been there for. That's their function in the company is to secure the assets of the organization. Where I view us going is if we can expand that purview into the engineering side of the world too, and allow the engineers to understand that code is not innately secure, it's your job to secure this code. And being able to lever a lot of these vibe coding assistants, you know, which are phenomenally powerful tools, the code that they're producing, not necessarily the most secure code. And that's not any fault of the vibe coder or of the engineer. It's just, you're taking again, probabilistic, you took the average of 10 years of stack overflow, and that's what we're getting.

23:56So now, where we have this code that's being generated and we're not able to confirm out of the gate is this stuff secure so i i really see us moving in the direction of fitting within that model of code being produced by ai and then checked by ai as well but from an offensive perspective okay well that sounds very exciting sounds like a future that's both very fast, very flexible, and also secure. And that would be lovely because currently I feel like development isn't any of those things. So that sounds like a much better future. All right. Last question before we go is just how are you guys doing on landing first customers going out to the market?

24:40How are those conversations going when the rubber meets the actual enterprise road? It's going great, right? So we have, we've had a lot of really great conversations with design partners, as well as companies that are really just starting to get a sense of the new landscape. And we've been able to identify certain companies that style, right? Who is our ICP? And they tend to skew larger, they tend to skew towards finance, They tend to skew towards those critical regulatory environments because that's the kind of company that truly understands that the problem is there and they have to deal with it because they've been working in these frameworks forever.

25:28And we're also partnering with a couple of hyperscalers, which is incredibly exciting. And being able to offer innately our services if you are, you know, signed up with this particular hyperscaler. Well, there's only so many. There are only so many, but that's... Blink twice if it's Azure, blink three times if it's Google, or four times if it's AWS. No, no blinks at all. All right, fine. Well, actually, I was just thinking that it'd be a really smooth way to sell this if you had a partner like Sierra, which builds AI agents and just kind of work with them as like a preferred, essentially, software vendor.

26:07And then you could probably just grow alongside some of these other companies that are doing well. But I think going the hyperscaler routes even better than that. All right. The website is ziosec.com. Just before we go, what is a role you guys are hiring for where you desperately want that excellent candidate? Andres, you want to take it? Yeah. So at this point, we're looking for security analysts, especially, you know, ones that have been doing band testing in this space. There are not many of them. So those kinds of people that we need. We're also looking for a couple of full stack engineers, particularly with the expertise in Rust and TypeScript on the front end.

26:43All right. Well, guys, thank you very much. When you have those first couple of enterprise customers, please come back and tell me all about them and we can wrap on about business models. But in the meantime, good luck. And here's to our agentic future. Thank you. Thank you. I hope you're ready for us to stop saying AI every third word for the back half of this show. But if you're curious about where all those GPUs are going to get their juice, well, look up to the sky and then say the word harabase out loud let's go if you are a long-time listener of the show you have heard jason and i bang on ad nauseum about ai and the enormous enormous energy inputs that doing all that number crunching is going to take now some people think we need to bring back nuclear cool by me some people think that fusion is closer to a commercial viability than other people do cool but one thing we do have today is a whole lot of sunlight and we have amazing technology, we as humans, to capture that energy and turn it into juice.

27:36Juice that can power data centers, homes, whatever you want to call it. Now, we all know this exists. Why isn't it everywhere? Well, there are some labor issues, some supply issues, but there's one company called Terabase that is doing excellent work to try to bring mass-scale solar to the world using automation out there in the field. I wanted to learn more. I added them to the Twist 500, so please join me in welcoming Matt Campbell from Terabase. Matt, hey, how you doing? great how you doing i'm i'm doing pretty good because when i was prepping for our chat today i was surprised at just the scale of progress in installing solar power in the united states the the narrative out there is that it's all happening in china and that we are infinitely far behind but according to the solar energy industry association 2024 report we installed 50 gigawatts here in the united states of solar power last year which is more than i expected and And I thought rather encouraging.

28:28So from a high level perspective, are we doing okay at overall solar install here in the States? Fantastic. Yeah. I mean, I think like you said, 2024 was a banner year. I forget the exact stat, but it's like more than 80 % of the new generation in the US was solar. 66%. Oh, 66. Okay. But still, I mean, it's a huge percentage of the new energy coming online. And I think it bodes well for what's ahead. so terabase is a company that melds uh software both for planning and operations and then in the middle there is an element called terra fab that helps you guys as far as i can tell construct solar panels out in the field for folks who are not familiar with terabase can you just walk me through the product mix really quick sure yeah so terabase is focused on and let me just start explaining the utility scale solar market so these are good point yes giant solar power plants out in remote areas that span thousands or tens of thousands of acres um and so 100 megawatt to multi-gigawatt type sites and so what we do at terabase is we build um digital and automation tools to design build and operate these giant projects more efficiently so there's three parts to this there's the engineering work that goes in before boots are on the ground then there's the construction element of this which is terra fab which i want to start with and then at the there's stuff to help run the solar farms.