00:06:00 · Hanah-Marie Darley
[00:06:00] S0: It was such a step change from requirements and documentation and process to a meritocracy where I think Dark Trace, especially when I joined, still very much had a startup mindset in that if you thought you could solve a problem and you had a good idea, you could probably go run with it, which was massively appealing to me because I was very used to government processes. I was very used to you have to formulate a whole process of requirements and you have to check those requirements of stakeholders and so on and so forth. And so the idea that I can see the problem, I have a solution, I can go implement that solution was like groundbreaking to me. So I was really excited by that freedom. I think the other thing is the intersection between security and AI was probably brought to the fore the most for me there at Dark Trace, more than it was in government. Um, those were always very bifurcated topics in government. You know, AI was kind of over here on the research side, and cybersecurity was more on kind of the offensive or defensive side. And it never the twain shall meet, really, at the time that I was working in government. Whereas with Dark Trace, it was very much an intentional injection of both cybersecurity using AI and AI for cybersecurity and really thinking about those two paradigms in one and how do they coexist and how do they make each other stronger. And so that was really impactful for how I also think about AI, how I think about the different applications that you can use as like a toolkit. Um, all of that's really informed by the building blocks from government, but it was definitely influenced by some of the different tactics and techniques that I saw, not just at Dark Trace, but in industry that worked really well for customers and the ones that didn't. But I think the the kind of again, that red thread through it was how do I think a bigger picture about systems and technology and people when it comes to solving a really complex challenge that isn't just a binary yesno or isn't just uh malicious, not malicious. And I think that sometimes, especially in security contexts, we go really, really down to lease principles, which is like bad, good, um, which is is really helpful from a sock decision perspective, but it's probably unhelpful when it comes to planning our systems.
00:07:57 · James Wise
[00:07:57] S1: Darktrace was founded on the idea of sort of machine learning was going to unlock new types of cyber risk, but also solutions, defensive solutions to those risks, uh, and obviously there for a few years. And then a new form of um artificial intelligence became available through through through Gen of AI, um which added to the plethora of opportunities and threats. Was it the beginning of that technology that inspired you to think actually there needs to be a new product set? Or was it more that you know you met brilliant people at Darktrace who pulls you out and said, look, you know, let's go and build something?
00:08:29 · Hanah-Marie Darley
[00:08:29] S0: No, I think it was definitely looking at where the market was going. I think it it was so obvious to me that generative AI in its 2022, 2023 form, wasn't going to deliver the productivity returns that most businesses wanted from it. And so it necessitated some sort of a vehicle to give the model context and operational power. And that came in the form of agents. And so I think very naturally, as soon as agents started to be kind of discussed in research circles and started to be kind of uh almost a slight ripple in the market. For me, it was an obvious flag of like, this is going to be where things go because it gives the models operational context and capability in a way in which we'll actually deliver transformational change for businesses that was very much promised by large language models themselves and by small language models, but never quite got there because it was so interactive. I always describe LLMs as kind of like, you give me cookie, I give you cookie. Like it's it's all just a back and forth, but it never really moves the needle on operational outcomes or on changing the way that work looks or or on any of the kind of business realities that will bring you true return on investment and productivity. And so again, a lot of that comes from how I thought about in government the future forecasting. I looked at some of the signals and I said, okay, well, if agents are going to be transformational to every aspect of work, if that's true, what are the things that I need to look out for? And some of it is early adoption, some of it is good returns on actual research that that showcase how it can work, especially autonomously rather than more kind of user-directed. And those things started to fall into place. And so my co-founder Henry and I had a funny conversation before we founded the company. He's like, I think businesses are probably going to take 12 to 18 months to get there. And I'm like, Henry, they're going to be there in six. Um, and you know, it was a it was a really hot take. Uh, I'm kind of infamously known for hot takes. They're not always right. But but in this case, you was in this case it was, and I definitely should have bet him money on that. But um, I think it it really, it really shows the pace of change. And I know we talked about this earlier, but the pace of change used to be measured in years.
00:10:30 · Hanah-Marie Darley, James Wise
I mean, if you talked about cloud migrations, some people are still in the midst of a cloud migration. The technological rate of change and the competencies that you had to develop, you had so much more time as a business. And now it's measured in days, weeks. And I think it's true of the threat landscape. We often talk about vulnerabilities and exploitations and threats are now measured in hours, not weeks. Um but but the same is true for the pace of technology, and and so the two are Gemini linked, I think. [00:10:57] S1: Yeah. And when we uh sort of first looked at Geordie, and obviously we we tracked um Henry and yourself and knew Benji from Snyk as well a little bit, uh, you know, you you were talking quite in a way at the time in in early 2025, it was visionary about uh, you know, agent adoption. But when we did references, a lot of security professionals were like, well, we're not using it. [00:11:15] S0: Not there yet.
00:11:15 · Hanah-Marie Darley, James Wise
[00:11:15] S0: Not there yet. [00:11:17] S1: And then there were two different perspectives we also got from experts in the space. One was, this is just another endpoint, we'll deal with it in the way we've dealt with everything else with our traditional uh cyber security stack. And the other one was, well, it's completely different to everything else. We'll just put it in a gateway and we'll just own all the data that way, and we'll do some more observability around it, and it it will be a completely different thing. Whereas, you know, you guys from day one came at it with a more broad perspective. Could you share a little bit about the, you know, the founding product principles of Geordie?
00:11:47 · Hanah-Marie Darley
[00:11:47] S0: Yeah, I think one of the most important things is that we always treated the agent as a completely new entity, as a new focus and as a new paradigm. And so I think anything that you're trying to adapt backwards from either endpoints or even from gateways, because the majority of how we tackle user challenges is gateway oriented. We put some sort of a gate in between the thing that we don't want you to have and the thing that we do want to keep protected, and therefore that will be a central access point for data, but also a central enforcement point. But I think the real challenge with agents is that they don't operate like human users. Um it's the same challenge with AI more broadly. We don't really have a category as humans for anything that has human principles but isn't sentient or isn't human. And so we sometimes over-anthropomorphize and we kind of escalate what AI is and what it isn't. And the same is true for agents. Agents operate very, very differently to human users. They have some commonalities, but they have machine speed performance. They have the capability to operate across multiple surfaces simultaneously. They have the capability in many cases in harnesses now to have their own sandboxes and to run some micro experimentation within their own operational cycle. And that would never happen from a human user. There's just no way to kind of parallelize that. And so all of the systems that are designed around how humans behave and how humans operate and controlling essentially human traffic is never going to work for agents because their operational pattern is completely different. So those were the founding principles that we came at it with. And so when we thought about how do you provide oversight, how do you provide certainty for security teams that needed evidence to move forward and to enable the business? How do you operationalize this across the business in a way that's secure, but also in a way that provides value, provides the investment back into that business in terms of either OpEx or CapEx? And really, it was fundamentally work from the agent, work directly at the agent and now the harness layer, and essentially think about the agent's execution loop.