Cyber Investing in the AI Exploit Era

What happens to security investing when vulnerability discovery becomes continuous and exploitation windows shrink from weeks to hours? I sit down with Chenxi Wang of Rain Capital to dig into it.

Chenxi is the Founder and Managing General Partner at Rain Capital, a venture fund focused on early-stage cybersecurity companies. She's been a Carnegie Mellon professor, a Forrester VP, and a strategy leader at Intel Security and Twistlock, and her portfolio includes companies like Claroty, ProjectDiscovery, Ox Security, runZero, and Straiker. She closes out my July run of conversations with security investors.

In this episode:

• The AI Exploit Age and why vulnerability discovery is becoming continuous
• Guardian Agents and the case that it takes an AI to govern an AI
• Separating AI agent identity from traditional machine identity
• The signals that predict enterprise adoption for early-stage security startups
• The barbell funding market and the squeeze on Series B and C
• What security leaders should do differently over the next twelve months

Connect with Chenxi:
 LinkedIn: https://www.linkedin.com/in/chenxiwang88/
Rain Capital: https://raincap.vc/
Rain Capital Insights: https://raincapital.substack.com

Subscribe to Resilient Cyber for more conversations with security practitioners and leaders: https://www.resilientcyber.io

Transcript

SPEAKER_01 0:00

If the AI market crashes, I'm not a alarmist, so I'm not trying to say that. I'm just saying from an economist point of view. No, that's it.

SPEAKER_00 0:09

We are seeing the industrialization of vulnerability discovery with AI. Finding things has kind of been become cheap, but exploitation that hasn't taken off yet.

SPEAKER_01 0:17

I think that economics is what's stopping the bad side today, but it may not stop them tomorrow because we know cheaper models are coming.

SPEAKER_00 0:26

The AI exploit age, where vulnerability discovery moves from scarce to continuous, and the window from discovery to exploitation is compressing, you know, from weeks to hours. You know, what convince you that this shift is kind of structural rather than a hype cycle? Like what makes it real to you?

SPEAKER_01 0:40

I think Mythos has proved it's feasible. It's feasible that it'll discover vulnerabilities even beyond humans' capability and certainly at scale.

SPEAKER_00 0:50

I think we are gonna see kind of widespread autonomous exploitation in the future.

SPEAKER_01 0:54

I know a lot of those CISOs are losing speed.

SPEAKER_00 1:11

Thank you for joining Resilient Cyber Show. My name is Chris Hughes. Today I'm joined by Chanxi Wang. Chanxi, thanks for being here.

SPEAKER_01 1:17

Thank you for having me.

SPEAKER_00 1:18

Yeah, I'm excited to chat. I've had you on in the past, I think it was around this time last year, actually. Um, you know, you had some great materials on, you know, state of venture capital, you know, services as a software, AI, cyber, everything in between. I follow your content for quite a bit now on LinkedIn and so on. But for folks that don't know you, don't know Rang Capital, can you tell us a bit about yourself and the team?

SPEAKER_01 1:37

Sure. Uh we are an early stage venture firm based in Silicon Valley. We invest in cybersecurity and AI. I'm the founder of Rain Capital. Uh, Sidra Almond uh is my partner. She came from uh Munich Ree, Munich Reventers, that is, and we are building a early stage um cyber AI and health tech investment firm uh in Palo Alto.

SPEAKER_00 2:02

Very cool. Yeah, and you know, you have one of the more interesting kind of career trajectories or arcs in the industry that I know of. Um, you know, you went from kind of academia in the community at Carnegie Mellon, where I know you're spending some time at this summer again. You've been on the kind of the analyst side, you know, of things, and you've been a strategic leader for you know large uh incumbent cybersecurity companies, and now of course on the venture capital side, you know, what pulled you from like the operator side, the analyst, the academic side to the venture capital space?

SPEAKER_01 2:31

Yes, I I joke that if you look at my career trajectory, it's clear, some it's clear it is someone who doesn't know what she wants to do when she grows up. I think I'm just like attracted to different aspects of things. Um, so I came from academia, and then after a while, I said, you know, it would be nice to try industry. I went to industry, I did a number of different things. Then I say, hey, you know, it would be nice to try working with a portfolio of companies versus just one company at a time. And so I tried my hands in consulting and venture capital, and now I'm a VC. So it's kind of opportunistic in at some level, but also intentional in the sense that every decision that I made, it is to learn new things, to grow as an individual and a professional.

SPEAKER_00 3:24

Yeah, I think it's uh honestly, it's probably super beneficial at this point because you've seen the industry from many perspectives. You've seen the kind of research, the academic side, you've seen the practitioner, you know, side of things, uh, at large uh industry leaders, right? Like, you know, providing services and capabilities and products uh to the industry. Now you're kind of funding those uh those entities, especially the ones that are trying to disrupt the status quo and be innovative. Um, and I actually follow your content, like I said, quite a bit across different platforms, one of which I caught you on a podcast called Venture with Grace in May. And you've talked about some of the signals that predict uh enterprise adoption for early stage start uh security companies, right? You know, when you look at a seed stage AI security company today, what tells you, you know, what enterprises may buy? And how do you how do you use like your pulse of the industry to kind of gauge that when you're speaking to some of these founders?

SPEAKER_01 4:10

Yeah, and first of all, I um uh I do a lot of work with practitioners, uh, meaning that I I chatted with them all the time and I try to understand where uh their challenges are today and where they're spending time on. In my travels, those things became uh sort of uh problem sets that we intentionally looking for founders to solve. And obviously, we have a list of problems, and there are founders come and pitch on a slightly different uh set of uh uh problems. And so we try to match them, right? And sometimes it's a perfect match, other times we're kind of trying to triangulate with the signals. Um, so one of the things we're trying to decide with founders is how compelling the problem is. You know, is this something that people will buy today, in three months, in six months? Uh that's a very one of the first questions we ask. Uh, another one is how big of the market, right? So is it you may be solving a very compelling problem, but you know, maybe you have a handful of customers and that's it. And and that's not a venture backable business. So those are the things we look at in addition to the T.

SPEAKER_00 5:25

Yeah, I was gonna ask when you said that there, of course, like there's the TAM, the addressable market, you know, the potential for revenue, scale, and size. Um, you also made a comment about three months, six months, nine months, like, you know, in down the road. Um, you know, has that changed? I'm curious, like when you look at the kind of pace of innovation right now with AI and how quickly people can build. And it seems like every every like four to eight weeks, there's something new, the harness, the model, the this, the that, like, you know, MCP's dead. No, MCP's not, you know, like there's all these kind of debates in the industry. Like, how do you kind of orient uh yourself when you're looking at these founders? Because it's so hard to predict what things may look like in six months or 12 months and so on.

SPEAKER_01 6:00

It is very hard to predict, especially with AI, right? The AI infrastructure world. I think it's changing every six weeks. Um, I was just listening to a podcast um on the way to work this morning, and uh they were talking about the memory layer for AI agents, and that has changed so much. And now the buzzword is semantic memory, right? And six months ago, it wasn't that, or maybe nine months ago. Um, so it's uh uh in the security industry, I think it's a little bit less frequent in terms of uh uh uh one wave of technology replace the other. What is interesting is how do you leverage these new tech that comes from uh the AI world, the new innovations, right? So often I get on a call and I'll talk to the uh the founders that are pitching a security solution, and I'll ask them how they build their stack, right? And sometimes you can tell whether uh the founders are very AI native or they are AI oriented, but they're not AI native. So you want to get someone who is really plugged into the bleeding edge of AI because you know, if they build their stack on the older generational technology, then it might be hard for them to pivot and to compete. So I these days I always look for that, right? I always say, who's the AI guy on your team? And where does this person come from? And where has she or he done in the past? And so trying to find a team that moves fast and pivot fast is one of the key ingredients today.

SPEAKER_00 7:40

Yeah, it's a great point. And anyone who's followed the space, you'll see many great companies, you know, that are now incredibly successful and large. Like they didn't necessarily start with the specific problem or thing that they do currently. They've pivoted along the way, they've changed, they've evolved, you know, and that's like a critical skill set they need to have. And, you know, speaking of kind of trends and AI stacks, like I know you have a newsletter as well. And in the June newsletter, uh, you were making the case, you know, for the concept of guardian agents, which is kind of like, you know, who's watching the watchers? Who's gonna, there's so many agents in the enterprise, you know, across every category. We're looking to use them now, whether it's SecOps or GRC or AppSec and everything there in in between, these kind of specialized AI systems that supervise and enforce boundaries on other AI systems. You know, what does that supervision layer look like uh from inside an enterprise? And then I'm also curious how you think about this because I know you come from the academic side and keep in touch with practitioners and researchers. It seems like there's like a circular logic where all the problems of AI are AI. Like we need more AI to watch the agents. We like the AI creates a problem. Now we need to use AI to solve the problem. And that kind of uh it makes sense, right, for scale and we don't have enough humans and eyeballs and all these kind of things and workforce constraints, but also like it kind of there's hallucinations and like, you know, all these kind of kind of uh problems like of LLM as a judge and so on, that when you use AI to secure AI, it's kind of like a circular logic uh involved there. How do you think about that?

SPEAKER_01 8:59

It's interesting that uh, you know, I'm sitting in Carnegie Mellon uh university this week and yesterday actually gave a seminar and we talked about uh verification of agents. And indeed, I talked about LM as a judge. It's funny that you mentioned it. And they did not like that at all. The audience was very skeptical of this approach now. But I think if we take a step back, when we first had cloud, we need cloud security, right? So cloud was a new infrastructure, you created a new uh attack surface, you need to secure it. And and guess what? The cloud security solution ran out of cloud too, right? So we used a cloud to secure cloud. So it's not a new thing. Anytime we create a new layer of infrastructure, uh, new set of security problems arise just because it's new and we haven't really thought about it or really had a deep intuition about how to secure it. Now, how do you add on the security, safety, trust layer? That's a design decision. And let's talk about do you use AI to secure AI? Do you use deterministic policies to secure AI, which is a um a discussion point I hear a lot in the security community, right? So one of the fundamental things we should get out of the way is when deterministic policy works, we should absolutely use deterministic policies. But we know in many scenarios it does not work, does not work well, uh, especially in edge cases or in things that doesn't have a yes or no answer, right? So in those scenarios, what do you do is a big question. Okay. So the um uh one of the biggest um incidents that that we saw recently was the open AI and Hugging Face, right? And so um, if you went on that call with Hugging Face uh security team, they said that the agents were just persistently trying to do something. You tell it no, it'll come back up and it'll try something else. And what happened was the agent was given a goal, but was not given any kind of feedback on how it should accomplish the goal, right? And so it is actually running into they were using deterministic guardrails and other things and sandboxing trying to stop it. But what happened was when an agent is given a no answer, it'll say, okay, this doesn't work, I'm gonna try something else, right? And but still uh with the goal in mind. So, you know, if um to to give an example, if you tell a a kid that is not trained to go to a restaurant to pick up a um takeout, but don't doesn't tell him or her that you know, if there's a line, you have to stand in the back of the line, you can't jump the line. And also don't elbow away that old lady and there's a puddle on the floor, you should not be uh running across it because you may fall. All these things you can't possibly specify in a prompt or in uh uh in a policy, right? And but AI agent is essentially an untrained uh teenager or untrained kid. It'll try all kinds of different things. So when you have a deterministic policy, what the AI agent receives is a yes or no, has no feedback. When you have an L Amazon judge, what could happen is you actually given feedback to the agent about the thing you you're about to do is not correct or not accurate because these are the reasons. How do you try different things or maybe try a different method, but with with these considerations, things will be better. I know of a project with a fairly large AI company, they're experimenting with LIM judge internally. Uh so what they do is they ask the agent whenever they are, first they use a deterministic policy to look at what the agent's trying to do. When the agent's trying to invoke or do something uh or perform an action that's not a clear yes or no, they'll send to LIM as judge, but then ask the agent to issue a justification. Right. And uh, you know, a free form justification is not a deterministic policy can handle, but LM's judge can handle that really well, right? And the LIMS judge can like iterate with it with the agent until the agent sort of submits something that considers safe and has the right justification. And that can done at scale. Uh now, the cost-wise, I'm not so sure. Uh, but that's a very good example of why LM judge uh is an interesting alternative to consider.

SPEAKER_00 13:59

Yeah, no, I think you're spot on. It's not necessarily like uh either or it's gonna depend on the use case. And like, you know, we know in some cases we need deterministic, you know, kind of hard boundaries as as many have started to frame them, deterministic controls and the environment that, you know, this can't happen underneath any circumstances and it's outside the reasoning loop and so on. And then in other cases, like, you know, maybe you need some uh uh flexibility, you need non-rigid controls to kind of examine intent and the agent's behavior and things like that uh to kind of make decisions. So it's not either, or I think it's gonna be a combination of both as the you know, as the space starts to continue continue to mature. Um, another thing I've seen you write about in uh your newsletter in May, I think it was, is uh, you know, this this kind of convergence of uh AI's industrialization of vulnerability discovery and soon like you know maybe autonomous exploitation. Uh you talked about writing uh entering the uh AI exploit age, where vulnerability discovery moves from scarce to continuous, and the window from discovery to exploitation is compressing, you know, from weeks to hours. You know, what convinced you that this shift is kind of structural rather than uh a hype cycle? Like what makes it real to you?

SPEAKER_01 15:05

That's a good question. Um, is it a hype cycle thing? Potentially. I haven't thought along those lines very deeply, to be honest, but I think Mythos has proved it's feasible, right? It's feasible that it'll discover vulnerabilities even beyond humans' capability and certainly at scale. What hasn't become the reality is to do it cost effectively. And so an attacker today using Mythos is going to be very expensive, might be more expensive than what they can get from the attack. So I think the economics is what's stopping uh the bad side today, but it may not stop them tomorrow because we know cheaper models are coming, right? Um, so because of that, the defenders I think have a limited window to really beef up their defense posture. And how do we do that is a big, big question, right? I know a lot of CISOs are losing sleep as we speak about this. And I'm sure we'll hear a lot in um at Black Hat on this very subject.

SPEAKER_00 16:12

Yeah, I think you're spot on. Like going into Black Hat, you know, especially with like this recent incident with Hugging Face and Open AI, everyone's gonna be talking about it. And there is it's one of those things where we never discuss things with nuance. Like there is nuance to it. Like on one hand, we are seeing the industrialization of vulnerability discovery with AI. Uh finding things has kind of been become cheap, but exploitation that hasn't taken off yet. Like if you look at uh uh first, you know, the organization first that does the CVEs, like their mid-year update, we're gonna hit 70,000 plus CVEs, but you know, exploitation is still relatively flat. And then there's another uh uh organization called Volm Check, I follow for like, you know, kind of vulnerability intelligence updates and such. And they showed the same. Like the exploitation rate is still about one to two percent, which is the historical norm. But that doesn't mean it's not coming, to your point. As uh these models get better, cheaper, more efficient, become kind of widespread and available to everybody, uh, that could change. I think we are gonna see kind of widespread autonomous exploitation in the future. Uh, if anything, we should be kind of thankful that it's not here quite yet. But gives us like your point, time. We have a little bit of time, you know, to address all this technical debt.

SPEAKER_01 17:16

So I'm uh part of my work at CMU is I have a venture partner here who's a professor at CMU, and his name's David Brumley. He published uh a piece of work called Exploit Bench. Uh, and this is uh um a really interesting study that compares the different frontier models in their ability to um carry on the journey of exploitation from vulnerability discovery to uh leveraging the vulnerability to crafting the exploit, to execute the exploit as a whole journey, as opposed to methos only discovers vulnerabilities, right? And so his conclusion, uh at least uh in um in his uh orchestrating of the models, his conclusion is that Mythos is still the most powerful in terms of being able to actually carry all the way to exploitation. I think that is a a piece of academic study which will become real. Uh I I don't want to put the number month on it, but I think it's already been uh demonstrated feasible. Um, yeah, so that that's that's kind of scary, Chris.

SPEAKER_00 18:30

Yeah, it definitely is scary. And I I saw that too, and I loved it because it again, like to my comment about nuance, like it went beyond the initial hype and headlines and dug into like, you know, what can they do beyond just finding something, like true exploitation, lateral movement, persistence, like a broader uh kill chain of activities and so on, a tech chain. Um, so that was an amazing piece of research, and I shared it on LinkedIn and people loved it. Um, you know, I was gonna ask you about another topic is like uh in the age of AI and agents, another topic that's been getting a lot of attention is uh, of course, identity and access management. Like we've struggled with it for years and decades, uh, concepts like least permissive access control, not least autonomy. Uh we had our kind of wave of NHIs or non-human identities, of course. Um, you know, how do you think of identities for agents? Are they NHIs? Are they something different? Like, how do you think of uh you know AI and agents uh impact on IAM?

SPEAKER_01 19:19

I think uh agent identity has um similarities to human identities and also have similarities to uh NHI, right? Um so the the thing with the similarity with humans is is um you know it takes on humans' ability to perform actions, um, and often those actions are performed on behalf of human identities. And also it can be fairly long-lived agents, right? Um but it's also a it can be very ephemeral, just like service identities. And so I don't know. Um this if you were if you asked me to pick one, say, is agent identity definitely belongs in the human bucket or then the service bucket, I'd probably put in the human bucket.

SPEAKER_00 20:12

Yeah, I mean, we're starting to move away from the traditional software. Now it can have autonomy, it can take actions, it can do something in the enterprise use tools. I'd I'd agree with you there. It's definitely like kind of a blur of both. And I think that's why the industry is like, you know, you see so many startups in this space, you see like uh discussions around new protocols like agent auth and so on, uh, because we're trying to grapple with like, okay, how do we how do we handle this new kind of entity in the enterprise? It's not a service account, it's not necessarily a traditional NHI, but it's not a human user either. Um, and so it's really problematic and interesting. I was gonna ask you as well, like, you know, the we talked about you know autonomous exploitation, the rise of uh AI-driven discovery of vulnerabilities. On the flip side, on the defender side, knowing you're talking to a lot of startups and practitioners, how far do you think we are from seeing enterprises try to leverage, you know, uh autonomy or autonomous defenses uh in the enterprise side in production, try to kind of keep pace with what's happening now and what's coming?

SPEAKER_01 21:05

I think it's still early stage. And we see that uh AI-driven pen testing is happening. Now, are companies using the pen testing results to close the loop with their defense layer? Not yet. Right. And so we are um one of my uh investments is a company called Stryker, and and Stryker is one of the companies that actually does that. So they have a very strong AI pen testing layer, and that uh that layer is integrated with their guardrail and the defense layer. The defense layer takes information from red teaming findings directly into their policies. They are one of the few can do this on the market today. And certainly, if you look across the practitioner space, like companies, very few companies have this today. So I think that is something. Something absolutely have to happen. And another thing that needs to happen is not just pen testing from outside. We have to pen testing from inside as well. And before anybody else externally finds anything, we have to continuously like look at ourselves and say, hey, what's what's missing? What's a gap? We have to fix it. And the window to remediate things is really diminishing.

SPEAKER_00 22:28

Yeah. Yeah. Well said. And uh, you know, speaking of striker, I like I said, I kind of consume a lot a lot of uh research and things like in publications and so on. They had a state of agentic threat report I found that was really awesome. People should check that out on their website. Um it was really good re uh report, a lot of great insights in there into the real world environments, not just hypothetical type stuff. So give it a look on their website. That's uh striker, s t-a-i-k-e-r.ai. So give that a look. Um, I wanted to ask you, you're my fourth investor as well. Like this month, I've kind of made a little bit of an investor sprint on the show. Um I had uh Ed Sim, John Sakota, uh Sid Trevetti as well, and you're you're kind of the fourth and final you know venture capitalist guest on the show. Um, you know, I'm curious, there's been a lot of uh shared optimism about AI security as a category. Um, your own newsletter has noted capital concentrating in the extremes, while Series C uh C and B you know stays somewhat constrained. You know, do you think we're funding a generation of seed stage AI security companies that may struggle to pass that gap? Uh, just seeing some of the seed round sizes and things like that that are happening. Like what's your perspective on the kind of venture landscape of AI security right now?

SPEAKER_01 23:35

I think there is a danger of concentration for sure. Um, but that doesn't just exist in the security space. Um, historically, we have not seen this much concentration of capital in one category, which is AI. Right. And if you are uh coming from a just a pure finance point of view, you said you could look at this and say the last time it happened was actually in the 20s when railways came to the market, and there's a huge concentration of investment. And that market back then is really small. So the you know, the percentage of the capital uh concentration is about, I think it's the same as today's concentration AI. So there's a big risk. Um, and hence there's a risk in sort of a downstream risk of concentration AI security companies. And those large seed rounds are, I think, historically an anomaly, right? So, because you want in the seed investment, in the early stage investment, you want to take risks. You want to give different ideas a chance to foster themselves. And some of them won't work and others will work. So you give a little bit of capital for them to see uh where they'll uh progress. But these days, you have a concentration of capital behind a very early stage risky idea. I don't know. I'm not I'm not a fan. I also think there's a big market risk in um, you know, we saw um the concentration of four or five frontier labs taking most of the um capital investments. But you would agree there's a pendulum swinging back a little bit. Folks wanting to have own their own models and have their own small language models or smaller parameters on premises that they can control, train their own data. And certainly the open way models are coming up. The Thinking Machine Lab has a new open way model and K3, they're all becoming better and better. So the pendulum is swinging back. The question is how is the capital distribution going to happen? So is it gonna be a smooth transition to a concentration of capital of five, six labs to 20 providers, let's say? Or is it going to be a big boom crash of the market? No, right? And so security is kind of at the sort of receiving end of the market dynamic. And if the AI market crashes, I'm not an alarmist, so I'm not trying to say that. I'm just saying from an economist point of view, you know, that's a a risk. And if that happens, it will really impact the security market. And so I'm catching it, you know.

SPEAKER_00 26:25

Yeah, you're absolutely correct. Like, and I've been watching that too. And like I, you know, I find myself now following, I don't want to say uh alarmist or doomerism kind of stuff, but people who are a bit more skeptical of the spend and the build-out and the level of capital, concentration of capital, because I do want to also consider the alternative perspective that this just won't go on forever and it won't, like this isn't the norm. And to your point, I love that you you kind of said that uh security is uh, I forget exactly how you framed it, but you know, we're at the mercy of the broader market. Like security is a subset of IT spend, right? Uh in terms of budgets and and priorities and emphasis and so on. It tends to be somewhat resilient, but it's not entirely resilient. It will inevitably be impacted if things do change. Uh, and you you called out something that I guess I want to double down with you a little bit. I didn't intend on asking this, but uh the pendulum swinging. It's definitely a uh a reverse of trend lately. We had the the gating of uh mythos, the export ban, the you know, the uh you know that. Then there's the gating of the uh other frontier models from open AI and so on. There's definitely been uh a big outspoken uh ness among the community now around open open weight uh models. You know, Jensen Wong made his first ex post this week, I think it is, on the topic. There's been a whole coalition around this. Um, but you know, when I look at it from the security perspective, I think we're needed in both scenarios. We needed to secure frontier uh model usage and consumption. We're needed, of course, to secure uh self-hosted open weight models and their utilization in the enterprise. So, you know, thankfully we do, I won't say benefit, but we are needed in both scenarios. Um, when you look at that and see enterprises like how do you, when you talk to practitioners, for example, how do you see them kind of evaluating which to use for which use case, whether we self-host and or whether we use an open model, whether we use Frontier, and what kind of drive, what is driving some of these decisions for these enterprise uh tech and security leaders?

SPEAKER_01 28:09

I think it's a difficult decision to make. Uh, you really need someone who is versed in AI and AI infrastructure to help you make the decision. And also some of these open way models are not that walk in the park to just like host, you know, it's not like plug-in and works. Um so you need a team of people to babysit it. Um so if you don't, then you kind of up the uh creek without a uh paddle. Um, but if I rise up a level a bit, I'm in the camp of we must support open weight movement for the reason that if US doesn't do it, the other countries will do it. We can't stop them, right? So we essentially give them the opportunity to innovate. Um I think that's a about as stupid a thing as can happen. Um so we as an industry, as a country, we must support open weight movement as well as continue to invest in frontier research. Now, uh that said, from a specific enterprise point of view, how do you make decisions? Um if you have, if the tasks you're trying to do does not touch proprietary data, does not involve proprietary workflows, maybe a frontier API is all you need. And because you're not gonna be able to build a model that competes with them for the depth of the knowledge, for the number of years they've spent training the model. But if something that you're trying to do is very specific, involve data or or knowledge that they don't necessarily have, right? Then you probably should look at open way of models. And and if you don't have the capabilities to, maybe you should bring a third party to help you do. Uh and that's a sort of easy kind of answer, but there are lots of shades of gray within that answer as well.

SPEAKER_00 30:08

So yeah, no, it's uh I agree with you completely. And it's like it's again, it's not the non-either or there's nuance, like organizations inevitably are going to use both. They're gonna use frontier models for certain use cases and capabilities, self-hosted in certain scenarios. And to your point, it's not just an easy button, like, well, we'll just use open source. Like there's total cost of ownership, there's deployment infrastructure, maintainment, uh, maintaining it, you know, securing it, et cetera. Like, so it's a big consideration. And I think we're what's exciting to me is we're such like early in this whole kind of conversation. Like there's so much work to be done and so much innovation still to come. And uh, but I agree with you, like we shouldn't cede the lead to other nations by you know taking that blocking type approach or or being restrictive, because as they say, uh constraint breeds innovation. I think we've seen that with uh chip constraints. And if we do that with the models, we'll see similar outcomes there. Um, so I do hope we see kind of an open innovative ecosystem in the United States here around uh model development uh and so on. Uh I wanted to bring it home with the last question for you, uh, you know, maybe back to the practitioners, like for security leaders listening who maybe they don't invest or they don't work at a frontier lab, you know, what should they be either doing differently or thinking about doing differently over the next 12 months to get ahead of some of this AI-driven exploit, you know, uh dynamics that we've kind of been talking about throughout the conversation.

SPEAKER_01 31:22

Yeah, so I think as a uh technology leader, right, first of all, you should look at where in your company AI is being used. Um, because it's definitely being used. And then you should say, okay, is it being used in the most effective manner? Um there places gaps that we could use AI to leverage that we are not, or places we're not should not be experimenting with AI but we're throwing all kinds of token costs, right? So doing kind of reconciliation, if you will, uh as a home as a homework. And then looking forward to the next 12 to 24 months and say, hey, where are the places AI can create value, not just like making it more productive or making it more cost effective, but really create net new value, a new strategy, new products, new services. That's where I would concentrate my effort on.

SPEAKER_00 32:20

Yeah, yeah. I love that answer because it's not, you know, efficiency and cost reduction and like, I mean, not that those don't matter, they're not relevant, but it's a highly hyper-competitive market. What can we do differently? How can we innovate? How can we differentiate? How can we provide more value to our customers and stakeholders? Um, so I love that kind of answer right there. Well, thank you so much, uh, Chenxy, for jumping on. We had a bit of uh some snafu with some tech stuff early on, but I'm glad I got you on here and I'm looking forward to seeing you next week at uh Black Hat as well.

SPEAKER_01 32:47

Yeah, travel safe and I'll see you in hot vegas.

SPEAKER_00 32:51

Yes, yes, ma'am. Talk to you soon.

SPEAKER_01 32:53

Okay, bye.