[00:00:00] [00:03:55] Zainab: I'm Zainab I am one of the co-founders of asymmetric security. [00:04:01] Jacob Haimes: Awesome. And in one sentence, how is cybersecurity relevant to AI safety? [00:04:10] Zainab: Good question. Um, I like to think of cybersecurity and AI safety as kind of two sides of the same coin. It's different flavors of risk management, [00:04:26] Jacob Haimes: Awesome. That's the best, uh, one sentence answer I think I've gotten so far. Most people try to extend it a little bit. Um, okay. [00:04:34] Zainab's Background --- [00:04:34] Jacob Haimes: So I know, I know some of your story here, uh, because we, we've worked together previously. Uh, but there are also bits that I'm less clear about. Uh, and it would be great to hear just on your side, like. [00:04:50] Jacob Haimes: What was the path that you took to become co-founder of a startup that raised 4.2 million pre-seed. [00:04:59] Zainab: Yeah. So, um, where to start here? I think, let me, let me rewind a few years. [00:05:07] Jacob Haimes: Okay. [00:05:08] Zainab: so my, my background is actually originally in physics. Um, I kind of zeroed in on computational astrophysics at college, and the reason I mentioned that is I was like always broadly interested in statistical analysis and, you know, anomaly detection and, and thinking how, thinking about how systems can work and how to break that down mathematically. [00:05:31] Jacob Haimes: Okay. [00:05:31] Zainab: I then, after my master's left academia and kind of sidestepped a little bit into a career in cybersecurity, um, but maintained that. statistical frame on the problem. And the part of security that I, um, ended up working in and, and very much chose to work in is this field called Dig Digital Forensics and Incident Response, um, which a lot of people are less familiar with, but the, the way to to think about it is when a company or individual gets hacked, um, their nine one one call will be to a digital forensics and incident response team. Um, and they will be responsible for guiding the person or organization through the hack, figuring out what happened, what data was touched, if the hackers are out, or if they've managed to persist in some way. Um, and just answer that will obviously be, um, top of mind for people who are impacted There. Um, [00:06:37] Jacob Haimes: Gotcha. [00:06:37] Zainab: that at a company called Strass Friedberg, which, which specialized in DFIR specifically [00:06:44] Jacob Haimes: Mm-hmm. [00:06:44] Zainab: and got the opportunity to work on some of the biggest hacks of the past decade. Um, unfortunately a lot of them are private. I would love, love to share [00:06:53] Jacob Haimes: I was about to say, can you tell us the details? No. Okay. [00:06:56] Zainab: Um, but the one that is public is we did a ton of work, uh, with Facebook in the wake of the Cambridge Analytica, um, [00:07:04] Jacob Haimes: Okay. Okay. [00:07:05] Zainab: sort of blow up and, and a lot of that was figuring out, you know, what, what, what actually had been accessed there. Um, so, so I did that for the better part a decade. Um, and then got more and more interested in ai. On the side. [00:07:25] Jacob Haimes: Mm-hmm. [00:07:25] Zainab: but I suppose viewing it through more of a security lens of what, what are the, are the risks, but also, you know, um, how can we, how can we leverage this technology to make our defenses stronger? [00:07:42] Jacob Haimes: Mm-hmm. [00:07:43] Zainab: Um, this sort of, this like defensive accelerationist theory, um, be behind some of the thinking that I was, that I was doing then or in, in tune, in tune with the thinking that I was doing Then, um, and when I kind of was, was doing my, my day job of pretty manually scrolling through. of log files figuring out, you know, what, what in here looks suspicious. [00:08:11] Zainab: What in here is, is probably fine. Um, realize that this sort of new wave at that point of, you know, the chat, the chat GPT movement, um, of LMS could potentially really change the game with [00:08:25] Jacob Haimes: Mm. [00:08:25] Zainab: regards to how this work is done. Um, and these situations are fundamentally crisis situations. So if you can do those investigations. Significantly faster, uh, significantly better, and for significantly less money. Um, it's not that hard to see why that would be incredibly useful. [00:08:49] Jacob & Zainab's History --- [00:08:49] Zainab: Um, so was sort of thinking, thinking about these things, was getting more and more interested into the AI side of things. Um, and then around that time, uh, met you Jacob when I was doing, um, one of the blue.courses to get more of a lay of the land of what AI safety and security looked like. and at that point was considering kind of pivoting to something that was more research focused, among, among, among kind of many other things. and then I remember, uh. Very, very vividly. It was a conversation that you and I had of sure that I was aware that the like security background that I'd come from was very much like a strength that could be brought to the AI safety space rather than something that I was potentially falling into the trap of wanting to explain it away because I didn't have a typical background. [00:09:56] Jacob Haimes: Yeah. [00:09:57] Zainab: so yeah, big, big, big thank you on that front. [00:10:00] Jacob Haimes: I'm glad that I was helpful there. Um, so then, you know, we, I, I, we meet and I'm like, no, you should pursue this. And you're like, okay, yeah, maybe I should. And then you take it and you go, you go with it. Uh, and I, I believe, I mean, I'm, you, I'm sure you're doing other stuff, but I, I reached out 'cause there was a, a team that I was working with, uh, at the time that, uh, could use someone with your experience and expertise. [00:10:26] Jacob Haimes: Um, and that's when we, we looked into Cybersec eval, uh, and like how it could be improved, I guess. Um, and then, uh, yeah. What, what was that time period like for you? [00:10:44] Zainab: Yeah, I think that was, that was super interesting. I think that was like my, my foray into just like the more academic part of the field, but also just like. AI safety, um, more generally, like getting, getting deeper into the ecosystem there. [00:10:59] Jacob Haimes: Mm-hmm. [00:11:01] Zainab: I think what I realized quite quickly is that the team that, that we were part of, um, so Jaime and Suhas or incredibly talented peers, um, uh, their wheelhouse was more on the engineering side of things, which was super valuable. [00:11:20] Zainab: And then I kind of, um, joined as, as a bit of a cyber subject matter expert [00:11:26] Jacob Haimes: Mm-hmm. [00:11:26] Zainab: understood the value that that could bring pretty quickly. 'cause a lot of, when we, when we picked apart Cybersec eval and for context, it was, um, one of the first like Lama papers that, that, that meta, that meta released. Um, some of the issues that we found found with it weren't just. about, you know, this, this, this code could be better. It was that the fundamental axioms that relate more closely to, um, the kind of security theory [00:12:00] Jacob Haimes: Mm-hmm. [00:12:01] Zainab: um, were kind of mis misplaced. Um, and it was sort of over reliance on certain tools that, um, security testers were using quite a lot. [00:12:15] Zainab: And, and not thinking about how these problems are, are carried out in the real world. So I think it was super interesting to start like weaving those threads together and understand the value of, I mean, people, people talk about cybersecurity all the time within the like, AI safety space, but then not that many, this is now changing, but I think at that point there were fewer kind of, you know, cyber SMEs involved. [00:12:43] Jacob Haimes: Mm-hmm. [00:12:43] Zainab: Um, and the byproduct of that is like. Engineers kind of to overfit on, on certain cybersecurity principles without grounding it in, in what's actually going on. Um, so it was super interesting to realize that that was the case, that that was the case, and also really like exciting and fulfilling to like play a role in, in closing that gap for, for one research project. [00:13:10] [00:16:03] Founding Asymmetric Security --- [00:16:03] Jacob Haimes: And then, and then what happened next? [00:16:06] Zainab: So what happened next was, we, we continued working together on that team, um, published, uh, I was. Working on a different research project at the same time. Published a couple of papers, um, at Europes and just got more and more involved, um, with people within, within the sort of space. And like went, went to a bunch of like networking events in London and kind of applying the philosophy of like, if you put yourself in the right places enough times at some point it'll end up being the right time. [00:16:40] Jacob Haimes: Mm-hmm. [00:16:42] Zainab: and through, through that ended up meeting Alexi, who is one of my co-founders. Um, he was on entrepreneur first, um, startup accelerator at that point, [00:16:55] Jacob Haimes: Mm-hmm. [00:16:55] Zainab: and wanted to build a company in the AI cyber space. his background is more on the AI side, um, and was looking for like a, essentially a cyber SME to, to, to work with. [00:17:10] Jacob Haimes: Mm-hmm. [00:17:11] Zainab: we then started working together, um, and things went well. Um. [00:17:17] Jacob Haimes: Quick question, do you know, do you know who, uh, um, who suggested you [00:17:24] Zainab: Um, so the person, so sort of two, two degrees of separation. Um, so Espin was the one who actually connected me and Alexi and he was in touch with a few people on entrepreneur. Um, so [00:17:39] Jacob Haimes: and who, who do you think suggested you to Aspen? I, [00:17:43] Zainab: I actually, was it you? [00:17:46] Jacob Haimes: yeah. [00:17:47] Zainab: way. [00:17:48] Jacob Haimes: Yes. [00:17:50] Zainab: I did not know. [00:17:52] Jacob Haimes: Yeah. But [00:17:53] Zainab: Um, [00:17:54] Jacob Haimes: uh. [00:17:55] Zainab: thank you so much. That is [00:17:57] Jacob Haimes: O of course. [00:17:59] Zainab: Um, this is a full circle moment. Um. [00:18:03] Jacob Haimes: So you got in touch with him. And from what I know from the side of the story is basically like what Alexi said to Bin was like, was Zainab is just leagues ahead of all the other candidates that I've been, uh, you know, looking into. So thanks so much for that. Right? So that was, uh, extremely valuable for him. [00:18:27] Jacob Haimes: Um, and then you guys just sort of hit it off and, and he was already an entrepreneur first, and so it just started going. [00:18:34] Zainab: Yeah. And then, yeah, I joined the program. Um, our, then we, so, um, the cohort that we were on was, um, effectively run by, uh, somebody called Pippa. Um, she is our third co-founder, so now it's a team of three, three founders. Uh, we all forces and yeah, we're and building and, and, and doing a bunch of fun stuff. [00:19:00] Jacob Haimes: Awesome. Um, so that catches us up mostly, well, that actually catches us up to, I think, mid 2025. And so since then, you've just been continuing to, to move forward there. Is that it? I, [00:19:15] Zainab: Yeah, basically we, we, um, exactly that. Um, we've, we've been continuing down that vein. Um, the work that we do. Um, this is a good time to, to, to intro that is, um, building a kind of AI native digital forensics and incident response firm. Um, and what [00:19:39] Jacob Haimes: okay. [00:19:39] Zainab: like is like providing the service, um, that I used to provide, um, in my previous job. Um, but building AI tech to that faster, better, cheaper. [00:19:55] Jacob Haimes: Gotcha. And before we, I like, we're, we're definitely gonna get into your methodology and, and what you're doing there, but before we, we do, um, I do want to sort of take, uh, a meta stance and ask a couple of questions more generally about cybersecurity to make sure we're all on the same page. [00:20:14] Zainab: For [00:20:14] Jacob Haimes: What does an actual cyber attack look like? [00:20:17] Jacob Haimes: Like what does, what does that mean in practice? [00:20:22] Zainab: Yep, for sure. Um, so I think that there, there, there are some which are more common than others. Like a, an easy one that comes up time and time again is like a user clicks on a phishing email that allows an attacker to gain initial access into a network. Then they'll be able to, um, once they're in, they can, do whatever they want within the user's email. [00:20:48] Jacob Haimes: And how is that leverage? Like what? What's the incentive here? [00:20:53] Zainab: yeah, for sure. So there are a few, um, and they're kind of two, the two categories. One is sort of spray and pray attackers [00:21:02] Jacob Haimes: My favorite, [00:21:03] Zainab: who are like playing a volumes game and they're not that sophisticated, [00:21:08] Jacob Haimes: thank you. [00:21:09] Zainab: to just like. Hack people get them to click on phishing emails if they, and then, um, one, one mode that we see quite a lot is a user will click on a phishing email. The attacker will keep emailing out phishing links to more people until they get to someone who has some kind of purchasing power. [00:21:29] Jacob Haimes: Mm. [00:21:30] Zainab: then they'll send an email saying that, oh my, my bank account details have changed. Send the money here. Um, so they'll effectively manage to wangle through some payment diversion fraud by doing that [00:21:45] Jacob Haimes: Okay. [00:21:45] Zainab: in a similar kind of spray and pray bucket. These tend to be financially motivated as you can, like, see from the [00:21:52] Jacob Haimes: Mm-hmm. [00:21:53] Zainab: for example. Um, you also get ransomware attacks. Um, and what that looks like is often the initial access is something like, uh, a vulnerable VPN. Um, you can also have like an initial phishing email as the root cause for that as well. But the attackers get in instead of just like trying to divert a payment, they will access to and servers on the network. Um, and then just, you can kind of the like class, the classic hacker meme of like your lap, your laptop is, has been locked, transfer X million pounds in Bitcoin to this address [00:22:33] Jacob Haimes: Mm-hmm. [00:22:34] Zainab: exactly what happens there [00:22:36] Jacob Haimes: Okay. [00:22:36] Zainab: all your files, you won't be able to access your data. They'll probably also steal a bunch of files and try and extort you with the threat of leaking that data too. Um, but effectively just like holding your digital data to ransom, [00:22:52] Jacob Haimes: Gotcha. Okay. [00:22:53] Zainab: like, that's bucket one, that's [00:22:56] Jacob Haimes: Okay. [00:22:56] Zainab: and pray and the motivation is like almost always financial. And these guys make a bunch of money. Um, they kind of operate like businesses themselves. bucket two is sort of, uh, people call them kind of a PT advanced persistent threats. those are the more sophisticated attackers who may try to, have much clearer targets. often the motivation there will be, let me steal some data, or let me gain access to some data [00:23:34] Jacob Haimes: So [00:23:34] Zainab: be. [00:23:35] Jacob Haimes: this is like nation states and or the like, uh, I guess hacker groups that have a presence. And it's not like, because we're criminals, that is because like we go in and find data and expose it or something like that. Like that's. Who these people are. Okay. [00:23:55] Zainab: Like, and nation states, I think are the key, the key ones to focus on here. [00:23:59] Jacob Haimes: Yeah. [00:24:00] Zainab: Um. [00:24:00] Jacob Haimes: A lot more of those, [00:24:02] Zainab: Yeah. so their, their goal is always let me, let me steal some data. And they will, they might be in their, their cases I've worked where people have been in there for years and gone undetected. whereas you can imagine sort of in, in the previous bucket of like the ransomware email compromise, they're pretty noisy. [00:24:25] Zainab: Like it's harder to be more noisy than locking down every laptop on a network and saying, pay me some money. [00:24:30] Jacob Haimes: right. [00:24:31] Zainab: Um, whereas for nation states and advanced persistent threat groups, um, will be as quiet as possible, as stealthy as possible. Um, and you really need to kind of follow the smallest of digital breadcrumbs to understand what's happened. [00:24:49] How to Know Who You Can Trust --- [00:24:49] Jacob Haimes: Gotcha. Okay. And I guess the last like more meta thing about at least just cybersecurity is so I, you know, I'm someone who doesn't trust corporations, uh, at all. And when I think about that in the context of cybersecurity, um. Vendors are incentivized to make things seem scarier than they actually are, right? [00:25:13] Jacob Haimes: Because then their services are in demand. If you say, oh, look at this scary thing that might happen. If you don't, you know, pay for my services, then that's good for business. And then model developers also do this and they benefit from these displays. And so we get things like Anthropic announcing, uh, that some group attempted to use their models to conduct a cyber attack. [00:25:35] Jacob Haimes: And then, like, just a week ago, or a week and a half ago, two weeks ago, something like that, uh, they announced, uh, their project Glass Wing effort in conjunction with the launch of Mythos. And there's a whole bunch of press about it being, you know, like too strong to release. And then there's also now like counter press and, you know, there's a whole, uh, blow about that. [00:29:15] And it's like, well, okay, who, who are we gonna listen to? Who do we trust here? Uh, and then especially given the additional context of like, oh, and they're also incentivized to. Make it seem like a bigger deal than it is. Uh, how do, how do you determine when something in, in the news or by that someone else in a, a cybersecurity company, uh, is saying is, uh, you know, accurate is, is truly representative of the situation? [00:29:47] Jacob Haimes: Uh, and then, and then the, the next thing is like, how can, how can we trust you because you are also in that position? [00:29:57] Zainab: Um, interesting. So I think there are a few, there are a few things. So I'll start with how I, how I kind of. Fact check, um, [00:30:06] Jacob Haimes: Okay, [00:30:06] Zainab: build or build my own conviction in something. [00:30:09] Jacob Haimes: sure. [00:30:10] Zainab: Um, I think that contextualizes the next part of the question as well. [00:30:15] Jacob Haimes: Okay. [00:30:15] Zainab: I'm, I, I mean, I'm just, will to make sure that a claim can be substantiated. [00:30:22] Zainab: I just, I need to see some, some kind of hard evidence behind that. Um, if we use Mythos as an example and kind of the, the increase in, in, in cyber capabilities, um, for models, um, over, over the, over like very recent times. Um, basically I think you, when you look at, when you look at the, the trend line. Of, you know, how, how many CTFs are actually being solved here with, with success? Um, I think AC issued a, a, a kind of paper on, on, um, the mytho capabilities where, you know, you can very clearly see a trend line going up and there is like data to substantiate it. And when you dig into it a bit further, this, these attacks are representative of what cyber attacks tend to look like. [00:31:18] Zainab: It's not just like a shoehorned subset of the easiest attacks ever and kind of trying to inflate, inflate that and like really kind of cast a critical eye on what, on what the data is saying. and I think just like making sure that those claims are grounded in, in facts and not just like mongering without any substantiation is like a key, key thing for me. I think a a point at which. One, one thing that I watched, which, which I thought was a really good resource on this, was, um, Nicholas Kini, who is at Anthropic, uh, did a talk at a conference think a few weeks ago now called Unprompted, [00:32:04] Jacob Haimes: Mm-hmm. [00:32:05] Zainab: demo of, of, uh, kind of Linux vulnerability, uh, being found, um, by a model. Um, I think things like that of, you know, see, seeing, seeing things happen in real time and understanding what, um, what the consequences are of that, and just like the steps involved in these attacks, um, can be, can be really useful to conceptualize exactly what the risk is here. Um, and it's things like that that I find most, most helpful to just like build conviction that this is genuinely something to worry about. And I think even though that. Sort of of corporations aside, there are individuals, um, who may or may not be affiliated with, with those corporations. And you know, I get, I get that it's, it's, it's a gray area when it comes to trust, um, who opinions that'll hold in high regard. Um, and just like understanding the broader landscape through that lens as well is something that I also find helpful. [00:33:20] Jacob Haimes: Okay. [00:33:20] Zainab: like gut reflections on that. Um, possibly a more trusting, um, than yourself Jacob. [00:33:29] Jacob Haimes: I, I mean, I think that's definitely true. However, that's not to say that that's wrong. [00:33:35] Zainab: Um, then how, and then the second part of your question of how, how can we trust you? [00:33:42] Jacob Haimes: Mm-hmm. [00:33:44] Zainab: Um, I think there are a few parts to this. Um. It ties it, the con the context that I've given before of like making sure things are substantiated in real world examples. And, you know, I feel like an AI real world can cannot really mean real world. [00:34:01] Jacob Haimes: Mm-hmm. [00:34:02] Zainab: but making sure that the how, you know, make sure your evals are actually realistic. So at asymmetric we run a bunch of evals, um, on, sort of um, log files that are from, uh, real world cyber attacks. Um, and the real world is like key there. It's not just like synthetic data that's been shoehorned to make sure that the, the model that we're using is, works really well on that and makes us look good. It's just like this is like a wide variety of. Attacks that we've seen, and this is like how, how our tech does. Um, so I think like the key a sort of crux of this is like making sure that there isn't sort of a dissonance between the, like how, how realistic the thing that you're testing for is and, and what that actually looks like. and then I will also ground like my, my job is very much giving people the necessary recommendations when they're mid cyber attack and trust is a massive component of that. [00:35:13] Jacob Haimes: Mm-hmm. [00:35:13] Zainab: like I am not doing my job right if my client doesn't trust me. Um, [00:35:18] Jacob Haimes: Right. [00:35:19] Zainab: so, so I do, I do think about it quite a lot and I will ground whatever recommendations I give them in the attack that they're facing will also. Caveat when necessary of, you know, these are the risks. This is what we've seen happen so far. Ultimately, is a business decision to choose which level of risk you're willing to accept, [00:35:50] Jacob Haimes: Aim it always. [00:35:53] Zainab: um, and, and make sure that it's clear that, you know, these are the consequences if you go by A. Like if you're to issue everyone in your organization a UB key. [00:36:05] Zainab: So they have to put their fingerprint in whenever they log into something. Possibly that's like the best, the best way to authenticate. It's also a massive pain if you are like a hundred thousand person organization, [00:36:19] Jacob Haimes: Yeah. [00:36:19] Zainab: Um, so making sure that the trade offs are clear, is something that I think is like really useful for just like build building trust myself. [00:36:31] The Threats Asymmetric Is Built to Fight --- [00:36:31] Jacob Haimes: Gotcha. And then, so like, what, what does this, uh, intersection of, of, of cybersecurity and AI safety, like, what is it that you feel you are. [00:36:45] Jacob Haimes: pushing forward here? Uh, like what, maybe, like what is the, the, the threat that you see, um, and, and how are you addressing that? Is, is maybe the best way to put that? [00:36:56] Zainab: Yeah, I think I, I think there are a few, I think there's one matter thing of, of attacks always happen. The thing that we're addressing is how, how quickly you can respond to them. [00:37:07] Jacob Haimes: Mm-hmm. [00:37:08] Zainab: Um, and what, what that means is, I know, I mean, in the UK we had MNS Mark expenses had a massive attack. recently, Ja Jaguar also had a massive cyber attack recently. The reason I mentioned these is those companies were down for months. Um, and a lot of that downtime is because you don't know which systems are infected, so you don't know what you have to rebuild. And when, um, and if you're able to respond faster, then you can minimize that downtime. And if you're called in when you know, the ransomware detonation hasn't quite happened, so it's sort of pre big event, but something has gone wrong and you're able to respond super quickly, um, you can kind of stop the big event. Um, so that's sort of 1, 1, 1 part of what what we're doing and what we're solving for, of making, making that process much faster. The second part is, I think where kind of in an era where, um. vulnerabilities in code that attackers can exploit, has never really been easier. Um, and I think is only going to accelerate, um, my career in cybersecurity, they've kind of been these few and far between watershed moments where there's been a massive surge in incidents because some vulnerability has materialized on, um, an application or on, you know, a, a type of server that is super widely in use. [00:38:56] Jacob Haimes: Mm-hmm. [00:38:56] Zainab: Um, and I think that the, the future could, could involve more of these surge events. Um, and if you're able to kind of respond much faster and build tech to, to make sure that, that the quality of that response doesn't waiver, [00:39:18] Jacob Haimes: Mm-hmm. [00:39:19] Zainab: we're trying to build a team that is and built for those surges. Um, the upshot is the tech scales really, really well. [00:39:31] Zainab: Rather than, uh, incident response analysts don't sleep for the next five days. [00:39:37] Jacob Haimes: Right. Although that might happen too. [00:39:40] Zainab: No, no promises. [00:39:42] Zainab: Yeah, I think, I think it's interesting 'cause there are definitely these very two distinct buckets of like spray and prey and then, uh, very high sophistication. And from the perspective of an analyst, those two cases look super different. [00:39:55] Jacob Haimes: Okay. [00:39:56] Zainab: in the former you are, you kind of know what you're looking for, the moment you get the call, um, the will often just use the same, um, tools over and over again, which translates into the same kind of artifacts or like digital fingerprints that you can follow. [00:40:15] Jacob Haimes: Because the main vulnerability is the human. [00:40:20] Zainab: um, possibly to get initial access. But it's also just like. If, if, for example, they got access to my laptop and wanted to escalate to, you know, a server, um, they will use the same types of like, malware, um, to go through that path. [00:40:38] Jacob Haimes: Okay, [00:40:38] Zainab: will always use the same. There, there are a bunch of tools that, um, can be used for remote access. [00:40:45] Zainab: They'll always use those. [00:40:47] Jacob Haimes: gotcha. [00:40:48] Zainab: they won't, they won't really be operating at the cutting edge. It's sort of like, if it's not broken, why should I change it? [00:40:55] Jacob Haimes: Yeah. [00:40:55] Zainab: Um, so you have that on one side and then on the other side you have, you, you basically don't really have any idea what you're looking for. And the hardest part of those investigations is like finding the initial investigative threads to pull on and just finding like bad activity number one, that you can then. Pivot off. Um, in the first case, you could probably get a significant of the way there by just like searching for some keywords. I'm oversimplifying a little bit, but just as like a bit of an intuition pump of like, this is the name of a, a piece of malware that's always used. Is it there? Um, in the second case, it's it's far more complex. You need to reason a lot more. You're kind of building, building confidence yourself. It's sort of like stacking lots of conditional probabilities on top of each other. And then until you get to the point where like, this is definitely bad for X, y, z reasons. [00:41:56] Jacob Haimes: Okay. [00:41:57] Zainab: if we think about it in the kind of. Context of AI what AI will do for defenders is, um, at asymmetric, our initial focus is on the kind of lowest sophistication, but high volume attacks and getting good at those. The work that I'm most excited for is getting good, um, getting AI agents good at the sort of more sophisticated, um, lower volume attacks because the work is just harder. [00:42:27] Jacob Haimes: Mm-hmm. [00:42:27] Zainab: Um, and the thing that we need to build is the ability to kind of manage those surge events that I was talking about previously on both ends of the spectrum. [00:42:41] Jacob Haimes: Okay. So if I were to summarize what you just said in relation to the, the question, it's, it's basically like sometimes yes. Uh, sometimes you, you will, uh, like at least ideally, uh, groups will be able to deal with the, the surge of advanced threats. Uh, however, that's probably like, at least in your eye is not gonna be, it's not like all threats are going to become advanced. [00:43:10] Zainab: Yeah, [00:43:11] Jacob Haimes: Okay. That makes sense. That makes sense. And that's maybe a, a really good like segue into what does it, you actually do. So asymmetric security and like your job, but within asymmetric security, what are you doing? [00:43:26] Zainab: yeah, for sure. Um, so we are, uh, sort of AI native digital forensics and incident response firm. And what that means is, um, we will, we. I think the easiest way to think about this is, uh, in the context of my previous job at straws. [00:43:44] Jacob Haimes: Mm-hmm. [00:43:45] Zainab: the way the process would work is, um, a junior analyst would kind of take the first pass through all of the data, find initial investigative threads, figure out what's, what's gone wrong, and then if there was anything more complex or needing needed digging into further, um, that would get passed on to a kind of senior analyst. [00:44:05] Jacob Haimes: Mm-hmm. [00:44:06] Zainab: Um, and that senior analyst would also be responsible for overseeing it and, you know, um, managing the investigation. we have built, um, so far and what we're building is sort of an AI agent that does that first pass. [00:44:19] Jacob Haimes: Mm-hmm. [00:44:20] Zainab: and it does that first pass, um, in, we're foc focusing right now on cases that are scoped to, um, email because that's a super common one. [00:44:32] Jacob Haimes: Mm-hmm. [00:44:32] Zainab: also. It's essentially the easiest place to start. Um, and those investigations would take a junior analyst a couple days, maybe a bit longer. Um, we have built tech that does that first pass in minutes. Um, and what that enables us to do is move much faster, and kind of cost a wide net when we do the investigation. [00:44:59] Jacob Haimes: Mm-hmm. [00:44:59] Zainab: if, if a client comes to us and says that, you know, I'm aware of one account being compromised, we will make, we will, look into every single account in their organization and check that everything's okay there. Um, a lot of traditional teams don't have time to do that 'cause it's so labor constrained. Um, and we're also able to do it far more cost effectively. 'cause again, it's less con constrained on. hours, we will always have a kind of senior analyst who is walking the client through the findings, um, and digging into things as and when necessary. Um, so the key thing is, you know, augmentation rather than automated. Um, and I think there's like an interesting here of even at the extreme of, um, an AI being able to do the full investigation perfectly. Um, because it's such a high stake situation, um, your client will want a kind of accountable person on the other end of the phone, at least for the next bit of time, I think. Um, so making sure that there's that close into play, um, between the tech that we're building and our analyst is like, and our analyst is a key component of that. [00:46:20] Jacob Haimes: Gotcha. And that's also then a good, you know, uh, shift to what, what is it that you do? Are you just doing that, uh, instant response, uh, like senior analyst position, or, or is there there more going on there? [00:46:38] Zainab: Yeah, for sure. So my role is very much kind of leading product and what product means for us is like leading, leading that delivery of service. [00:46:48] Jacob Haimes: Okay. [00:46:49] Zainab: so I think that the kind of two, two parts of my job, one is leading, leading our team of cyber analysts, um, when they're working these cases, [00:46:57] Jacob Haimes: Mm-hmm. [00:46:58] Zainab: helping them out whenever necessary. Um. We have, yeah, an exceptional team. Love working with all of them, um, and making sure that they're supported and unblocked whenever possible. And then the second thing is that the kind of two key teams that we operate are the cybersecurity analysts and our engineers. Um, and my, my, my, one of my key responsibilities is kind of sitting between those two, [00:47:26] Jacob Haimes: Mm-hmm. [00:47:27] Zainab: and making sure that there's a very, very tight feedback loop between the two. [00:47:32] Jacob Haimes: Okay. And then, so just to, I guess, like map out. How, how things are, are working. The cybersecurity analysts are the ones that are, are basically, uh, customer facing. Uh, and then the engineers are improving your internal systems, is that correct? [00:47:48] Zainab: Yeah. [00:47:49] Jacob Haimes: Okay. Um, [00:47:51] Jacob Haimes: and in practice, what, like, at least right now, what are, what are the limitations, right? You said, you know, you don't think that it could be a full, uh, like conduct the full investigation yet. Um, and regardless, there still needs to be, uh, some amount of human accountability. But like, um, what are the, the points where it's likely to fail that you're trying to address now? [00:48:18] Jacob Haimes: Uh, and how have those changed over time? [00:48:21] Zainab: Yep, for sure. So I think, um. Our initial focus was on email cases, and we're approaching, kind of approaching saturation there where, um, the tech that we've built has, does that first pass almost perfectly? I think accuracy, the on on our last set of evals was at 98%. [00:48:41] Jacob Haimes: Under what assumptions? Like, so does that mean you're assuming, like you've mentioned earlier, like typically those sort of spray and pay attacks use the same sorts of methods, so one of the underlying assumptions would be like, that stays the same, I guess, if, if that's correct. [00:48:57] Zainab: Yeah. So this [00:48:58] Jacob Haimes: Okay. [00:48:59] Zainab: um, this is an interesting part of it. So are a few, the thing that stays the same is kind of the arena in which they're operating in, and that's like specifically email. Often the initial access will be via phishing, but the way in which we investigate them is we will collect the various logs from the environment. [00:49:18] Zainab: So that would include a full list of every authentication. Over the past six months, seeing exactly where people have been signing in from, what devices they've been using, that kind of thing. [00:49:27] Jacob Haimes: Mm-hmm. [00:49:28] Zainab: so the anomaly detection process is sort of much broader than just like find the initial phishing email. It's, you know, where, what are the login events here that look suspicious? [00:49:39] Zainab: Are there any, um, applications that have been linked to the environment? Like people sign in with Microsoft or Google all the time on various applications. Is there anything there that looks weird? [00:49:50] Jacob Haimes: Mm-hmm. [00:49:51] Zainab: so it's that kind of analysis. So the assumption, to answer your question, the assumption, the assumption is that it's the, the attack is scoped to the email environment. [00:50:01] Zainab: They haven't gone anywhere else. Um, um, something has gone wrong. [00:50:09] Jacob Haimes: Sure. Okay. Okay, cool. [00:50:11] Zainab: but I think the limitation is that that is only one kind of incident and also the simplest kind of incident, which is exactly why we started there. 'cause it's both easy. Um, relatively easy and high volume. the, the limitation is that gets, this tech gets way more interesting when you think about ransomware or just like expanding the scope or, you know, moving towards this sort of, uh, super sophisticated nation states. [00:50:38] Zainab: Um, and that's where a significant amount of dev time is, um, dedicated now like, can we increase the scope of the data that we're reviewing So it's not just email and, you know, laptops, servers, um, a whole long tail of applications, GitHub, uh, infrastructure via AWS, um. Anything that you can kind of like firewalls, VPN logs, anything that you can imagine being in scope within sort of any corporate network. [00:51:08] Zainab: Are we able to build tech that is extensible enough to cover all of that and also kind of cover the most sophisticated attacks [00:51:16] Jacob Haimes: Is there a context problem there? Because I feel like, I mean, essentially what you just said is like, let's look at more data and different types of data. Uh, and also as you sort of went up the chain there at least, and I could be slightly off here, but. My perception would be that the volume of data would also increase, not just the ha having different types. [00:51:38] Jacob Haimes: And so, uh, is that one of the issues you'd be like trying to address there? [00:51:44] Zainab: Yeah, I think, I think context is an interesting one. Um, and they're kind of, they're like several layers to this. [00:51:51] Jacob Haimes: Mm-hmm. [00:51:52] Zainab: so the, for an agent to be successful in these cases, um. The kind of several types of context that need to be maintained. One is kind of open source of like, what's out there already? [00:52:09] Zainab: What types of attacks have we seen, um, what kind of pattern matching can we do from that? Um, and that needs to be kind of held somewhere. [00:52:17] Jacob Haimes: Mm-hmm. [00:52:18] Zainab: Number two, um, and this is an interesting one of if you have A-D-F-I-R team that's responded to incidents, time and time again, [00:52:26] Jacob Haimes: Mm-hmm. [00:52:27] Zainab: build up this tribal knowledge of like, these are the things that we look for. Um, these are sort of like, uh, interesting log files that people don't even know really exist that we can pull on. Or, you know, there are these like random artifacts and random places that people stumble upon by chance. Um, that end up being super useful to map out exactly what's happened. So there's like that knowledge as well. then. three is like maintaining context from the case itself. [00:52:58] Jacob Haimes: Mm-hmm. [00:52:58] Zainab: I know that on the 1st of April, that was when the first suspicious sign-in was. Um, I know that, um, the attacker was doing a bunch of stuff on GitHub. Um, they like downloaded a bunch of code, for example, [00:53:18] Jacob Haimes: Mm-hmm. [00:53:19] Zainab: that. And then there's also kind of, um, what you were gesturing at of like these log files can get pretty big. Um, so if you're trying to analyze all of them, for sure, you're gonna max out the context window, um, if you were just streaming everything through. so a key, a key thing that we're working on now is, you know, how, how do you. Layer this properly. Um, how do you equip the agents with sufficient tools to call on those pools of knowledge? [00:53:51] Jacob Haimes: Mm-hmm. [00:53:52] Zainab: and how do you structure system in a kind of smart, smart way to make sure that you're not just like a brute force streaming data? Um, but it's more kind of powerful querying. The last one is sort of can, can be a reasonably simple one of equip, equip the agent with the ability to run database queries, [00:54:14] Jacob Haimes: Sure. [00:54:15] Zainab: and sure that not like loading in all of the data at once kind of vibe. Um, that's sort of thinking about it on easy mode. Um, but hard mode gets activated when you kind of try and put all the pieces of the puzzle together. [00:54:31] Jacob Haimes: Okay, cool. I'm just thinking about like the. Implication of that is if I have any other questions about that. But I, I think that the other aspect that I'm most interested in is just like, as you were building asymmetric security, what made you go like, oh, that's great. Or like, what if I did this? And then you try it and it just works. [00:54:55] Jacob Haimes: Or something clicks. Like what were those aha moments and [00:54:59] Jacob Haimes: what, like what led you to them is also I think, quite interesting here. [00:55:06] Zainab: Yeah. I think, um, I think useful context here is like I've, I've spoken a lot about. Analysis, but it can be difficult to conceptualize what that actually looks like and the process you go through get to the point of like, attacker a entered on X date did X, Y, Z, and then got kicked out. Um, the way to picture this is literally imagine a person, me, me, two years ago, um, with a hundred different Excel files, literally Excel files open on my laptop, and each Excel file will correspond to, uh, a log file that shows that these apps were executed on in the past two months, or, uh, these 58 services were installed in the past two months. Um, so the key, the key thing I'm trying to, um, drive home is super distributed data and just like reasoning across. Various tabs and figuring out that is for sure malicious, um, because the, the, the legitimate user never signs in from London. and then figuring out, okay, if I know that this is bad, then everything else that happened during this time where this person was logged in is probably also bad. Um, and like going through that process. So it's a bunch of like, people always say pivoting of you switching, switching between things, understanding what's going on, pulling, pulling a thread, and just like doing that investigation, it's, I, without LLMs, it's I think impossible to, think about automating that. And the reason behind that is that. You need, you really need like a level of reasoning behind it. It's not enough to just statically point, point, point things at, at various files and be like, uncover these keywords. And the, the caveat, the like caveat here, or like the key thing here is, um, by impossible to automate, I mean like impossible to automate fully, not like [00:57:29] Jacob Haimes: Sure. [00:57:29] Zainab: help you out a little bit. [00:57:30] Jacob Haimes: Mm-hmm. [00:57:31] Zainab: Um, so I think when I was still at straws at my old company, um, I was simultaneously kind of getting more and more interested in AI and like doing a bunch of work, uh, with yourself, Jacob. And just like doing, doing these evals, [00:57:52] Jacob Haimes: Mm-hmm. [00:57:53] Zainab: or thinking about what these evals look like, and, and realize that, you know, if you can. A lot of people were writing about automating stuff on the offensive security side. but there was very little literature out there on the defensive security side. and I was sort of thinking about that. This is still just like scrolling through data. The, the like, uh, simplified version of it is you're scrolling through data and doing a level of reasoning. [00:58:26] Jacob Haimes: Hmm. [00:58:26] Zainab: Um, like what would happen if I were to put this random log file into cloud and prompt it with like find bad, um, like [00:58:36] Jacob Haimes: Besides, besides like a data breach. [00:58:41] Zainab: um, uh, put it put in anonymized data into Claude and [00:58:47] Jacob Haimes: Okay. Okay. [00:58:48] Zainab: um. And then what I did was I like generated a bunch of synthetic data. Uh, it was actually, it was genuinely anonymized. Um, not from my real cases. That would be crazy. Um, [00:59:03] Jacob Haimes: I assumed. [00:59:05] Zainab: um, and, and, and prompted it to, to, you know, find, find bad. [00:59:13] Zainab: This was when we were on like sonnet 3.5, I think. [00:59:16] Jacob Haimes: Okay. [00:59:17] Zainab: Um, so a little, a little while ago it found like one, one very obvious thing. Um, but not, not much else, but the fact that it found one very obvious thing. I was like, okay, there is like, there is something here. And the prompt that I'd given it was like terrible. Um, I was just kind of curious. [00:59:39] Jacob Haimes: Sure. [00:59:40] Zainab: and then kind of got to thinking of like, what's the level of scaffolding that's needed here? Um, started thinking about my workflow and decomposing it into various tasks of like, what do I actually look for first? What are the quickest wins here? [00:59:55] Jacob Haimes: Mm-hmm. [00:59:56] Zainab: and then began kind of playing around with prompting it more specifically. Um, and that gained a ton of value. so it was a lot about like decomposing the workflow, thinking about it as, um, when I was at straws, I was, um, of responsible for training incoming junior analysts, and I was like, let me see if I can about decomposing the tasks as I would there, but like giving it to a model. [01:00:23] Jacob Haimes: Mm-hmm. [01:00:24] Zainab: Um, that helped a ton. Um, I remember I had like three synthetic data set sets at that point. One was like easy, like, think of it as like easy, medium, hard, [01:00:33] Jacob Haimes: Mm-hmm. [01:00:34] Zainab: cases. [01:00:35] Jacob Haimes: Okay. [01:00:36] Zainab: I initially sort of hacked, hacked, hacked together a workflow that was able to kind of solve the, like EAs easy mode task. Um, pretty well. [01:00:46] Jacob Haimes: Okay. [01:00:47] Zainab: this was still like sonet 3.5, but it would miss stuff on like medium and hard mode. and the reason why kind of hard mode was, was, was harder is 'cause there was like, there were more ambiguous indicators, like the legitimate user and the attacker were logging in from the same place. Um, they were kind of using the same session at the same time, rather than having clearly like, demarcated activity and that that can be difficult, you know, for, for like an experienced human analyst to figure out as well. Um, I then got sort of media, media mode working pretty well and that was just, you know, a bunch of like b. [01:01:26] Zainab: The engineering and figuring out how, how, how I can cluster the data together beforehand. Just like a combination of like data engineering and like more, more, um, work on the LLM side. but the model still wasn't working well for hard mode. Then when we upgraded from sonnet 3.5 to 3.7, there was like a step, step increase in capability on my, my then like toy eval. Um, hard mode was immediately kind of solved to a much greater degree than possible. It still didn't find everything, but it was a significant improvement. And that was also a kind of key aha moment. Um, 'cause obviously like the discourse now and the discourse then is like, you know, capability's always increasing. [01:02:20] Zainab: This is like. Seemingly constant inflection point, like lots of things compounding at the same time. Um, but that was the moment where I was like, if you like set up, set up the harnesses in the right way, and build for, uh, a world where you can just like plug in the next model, um, there's like a lot of gain from that as well. Um, so I think like tying back to your initial question of like what are, what were the kind of moments as it were one, the first time I like saw a model, be able to find a bad in like a pretty convoluted log file. two, uh, kind of seeing the step increases as model models became better despite like, uh, not being trained on, on, on this, on this task specifically. Um, and three, um, when we've like got to the point as a company of actually getting to like very, very high accuracy on these still. On any email case, so like still hard email cases, still sort of boxed like email as the scope, but that like gives me sufficient conviction in addition to other stuff that we're doing now as well, that we can expand this to more complex types of incidents. [01:03:36] Jacob Haimes: Gotcha. Okay. Let's see. So there's one, there are a couple things I wanted to to like dig into there. The first thing that I thought of when you were talking about. You know, the Excel spreadsheets and, and all of them sort of like laid out, um, distributed data. It sounds to me like there would probably be a lot of value in being able to aggregate and overlay those, uh, in a, in a meaningful way. [01:04:08] Jacob Haimes: And that may be something that like language models could do relatively easily, uh, and create this sort of more easily, um, investigated artifact that is more comprehensive. And does that, is that something that's like typically done? Is that, uh, something that you're, that you do or, or is it like, oh, we don't even, like, that's not necessary, um, to do at all? [01:04:41] Zainab: Um. We, we do it is [01:04:45] Jacob Haimes: Okay. [01:04:45] Zainab: short answer. Um, and I'm glad you called it out because I think that two parts to the analysis process, one, like seeing everything in one pan and two, just like being able to analyze it well. [01:04:55] Jacob Haimes: Mm-hmm. [01:04:56] Zainab: what, and, and you're right in so much is language models are, are very useful for kind of creating this like master schema and like mapping, mapping things in the right way. was previous. I think that lot of IR teams have [01:05:14] Jacob Haimes: That [01:05:14] Zainab: tried to do a version of that. 'cause [01:05:16] Jacob Haimes: right. [01:05:17] Zainab: I, I used to write scripts, uh, in my old team to just like aggregate stuff as much as possible. So you're, there's not like tab fatigue. Um, for lack, for lack of a better term. Um, and, and we do that too. [01:05:31] Zainab: And it's super useful to just like you get these distributed data sources, almost all of them will have like some timestamp. And almost all of them will have like some kind of action of like login, delete, uh, execute, whatever. Um, and if you can get all of that in one place, um, then you're already able to work so much faster. Um, [01:05:54] What's Asymmetric Tackling Next? --- [01:05:54] Jacob Haimes: Gotcha. I guess I'm not, I'm not just thinking about like getting it all in one place though, but, uh, almost like a hierarchical, um, um, aspect to that. Right. Okay. Okay. When you're thinking about development, when you're thinking about the next thing, what kind of process are you going through? How are you trying to predict, um, what that next thing is, uh, and then engaging with it and like getting the, getting the engineering team to sort of get into it and then go about actually, you know, testing feasibility of an idea. [01:06:38] Zainab: Yeah. Um, I think that by, by next thing, the way, the way I think about it is like, what, what type of case can we bite off next? So we like started with email. Um, right now we're working on kind of insider threat slash. towards ransomware kind of thing. Um, and the way that I, there are kind of two axes, which are useful for figuring out what's helpful to bite off. [01:07:05] Zainab: Next is, how often does this happen? Um, what's, what are the volumes like here? [01:07:15] Jacob Haimes: Mm-hmm. [01:07:15] Zainab: we were to dedicate a bunch of eng time to, uh, work towards a type of incident that happens once every 10 years, um, probably not the best use of time. [01:07:26] Jacob Haimes: Sure. [01:07:27] Zainab: and the second access is like, how complex is it? So like how, how many more data sources do we need to add to our system to be able to bite, bite, bite the next thing off? as, as a bit of an intuition pump here, when the first thing that we moved on to after email, um, was looking at like laptops, as a data source and. The reason for that is that laptops and, and servers are very similar, forensically speaking, um, get hacked all the time. that is probably the bread and butter of most forensic analysts. and it's, they're reasonably contained. you know, everything is like limited to what's on your hard drive. It's not like hundreds of terabytes of like network data flowing in and flowing out. Um, it's just what, what programs were run here, what was installed here, what files are present, et cetera, et cetera. [01:08:27] Zainab: I'm simplifying a little bit, but just to like, give you a sense. Um, so that, that's, that's how I tend to think about it of, you know, what are, what are the bites that we can, we can take off here? Um, high volume are they and how tightly scoped are they? Um, and as we kind of keep doing that, um, does it build to this like super set of like full. Forensic analysis and it can rebuild it in such a way that it's extensible when we see a new data source. [01:08:58] Jacob Haimes: Okay. [01:08:59] Zainab: yeah. [01:09:00] Jacob Haimes: You also talked about building for the next model. Um, how do you do that when aspects of what's being provided are shifting significantly? [01:09:13] Zainab: Yeah, I think it's like there, I don't think there's a silver bullet solution here. Um, I think the key is to like be, be, be agile and like be, be kind of. O open to just trying new things and like avoid falling into the six song cost fallacy of I've spent months perfecting my prompts and now I don't need them anymore kind of vibe. Um, I think it's, I think it is hard, um, to do that, but I think that the, the, the kind of you can get on like levels of abstraction, um, better, the better it's gonna be. So, I spoke a bit about, you know, experimenting in the early days, and a lot of that was like prompt engineering. we do much less of that now relative to, writing quite like intricate skills files for, as, as one example. Um, and that operates at sort of like a higher, higher level of abstraction, I think. And that is like something that is more extensible, um, as new models out. Um. The way I like to think about it is, you know, um, is what we're building now, like actually gonna be useful months down the line. There are obviously like various sub components to that question. Or are we operating in, in such a way that, you know, this work is, this work doesn't end up being redundant. And I think like skills for skills files are like an interesting example here. Like whenever I use code code there's so many random skills files that I work with, um, that end up being super useful. [01:11:00] Zainab: And like in effect what we're doing is these agents with forensics knowledge initially in like prompting, further down the line and like context engineering. Um, and. can obviously to take various modalities, um, and making sure that the dev work that we're doing kind of as flexible as possible and, you know, well, well organized and clean and that kind of thing I think is like super important just to make sure that, [01:11:39] Jacob Haimes: Mm-hmm. [01:11:40] Zainab: that knowledge base can transfer. [01:11:43] Jacob Haimes: Okay. And then you also talked about, um, I, I think what. You target, right? And how you think about what you're targeting on a high level. But I'm also curious about how you go about like the innovation aspect of that. Um, is it, uh, from what you said it sounded like it's, it's mainly, you know, essentially using the same tricks, uh, but applying it to a new domain. [01:12:15] Jacob Haimes: So like, uh, context management, uh, you know, specific, uh, skills or, uh, information around certain types of attacks and that sort of things. Is it, is it at this point, like primarily that, like it, which is essentially an engineering problem? Uh, or is it, are, are there more innovations I guess, that are also happening? [01:12:45] Zainab: I think a significant portion of it is an engineering problem, um, at the stage we're at now. Um, but I think there, there are always two things that we're optimizing for, and with the goal of how can we make these investigations as quick and good as possible? Um, and those two things are one, accuracy, is where the engineering component comes in. 'cause the better your agent is. The more accurate it should be in shooting for an F1 score of one is like al always gonna be the goal there. and second, um, which is sort of more interesting in some ways, sort of hard, elusive, um, from an innovation perspective is how can we get our analysts to QC and understand this data, um, as fast as possible? [01:13:44] Jacob Haimes: Mm, [01:13:44] Zainab: And [01:13:44] Jacob Haimes: okay. [01:13:45] Zainab: there, what that looks like is like, it comes more from like a platform slash like what affordances this page actually include? [01:13:58] Jacob Haimes: Mm-hmm. [01:13:58] Zainab: an analyst is able to move very quickly through this data. Um, like how, how should, how should we be querying if you kind of think, take a step back of like [01:14:08] Jacob Haimes: It's more human factors. [01:14:10] Zainab: pardon. [01:14:11] Jacob Haimes: It's more human factors. [01:14:12] Zainab: Exactly. And you like take a, take a step back of like, everyone is so used to scrolling through like random Excel or like running slightly janky queries. Like how, how would we like redesign this process if we were to think about, you know, actually like QCing a lot of data. Um, and that can have low level implications of like, how, how you actually present the data on screen in the platform. Um, can have sort of, uh, implications to like the affordances of, you know, how are we filtering things, how are we organizing information? Um, how can we make sure that our analysts are able to go from findings to call as quickly as possible and [01:14:57] Jacob Haimes: So it's, it is kind of like streamlining the, uh, the interaction between a highly specialized, uh, sys like LLM based system that has lots of hours put into it and. The expert human and making sure that that, uh, interaction as a is as seamless as possible as well. [01:15:22] Zainab: Yeah. [01:15:22] Jacob Haimes: Okay, cool. No, that's, that's really helpful to understand like where it is you're doing and I guess, or where, where it is you're going and what you're doing. [01:15:33] Glasswing, Dual Use, and Power Concentration --- [01:15:33] Jacob Haimes: then my last like, I guess major question about, uh, asymmetric security, but also bringing in something we talked about earlier, uh, with, with Mythos, um, and Project Glass Wing. Um, I wanna use it as a like, contextualization about conversation, about like dual use information and technologies and consolidation of power. [01:15:55] Jacob Haimes: Uh, because I see this initiative as, uh, particularly concerning, uh, I don't think it's. Uh, unexpected, right? Like, uh, this is, like I said, I don't trust corporations. This is exactly what I would assume would happen, but the fact that the people who are currently in power are establishing that, uh, well, so, so they're saying whether founded or not, uh, that their system is so good at cybersecurity, uh, that they're not gonna share it with you, but their buddies can have it and they can access it, and they can, uh, you know, in theory, if, if we believe it, that would make their technologies and their access superior and, uh, would allow them to reap the benefits of the models before anyone else does, uh, in advance further. [01:16:44] Jacob Haimes: And what that says to me is like, that could be the, that could be a death sentence for asymmetric security if you don't have access to that model and they go all in. Now, I don't think that that's necessarily going to happen. Uh, because again, I'm more skeptical about some of the claims, but if that's what they're trying to establish, like that's kind of concerning. [01:17:12] Jacob Haimes: What, how, how do you think about that? Both from the cybersecurity angle of like, maybe this is good, maybe it's not good, or, or, you know, I don't know. But then also from the, uh, company standpoint of like, how does this impact us? [01:17:29] Zainab: Yeah, I think, I think it's an interesting one. I think the, I think from this, from the cybersecurity angle, these, what, what I think we need for these models is they need to be tested very rigorously. The like capabilities actually need to be understood very rigorously. Um, who does that I care less about then like, that's actually happening. I've seen a bunch of like positive press project last swing. [01:18:00] Zainab: I've seen a bunch of negative press around project lasting. I think people taking mythos seriously is like gen genuine is, is a good thing. and I think people like ma making sure that the model is like, but, but through its spaces is also a good thing. Um, I think. From like a company standpoint, I kind of see people will, in nostalgia space will often talk about, you know, the resources that incumbents have. [01:18:32] Jacob Haimes: Mm-hmm. [01:18:33] Zainab: Um, and that's like usually like manpower, tech, uh, massive teams, Um, and it will always be like a David and Goliath situation, um, with small startup and massive incumbent. And I think access to having early access to models is just like a ex, an extension of like the resources that those companies have available to them. [01:19:10] Jacob Haimes: Mm-hmm. [01:19:11] Zainab: And then I think this kind of feeds into the sort of classic like innovators dilemma. Um. Trope of like wear a speedboat, they're an aircraft carrier, um, which is the kind of organizational thrust that is required for a massive incumbent to dedicate, um, to go all in uh, innovation on like a very specific thing. [01:19:43] Zainab: So like AI enabled incident response, for example, is far greater for, you know, CrowdStrike Palo, um, the, the companies involved in project glass wing than it is for us. [01:19:56] Jacob Haimes: Mm-hmm. [01:19:57] Zainab: Um, if it was like, if the only thing, this, this came down to was like a battle of resources, then no startup would succeed ever simplifying a little bit. Um, but I think it's like part of a meta point of like. If the existential threat to your company is like more resources, then there's something that like you need to think about, um, a bit more, or like to rephrase a little bit. I think it's just, if the, if the, like if the only threat that you're facing is like your incumbent has like, access to more stuff, um, [01:20:41] Jacob Haimes: Hmm. [01:20:41] Zainab: then I think it's like you're not in a terrible position. Um. [01:20:46] Jacob Haimes: Sure, sure. Yeah, that makes sense. So, uh, on, on, like yeah. Mythos and, and Project Glasswing, I guess. So. You mentioned earlier that, you know, it does seem like there were more, um, things, I don't remember exactly how you worded it, so I might, uh, could be off here, but like, based on the, what you've seen, you felt like it was, uh, an increase, right? [01:21:16] Jacob Haimes: But then you also have a bunch of people saying like, oh, well we tried this thing on, um, like with, with similar, uh, not, not large as large models, like, you know, much smaller models that are more available and, you know, we were able to find similar things. So like is it really that [01:21:37] Jacob Haimes: you found or that like there's a, a step change here? Or is it that there is a recognition of, of a capability being there all along? [01:21:49] Zainab: I think, um, I think there is a step change in terms of, you know, how this, based on the kind of stuff that I've, I've read on what, what attacks was the, what attacks were the previous. Generation of models able to do versus mythos. Um, and there is a significant increase. I think the, the AC paper that I looked at was like capable of like full, full network intrusion, and that's super hard and there are like various steps that are involved in that and you know, there's like a bunch of stuff. so I think there is, I think, I think it's important not to underplay kind of, uh, change here. Um, I also do think there's an element of this, of like, when something is discovered once, um, for, for the first time, it is easy. It is easier to like replay it. Um, I'm not saying that everyone just, you know, copied exactly what had happened or, you know, looked at the, the exact chain of vulnerabilities that led to, mythos being able to find like some 27-year-old Linux vulnerability and, and kind of copied that step by step and got to the same answer. Um. I don't think, I don't think it was that. but I think, I think it was more a step change than just like eliciting capabilities. And I think a potential hypothesis for that is that a bunch of work on the models recently has had this like jagged pointedness towards coding ability. Um, and there is this like close interplay between how well one can code and how well one can write secure code. I don't think it's surprising that as we see like coding models get better and better. Like I've been coding with code, code for, for a little while and if, you know, see, see it get better and better, um, over time. Um, and I think that is the grounding that I have of, you know, this like. Everyone is saying this is a step change, but it makes sense why this is like probably a step change. Um, because the capabilities that have been prioritized to like a meaningful degree are similar operate in a similar arena. [01:24:21] The Relationship Between AI Safety & Cybersecurity --- [01:24:21] Jacob Haimes: Okay. And then we're almost, we're almost at the end of it. I, I do wanna ask just a couple questions more about, uh, AI safety and cybersecurity and like, what the, the differences are there and, and how you see that relationship progressing, uh, as well as maybe a couple lightning round questions. Um, so. It seems to me that like AI safety and cybersecurity are beginning to converge in, in a way. [01:24:51] Jacob Haimes: I mean, obviously there are aspects of both that are, that are not, uh, related to each other, but there are cont continuingly overlapping increasingly overlapping, uh, aspects, but these communities largely remain separated and distinct. Why do you think that is? [01:25:14] Zainab: I think, I do think, I think they're coning more, but I agree that there is like a distinction and like, part of the reason why I was so excited to build in this space that I'm building now is because it was like, there was very small, small like Venn diagram intersection there. why do I think that's the case? Um, I think I, I think it's just sort of like industry research disconnect that exists across sort of, um, various domains of, like historically AI as a field, as a field has been, dominated by like academic side of things. and cyber is like literally the opposite, um, of like, you know, a lot of the best practitioners, um, didn't, didn't go to college. [01:26:02] Zainab: Um, it's like, it's a skill, it's a skills based thing of like, you learn, you learn how to do this thing, you like, develop knowledge. Like are people who, work for, you know, the, the like, security as a field isn't, uh, kind of the, the glamorous PhD that you do. I'm not saying that that doesn't exist, but just like to, to sort of paint a bit of a picture there. Um, that's like, as somebody who sort of existed in both, that's like the difference that I see most, most strongly. Um, but I am like very optimistic, optimistic about the, the kind of convergence, um, and. is super hot in AI right now. Um, um, as, as we all know, and I'm like hopeful that part of riding this wave is like bringing those two disciplines close together. Um, I think the way that they've existed together in the past is like bolting on like ML [01:27:10] Jacob Haimes: Mm-hmm. [01:27:11] Zainab: security systems, and then more recently, um, more advanced lms, but like very much like bolting them on. so yeah, I think there's like a ton of exciting work that's happening at the, like, in, in intersection now. Um, and I'm super hopeful about like how that's progressing. [01:27:33] Jacob Haimes: Okay. And then sort of maybe in contrast to the hopefulness currently, what do you feel AI people misunderstand about cyber? And then also the flip side, what do cyber people misunderstand about ai? [01:27:51] Zainab: I think, um, this kind of ties into what we were talking about earlier of like so many of the attacks that we see are pretty un unglamorous and run of the mill. and possibly there is this narrative in the AI community of like, gonna be a ton of like polymorphic malware that is like LM based and changes based on static detection patterns that like don't exist yet. [01:28:19] Zainab: I'm using a very extreme example. Um, whereas like in reality, the, the bulk of the cases that, um, a lot of, uh, DFIR providers see are just like really run of the mill and super kind of but like obviously have significant impact. Um, so a lot of breaches on these like cool novel attacks. They're like. Pretty kind of the, the boomer, the boomer version, um, for lack of a better term. Um, and then I think on the flip side, um, thinking about how the cybersecurity world may misunderstand, uh, AI is, uh, the way security has like developed is that there's, people see new tech as like an increase to the attack surface. [01:29:19] Jacob Haimes: Okay. [01:29:20] Zainab: this happened with like cloud, um, a couple decades ago. this happens when like. There is like a new, new VPN vendor that gains a bunch of traction or, you know, just like whenever there's like new tech, um, the, the attack surface increases and you kind of need to think about vulnerabilities from a new angle. Um, possibly, um, people can think of AI in the same way of like, oh my gosh, now we need to worry about, um, you know, uh, just risk risks in terms of like LMS being misused. It's not just about like SQL database injections. We also need to think about prompt injections, um, and like that kind of thing. And I think the, the misunderstanding there is like with, with ai, it's not just like you have one more door to lock up. Um, it's, this is like tech that can change the various structure of how, uh, security systems operate. [01:30:26] Jacob Haimes: Okay. So it's opening a bunch of doors at the same time. [01:30:29] Zainab: Yeah. [01:30:32] Jacob Haimes: Okay. Yeah, that makes sense. Uh, uh, that's in line with my thought of like, oh, we should probably be specifying these, these technologies more anyways. Right. So, um, if we're introducing more narrow technologies, that wouldn't be as much of a problem, but they're intentionally, uh, over defined, I guess, or, or undefined to be Yeah, under defined. [01:30:55] Zainab: And I think it's, it, it's more than a bunch of doors. It's like a, it's a bunch of doors, but also this can fundamentally change how we do security of like, I think incident response is fundamentally better when, uh, you get kind of cutting edge AI involved. Um, [01:31:13] Jacob Haimes: Okay. [01:31:13] Zainab: I think it's like, it's like a, it's like a power, like a power enabler rather than just like, oh my gosh, this is like another, like window. Two attackers being able to get in. Um, there's like a sim a a similar paradigm around cloud of, like, CrowdStrike as an example, became such a big key player because they kind of redesigned how detection and response was done when like tech became a thing. [01:31:43] Jacob Haimes: Mm-hmm. [01:31:45] Zainab: and I think we're at a similar moment with ai, and it shouldn't be treated as like, this is like an additional vector that, that we need to think about. [01:31:54] Zainab: And like NIST has written a guidance about, it's so like job done. It's actually like attacks are changing massively, but like can change massively too. [01:32:05] Jacob Haimes: Gotcha. Um, okay, last questions. What about your job is something that like really annoys you, really grinds your gears, you don't want to be doing it. What do you not like? [01:32:21] Zainab: Um, what do I not like? Damn. I truly love my job, but there is, is definitely stuff that's annoying. Um, let me think about this one. I think, I think when I, when I, this is a double-edged thing. Um, I, I will only get called when somebody is mid hack. It is probably one of the worst days of their life. Um, I enjoy the fact that I can make that a bit easier. I don't like the fact that that can, sometimes bring out the, like, ies of personalities in people. [01:33:05] Jacob Haimes: I see that. That makes a lot of sense. Yeah. Okay. And then on the flip side, again, what, what's your favorite part about what you do? [01:33:19] Zainab: Um, I think, yeah, I like, there's, so there, this is way easier to answer than what, what do I hate about my job? Um, I think there are, there are a few things that stick out. One is, so fun to be able to, if like, reinvent this field, like we have [01:33:41] Jacob Haimes: Hmm. [01:33:42] Zainab: amazing tech that's like available to us and it's about harnessing it in the right way and the possible impact of that is massive. I think that's, why I wake up in the morning. That's why I'm all in. Um, like what, what makes me get out of bed and like genuinely excited to like go to the office and work. I think there are also smaller moments that are incredible and like super rewarding. When I, um, I think we, we closed out a case on Wednesday, um, and the analyst working on it literally did it in an hour and she kind of looked at me and was like, I've never done a case that fast. And that is an exceptional feeling of, you know, it ties into change changing how work is done, but like seeing it on a personal level of like, this is crazy. Um, this was unbelievably fast. Like the client called us, uh, you know, a few hours ago, um, and then got us access very quickly and we're done. We can call them, um, very quickly and say that you guys are good. Um, don't worry. You've taken the steps you've needed and then they don't need to worry about it. Um, because a stressful part of this on the client side can just be like. Wait, waiting for results, you know? [01:34:57] Jacob Haimes: Mm-hmm. [01:34:59] Zainab: so yeah, I think those are like two that stick out. [01:35:02] Jacob Haimes: And then what, what are the tips that you would give people? Like the everyday people, the, the how, how do you protect yourself, uh, in, in this, uh, era? Mm-hmm. [01:35:15] Zainab: Yeah, I think that like this easy stuff that everyone should do, use a password manager, never use the same would for more than one thing. I tell all my friends to do this as well. Um, um, put MFA multifactor on everything. Um, if you're super concerned about something, get one of those like fingerprint keys that does like, add a bit of like time and effort and like, possibly a little bit annoying, but like if you're super concerned about something that like have a physical key tied to it. the other thing that I'll say is like in this era of like. Not just AI enabled attacks, but like, there's like a new AI app popping up every two seconds and everyone will always like rush to try it out. Um, there is like a bunch of like tech sprawl connected to people's, um, So like, do an audit of the applications that you've like connected to your Google account. [01:36:14] Zainab: If you're not using one regularly, just disconnected it. I think we're gonna see an ops surge in like supply chain attacks and the less stuff that you're connected to, better. Um, the other thing is like keep your laptop aggressively updated. The moment you see a software update, just like get it going. And the reason for that is like, I think the time between, uh, a vulnerability coming out, um, like being exploited is going to narrow. Um, and if you would just like. Make sure that stuff is patched and your specific device is like vulnerable for, um, as little a period as possible. That would help too. Um, I think those are my, my initial, my initial thoughts. I, [01:37:08] Outro --- [01:37:08] Jacob Haimes: Well, Zab, thank you so much for, for joining me. I have been looking forward to having you on for a while. I know we talked, we talked about it like a, a while ago, um, initially, and, I've been looking forward to it since. So I, I'm really happy you were able to join me. [01:37:25] Zainab: me too. Always a pleasure to talk to you, Jacob. [01:37:27]