Jackson: And they have the compute budget for ten thousand actions a day. I think a lot of the nightmare scenario everyone's talking about is, my goodness, they throw ten thousand actions at me in one day, and this agent is so fast, or the swarm of agents overwhelms me. What I'm much more concerned about is they say, Okay, I have ten thousand actions a day, I'm gonna take ten actions across a thousand different environments, and my year over year ROI is gonna be greater overall as I achieve bigger impacts.
Aaron: the idea of scale and patience is terrifying
Jackson: Yeah. Yeah.
Aaron: because offensive defense, all hackers are somewhat lazy, right? In the sense like we love to automate things, we love to like it's just drives our whole industry, it always has, right? And really, really smart people given the superpower of patience terrifies me.
Jackson: people are shoving way too much intelligence at problems because it is easy. I I think probably everyone has had the case where they're like lazily trying to solve a problem and they're using like sonnet and it's not working and they're like, alright, fable, you go do it. I'm not gonna pay enough attention to go do this, like you solve this problem for me. And it's not a fable level problem. It's a sonnet and like one percent of your brain kind of problem.
Aaron: Hey everybody. welcome to the ZeroSum podcast. I'm your host, Aaron Mog. so great of you to spend some time with us today, listening and watching us. really excited to have Jackson Reed on the show. Jackson's a a new friend, somebody that I'm getting to know and learn about in when I heard what he was up to, was really excited to invite him to be on on the show. I get a lot of feedback about what we're talking about and AI and the future of AI, the future industry is is certainly an interesting topic. So, first of all, Jackson, thanks for being here.
Jackson: Yeah, of course. Yeah,
Aaron: Really appreciate it, man.
Jackson: glad to be here.
Aaron: So for those of our listeners that may not know you or know your background, maybe you could give us a little bit about your perspective and where you're approaching this from.
Jackson: Yeah, yeah. So I guess I'll I'll start near the very beginning with not being a technical guy. I did not think I was gonna go into technology, did not think I was gonna go into cyber. yeah, I was gonna
Aaron: Really. Interesting.
Jackson: go into diplomacy, I was gonna work for the State Department, do national security kind of things. That was the plan. that's what I went to school for. But then second semester of freshman year I took a course and it was all about space and international affairs. And I'd never thought of things like radio frequency is a scarce natural resource. You can't like carve it up into infinitely fine little bands. And so states of conflict, not just over like, this is my land, but also this is my band of of spectrum and I don't want you to use it. I never thought of that before. I never thought like wow, there are all
Aaron: Yeah.
Jackson: these all these negotiations and conflicts that are happening in a sector of the world I didn't think about. the professor was talking a lot about like how law was being made and policies were being decided. It's like a bunch of people in a room making a choice. They're like maybe this is the right call, I don't know, we'll find out. I thought that really interesting. Wanted to do that for cyber because I thought we're kind of at that stage now where people are making these five by the seat of their pants choices that are impacting law and society. Yeah, never. Yeah,
Aaron: Us in Cyber? Never. That's never happened.
Jackson: so I w I wanted to like do the policy work there, but I wanted to be technical enough to actually know what I was talking about. So I started getting more technical and then it just like sucked me in. So finished up undergrad with International Affairs, did grad school in computer science through I guess I'll give a plug for scholarship or service or cybercore if anyone has kids. Or if anyone is young enough to be thinking about this, it's an incredibly awesome program. Think Rotzi for Nerds. So that paid for most of undergrad and grad school. And then I did work, so it did think tank work, did some sock analyst work, discovered that I didn't love that. but it was a great great learning experience and good folks to work with. And then went over to CrowdStrike to do threat attribution and intel analysis work. that's where like I gotta bring the international affairs stuff and do China stuff. and then went over to the US Army Corps of Engineers to build and lead their red team. So everyone probably
Aaron: very cool.
Jackson: on this on this knows red teaming, but Department of then Defense, now Department of War, has like a particular process for becoming a certified and accredited red team. So did that, took the team through CNA. that was exciting, did that for about four years. Really fun to operate on like really interesting, cool systems that you just don't get to in normal life.
Aaron: Yeah.
Jackson: and also just a really, really fantastic team. So that's kind of where I learned CyberOps stuff, had a great time there. Went over to Target to rebuild and lead their red team. Did that for about four years and then in March launched Barding Defense. So we can talk about that to whatever extent we want, but also I guess I should say sort of overlap with a full career stuff. About three years ago started working on AI, saying, okay, I can no longer just sort of like watch this with mild interest from the sidelines. I need to actually get involved. So started doing, you know, sort of the babies first model, the hello world versions of, you know, I'm gonna train my own little classifier, things like that. and then started getting into mechanistic interpretability. with you know looking at like how does model actually produce the outcomes and can I at that time I was saying how can I break models and how can I use mechanistic interpretability to understand like better this geo break is going to work and I know it's gonna work because I like watched other prompts function. So that was exciting did that for a bit. and then started working with a group called BT6 doing assessments and geo breaking for Frontier Labs. And then sort of from all of that is what Barting comes from, the combination of social impacts, cyber, then AI. Does that give the is that like a covered review?
Aaron: Yeah, yeah. I was like BT six because it's like the most hardcore group of AI hackers with the most boy band name of all time. it's like we're
Jackson: Yeah.
Aaron: gonna hack your frontier model, then we're gonna go on stage and perform.
Jackson: Yeah.
Aaron: but that is a a group that made up of some some world class AI hackers and thinkers.
Jackson: That's pretty cool.
Aaron: so that's a lot of depth and a lot of understanding, I think, of AI and It's funny that somebody who's been doing it for three years has as much depth as anybody because it's all so new. but I I I think it's it's something people are are tremendously interested in. But let me take you know, take it a step back because obviously being a founder, you know, being involved in sort of the cutting edge of AI, one of the things that we like to talk about on this podcast is our industry and kind of where we're at and how we feel about it. And I'd love to maybe start there of like where like You obviously are diving in the commercial side
Jackson: Okay.
Aaron: of this of starting a company and you and you obviously have been doing cutting edge AI work.
Jackson: Yeah.
Aaron: Are you hopeful, upset? Also, just a level set for anybody listening, depending on your listening, we are it's Friday, September 18th, 2026. The hype day in the history of AI so far. The overwhelming amount of bullshit coming out of the labs and out of the news. Is it's just unbelievable. So we are also in this like panical era of just AI craziness. But like what do you think? Where are you at?
Jackson: Yeah. I sort of I feel like I have all of the emotions, in like different sectors depending on on the exact topic that's being discussed. I think probably my overall view is just like you said, there's an insane amount of buzz and hype and there's like a reality distortion effect around AI. There are people who are like very competent, very experienced people that will make statements to be like that is objectively and demonstrably false. Not like I think it's false, like Here's the data show that that is not an that's not an accurate claim. so it is a very weird field. There's a lot of strange statements. but I think we are in for a pretty bad time. I am not personally sold on the you know, swarms of self ref self replicating agents will take over the internet in some like utterly unstoppable way. I do think loss of control is a real risk that we should be talking about. I think that most of what we're labeling loss of control is like not actually loss of control and solvable through like fairly easy approaches. But yeah, I I think I think we are in for a bad time. I can talk a little bit more about what I think what I think is gonna drive that. There's yeah, I sh sh short short version on it is I think we're so so early in
Aaron: Yeah.
Jackson: this. you can look at right, history history is gonna come into play here. You can look at like every time there's new technologies people like really get fixated on, wow, this new technology is really transformative, which it is. and then they say, and therefore and the next step they make is wrong. and I think we're in that phase right now where we're saying, Wow, this new technology has come out. It's really transformative, that's all correct. And therefore And then that's where like all sorts of insanity occurs. I don't know if that's a a complete answer. Short version is like I feel bad, I feel good about some things. I think most of the most of the harms that we're talking about that we're really concerned about right now, I think are harms that are that mostly are real and mostly we're gonna figure out how to solve, like not too far away from now. the things that concern me I think are are harms that are gonna be harder to figure out solutions to. But that's probably the
Aaron: Yeah.
Jackson: My version of a short version of that.
Aaron: It's interesting that you use an analogy of weapons versus an analogy of discovery, right? Because the very interesting. I think it says a lot about where you're coming from. But like I also, you know, I think about it a lot like atomic energy, right? Where it's like so many good things can come from it and so many awful things can come from it. And
Jackson: Yeah.
Aaron: it's like it's it's all in our hands on how we use it, right? And but at the same time,
Jackson: Okay.
Aaron: you know, it's I saw something the other day where they were sending in cereal boxes, they gave kids little rings with little like radiation metal in it to
Jackson: Uh-huh.
Aaron: like hype the atomic era that like poisoned kids. And I'm like, that kind of feels
Jackson: Yeah.
Aaron: like AI right now. Like you're so hypey, you're so in a hurry to market it before you know what it is. And that's
Jackson: Yeah.
Aaron: kind of how I think about it. Like we don't know what this is yet, but we're we can't wait to package it and sell it and monetize it and like IPO it. Like, are we
Jackson: Yeah.
Aaron: really ready for that? Like, I don't like again, more business side.
Jackson: Just set it.
Aaron: But what you say you think it's gonna get bad or or and I've had like I had Casey Ellis on the pod, and Casey knows a lot about vulnerabilities. He's like, it's gonna be a disaster for a little bit, and then and then we'll figure it out. I think it's like a short term pain followed by long term winning kind of mentality, which sounds like maybe you have a little bit of that. what do you think is gonna get bad? Like the threat actors scaling, or like what what part of it makes you the most nervous?
Jackson: Yeah, so i in the short term, I think like, sure, there's gonna be a lot of like vulnerability discovery, there's gonna be a lot of hey it turns out you can throw a lot of agents at a thing very quickly, they'll find a bunch of attack paths in. and I think like there is going to be a band of time where these are problems. But I think that first off, defenders love seeing bursty anomalous behavior. Unless it's so bursty it overwhelms them, but these are just sort of like software problems that we're gonna resolve. Bursty anomalous behavior. really generally quite nice to handle. you would rather have that than slow normal behavior. Yeah.
Aaron: Slow quiet is the scariest. Yeah.
Jackson: So I think we are gonna be in for a period where like the burst of the anomalous behavior is like new, it's unsettling, and it takes our big industry a while to react. So I think there's gonna be that that happens. And that's I think like the initial bad phase and I think that's where a lot of the energy is being focused. The part that concerns me is what happens when attackers figure out, hey, I'm actually gonna create really effective purpose-built harnesses for operations. Because all the benchmarks, like operations suck for a benchmark. Because operations are not very like science friendly. They're not very reward function friendly. It's very reward
Aaron: Mm-hmm.
Jackson: function friendly to say, did you fuzz the binary and did it crash? Binary yes or no. Okay, now that you've done that, let's continue down that path until we have an exploit. It's very nice for benchmarks. Very nice for a lot of the Frontier labs are not cybersecurity people. The overwhelming majority of them are not. And so from like an academic shape or from like a I have my machine learning PhD shape, it's very nice to say, wow, look at all this thing we're doing in cyber. And then the Frontier Labs are already like problematically biased to believe that AI will solve all problems will be smarter than all people. So they
Aaron: Right. Sure.
Jackson: go to like, wow, this is this is the future. But I think also, you know, what concerns me is when attackers get harnesses that allow them to run operations not just where they're throwing 10,000 agents at a problem. And they find all the attack paths that one of them works. Where they're able to say, okay, our harness is now going to allow us to pick just the right path. We're gonna figure that out of time or we're gonna pursue it. So I shall summon Rob Joyce. So he ran offensive and defensive ops for NSA. And he said in public, he's written publicly about you know, what makes NSA awesome at doing what they do is not, as people often believe, the fact that they just have really great zero days and really great tools. What makes them great is the fact that they take the time and they have the ability. To deeply understand an environment they're operating in better than the people who designed it or are managing and
Aaron: Mm-hmm.
Jackson: maintaining at that time. And that's what is concerning me. he has another quote like a Usenet's paper where he says, like patience and persistence will get you places that a zero day never will. Harnesses can do things when it comes to like patience. It is like sucky for a human operator to sit on a machine and like dig through a bunch of files slowly.
Aaron: Right.
Jackson: agents are really great at that. and agents like it doesn't matter if an agent is picking up a task a month later or a second later, the context is just the context for them. Whereas for a human operator, I can't manage like ten different operations at once, where one is like three months old, I don't recall really what happened there. So Yeah. The the aspect of agents being able to move slowly, methodically, effectively. And also what happens to your log retention periods. If if you have an agent, you you have some sort of attacker, some criminal, they're trying to achieve some sort of nefarious aim.
Aaron: Right.
Jackson: And then the impact that has for us in terms of log retention periods, when an attack unfolds over six months, as opposed to like six minutes is really nasty. what it does to like the compute costs and expenses of managing a sim where have this huge storage requirement where you have to do correlative alerting not over you know a period of four hours but a period over four weeks. This is horrible stuff. where an attacker gets
Aaron: Yeah.
Jackson: on your system, just monitors it for a week and collects the baseline data around like what normal looks like and then starts operating within normal bounds. This is the part that I find is a lot more concerning to me than just like Bursty swarms of agents, voans every day.
Aaron: Yeah, it's it's interesting because
Jackson: It's a.
Aaron: I I agree. I think the the things that I'm the most nervous about, right, is and I'll kind of rephrase what you're saying, but like Like it really like that that is scary. And the other is sort of there's always been this issue of asymmetry with threat actors and and defenders, but now you're getting into things like model asymmetry. Like they have models without guardrails, we have models with guardrails. Like, how is that going to turn out? Like, there's
Jackson: Yeah.
Aaron: a lot of questions that I have. But you know, there's also and I I'm interested in this because I I know you've gone deep on some of this. And I've had people on the sh on the show, I just did I had Silas Cutler on who does threat research. It's like
Jackson: Yeah,
Aaron: AI doesn't necessarily create more hackers either, right? So, you know, I hear a lot of hypey shit in cyber about like 10xing engineers, 100xing engineers. It's like, you know, the same tech, 100x is hackers, right? Like we're
Jackson: Yeah.
Aaron: not really talking about that, but like it's it's terrifying. So you chose in your career to focus on testing,
Jackson: Cool.
Aaron: red teaming, things like that. Is that where you think the answer is? Like, is to understand. behaviors better and defend them? Or like how do you where do we go? I guess how do we solve this?
Jackson: I think that there are two when we're talking about capabilities, there are two failure modes for red teams. And one mode is like the the fully reactive emulator red team that says, Okay, a bad guy did thing X, we're gonna go do thing X as well. Defense has to be by our nature reactive. if you have no attackers, you have no defenders. So defense to some extent is always responding to the like approaches and the incentives that criminals have. We also get to shape them, which is exciting and we can touch on that as well. But you when a red team falls into that category as well, and a red team is like responding to what defenders who already are in a reactive role are seeing attackers do. I think that's a a state where like red teams are still providing value or have historically provided value, but they're way below what they actually could be supporting an organization in. The other failure mode Is when like red teams realize, hey, I shouldn't just be like following in the footsteps of attackers and do the thing like we already know they're doing. If we already know they're doing this thing, like that's what Baz is for. but then when red teams go the other way and they're like, well I'm be like the super secret squirrel, like most hackerman, most James Bond red team in the world. And then they like are proactive, but in ways that like red team or in ways that criminals won't be. you're
Aaron: Right.
Jackson: not paying attention to the actual incentives and requirements of criminal groups. So I think with all of that. I really am quite bullish on red teams. Cause I think proactive security testing where you're saying, Hey, this is where the puck is going and this is where we see like the incentive is aligning, the technology is aligning for criminals to be going in this direction. Now we need to start doing that, I think is extremely powerful as a way for organizations to see and understand risk and make decisions. Because when you hear like, hey, there's a risk on paper that goes straight into the trash heap of things that I can't do today and I have too many of them anyway. where you like see it and get to feel it, which red teams can do, I think that changes things. and for this, I'm actually gonna go back to CryptoLocker. I'll see if I can recall my recall my dates here. But the first ransomware event was in like the nineties. It was the AIDS virus and it was this like extremely feeble and poorly done piece of malware, it like floppy discs shipped to medical offices and you had to like mail physical dollars or like you know money orders to Panama. And that was our first ransomware and it sucked.
Aaron: Jesus. Sure.
Jackson: and then there was I think the paper was cryptovirology maybe. but like two two researchers, Young and Young, put out a paper where they're just sort like screaming into the void of like, hey ransomware's gonna be a problem. The fact that you can do this is a really, really significant concern. And it was years ahead of anything serious happening. And there were like a couple really crappy attempts at ransomware. There was sort of like period of silence. There were attempts at ransomware it was like like the key was in memory all the time, so you could like very easily unlock it. But then there were three
Aaron: Right, just reverse it, yeah.
Jackson: things that converged at once. You had CryptoLocker, which actually had like you know, reasonably professional crypto that it did for the first time, and none of the prior instances had that. You had like a reasonably good distribution network with Game Over Zeus, I think was I think anyway, whatever CryptoLocker used for distribution was like reasonably good. and then you also had like a reasonable way of cashing out. And that was the first time that all three of those things converged together and all of a sudden things just like caught fire. And then for years, ransomware was the hot stuff because everyone saw, wow, that really, really worked. And I think that we're in that space now with it with like AI and cyber offense. So we're just kind waiting for the right clever nefarious person to put together all those like factors that flow together and create the first crypto locker. I think there's a lot of discussion like, hey, AI and cyber could do this thing. Sure, there are lots of things it can do that it's not going to really be adopted by criminals for. There are some that it will. but I think like the the more that red teams can do and the more that like proactive security folks can be doing to say what really is gonna be the shape of the factors that sort of like flow together and what's gonna be the first real sort of crypto locker estimate. Crypto locker was technically possible for a long time before it actually happened. But then like some enterprising criminal did it. It caught on. They made a huge amount of money in a very short amount of time. nothing collects criminal attention then more than, hey, this thing worked really well. Yeah. So they jumped on that train.
Aaron: Formal? Yeah. Yeah.
Jackson: Anyway, I I don't know if that answers your question on security testing, but
Aaron: No, it I remember when blockchain, I mean Bitcoin, but just in general being the scariest thing to me in security ever, where I was like,
Jackson: Yeah.
Aaron: god, they have a way to get this money out, right? Like, you know, it was never a very easy way, like to your point about mailing money to panel. Like ACH sucks, like it's you're getting into the Fed shit, like it gets crazy.
Jackson: Yeah, yeah, yeah.
Aaron: Like just wire me a hundred thousand dollars of Bitcoin is so much easier. It never really happened though, in in my perspective. Like it didn't crush our economy like a lot of people were scared of like, God, every small business in America is gonna crumble because they're all gonna have crypto every other freaking week and have to pay bounties and ransom. That being said, you and I know on the actor side, like that's so much hassle. Like what a nightmare to manage that and all the C2 and all like tracking who's where and and like it just but like AI makes that possible. Right. So like large cyber ops become a little bit more manageable. Again, like maybe like I'm I'm just guessing.
Jackson: Well I I think there's an aspect From a criminal perspective it's I think under discussed, which is right now For for defenders, like cooperation can be a pain just from like general business politics. But it's generally like a fundamentally safe thing to do. But for criminals, cooperation is not generally a fundamentally safe thing to do. and I think one of the areas where like whatever whatever NX, whether it's like 1x or 10x or like 0.1x you think AI improves a human, you know, whatever that value of n is. I think criminals from an organizational standpoint have a lot to gain from the use of AI because it allows them to reduce their trust on untrustworthy partners. And so I think that there's some aspects there where, you know, for me, one of the things I'm concerned about is longer horizons, slower intrusions. Right now, criminals have incentives to like do a quick cash out, because if I don't do it, then like my my nasty competitor who I don't trust is going to steal the money first and I won't get any payout. But I think once you're once your competitor is like an agent that you're working with, then you like I I don't necessarily trust agents when my option is like a real trustworthy human being, but if my other option is like an actual criminal who actively wants to harm me at this time and will screw me over if he can, maybe an agent's like a much better approach for me.
Aaron: Yeah, I mean, even things like accounting and dealing with like business stuff, like every time you open that circle of trust to other parties to help you do your defairs act, like to your point, you're raising risk. I I haven't
Jackson: Yeah.
Aaron: really thought about this that way because like I use this in business, right? I'm sure you do too. Like, you know, an agent that's helping you with tax advisement or agent that's helping you with expenses or whatever. Like those are people I don't have to pay, which isn't a risk thing for me as much as just a money thing.
Jackson: Yeah.
Aaron: But the the ability for a threat actor to build essentially a corporation of support around them is fascinating to me. Like to that, I I haven't really thought about that. So
Jackson: Yeah, and I think like
Aaron: you think there'll like ransomware as a service and then like criminal enterprise as a service with like agentic solutions to like
Jackson: Yeah, it's
Aaron: CRM? Like, do hackers have CRM? And if so, do they are they gonna use AI to manage it? Like those are all super fascinating questions.
Jackson: Yeah. Yeah. I I think also like lower lower trust. You know, it's not just I don't have to pay this person, but like I don't have to trust this person to not screw me over, which I think is is a a much bigger problem
Aaron: Super interesting.
Jackson: for them than it is for for me or for you.
Aaron: Yeah, yeah, for sure. So red teaming. And I'm super fascinated by this. I actually I mentioned earlier I had somebody on a another pod that I do talking about red teaming and you know how we're using AI. And like obviously companies are scared of having AI do the full thing,
Jackson: Yeah. Yeah.
Aaron: which makes sense. So it seems to be I would say starting at the beginning and the end, right? In the sense of like AI is, I think, a really good place to think about opportunistic discovery and scaling up attack surface management, like the shit where you're not hurting anybody, you're just gathering
Jackson: Yeah.
Aaron: a lot of data, looking for patterns. Like that's really AI. And then reporting, right? Which is like nobody likes to write the PDF at the end of an engagement. So like having AI do that. But in the middle of it, like modern security ops, right, are are are built a certain way. Like, so this is something fundamentally I think a lot about. Like we're taking AI and we're applying it to yesterday's security paradigm, right? Like you think about like stock. Like logs. Logs are stupid to your point. Like, shouldn't there just be an AI sitting there doing the job? Like, like this is 1980s ways of thinking about AI problems. So I I think about the same thing with hackers. It's like, I'll give you an example, like from my days of red teaming, right? The draw out versus the go-in, right? Of we would get paid by a a bank or somebody to do a pen test, and we would immediately start scanning their attack surface. Maybe we'd send some spear phishing emails and we'd do like basic shit. But like it was all about them. And
Jackson: Yeah, yeah.
Aaron: then when we wanted to, we would create infrastructure that looked like them or look like their SSL VPN or whatever. And that always worked way better, but it's such a pain in the ass, right?
Jackson: Yeah.
Aaron: To make it look real and by the domains and all this. But like AI. We're already in this hall of mirrors in society, right? Where like, people don't know what's real on the internet. There's
Jackson: Yeah.
Aaron: a a a great comedian bit is like the internet's just not for you anymore. Like you don't this is too confusing.
Jackson: Right.
Aaron: But like what what happens in red teaming and corporate America where with a push of a button I can exactly copy your infrastructure, like at a very expert level. That's
Jackson: Yeah. Yeah.
Aaron: terrifying, right? Like, how are you thinking about that? Like When all of a sudden employees don't know what's real and they're being drawn into like weird systems. And like, how do we deal with this? How do build a trust in this new world?
Jackson: I think I think a huge amount of it like comes down to fundamentals and I realize that in the face of massive technological change, being like, hey, do the do the basic things right is neither satisfying
Aaron: Right.
Jackson: nor all that like comforting. but I think that a lot of it is the right answer because when it comes down to like how do I know what's true? Well like have an asset inventory to start with. That's probably like a really like this is this is not this
Aaron: We've been saying that for thirty years. Yeah. Right.
Jackson: is not like You you want to hear something exciting and you like, like use AI to counter the attackers and neutralize their threats, but like really it comes down to like, yeah, I mean, we got away with a lot of stuff because attackers couldn't do speed and scale and sophistication and stealth at the same time. And so you had like the people who were building custom infrastructure to like track what you're doing and be really sneaky. and that wasn't you couldn't do it like really, really well. And be really stealthy once you got into a network and be like fully autonomous and fully fast and really scaled. So we got to sort of live on this trade-off where attackers had to pick one. And so we got to be a little bit lazier and sloppy around how we did some stuff. And we optimized for certain things and our optimizations forced them into some of the decisions that the attackers made as well. But I think, yeah, it comes up a lot of the basics, like have an asset inventory. manage your systems, do identity, don't have like five thousand agents for every human being and none of them have an identity other than like Bob's agent one through five thousand. Like that's a bad way to do it. so I think that really is what it's gonna come down to is like we are suddenly dealing with the agony of all the things we like kind of knew we should do. It's sort of like we go to the doctor and like your cholesterol is too high and you're like, I knew I shouldn't be eating so much bacon, but like now I have to and I It's that sort of thing Yeah yeah, like now we know.
Aaron: Right. Yeah. You got the warning. Yeah. Yeah.
Jackson: and we have to do it. So I I think that's it. It's not like a super exciting answer, but
Aaron: Yeah, it's funny, man, because like I have found through this and my my other stuff that I do when talking to experts, I have found the more you're an expert at your field, the more that's the answer. Right.
Jackson: Mm.
Aaron: Of like not a hand wavy AI answer, but like, dude, just like do fundamentals. Like I
Jackson: Yeah, yeah, yeah.
Aaron: had a a guy on the show who helps run software security at GitHub, right? Which is at the epicenter of AI chaos.
Jackson: Hmm, cool. Yeah, yeah, yeah.
Aaron: And I'm like, what the hell are we gonna do with all this code? Like, it's too much code. Like, it's millions of lines of code every day that, like, you know, I I was I was using an analogy of like video, right? So, like every day we put more video on YouTube than the entire world created in the 90s, or some like analogy like that. We're doing
Jackson: Yeah. Yeah.
Aaron: that same thing with code now, where it's like more code written this week than was written in all the 2000s based like the
Jackson: Yeah, yeah.
Aaron: early 2000s. That's terrifying. And and it's like He had to say he's like his fundamentals. It's like supply chain security, reusable
Jackson: Yeah.
Aaron: code, reliable pathways, like all these boring ass things that we've been talking to developers about for
Jackson: Yeah.
Aaron: 20 years. But like we also have to AI it, right? We also have to scale up with it. And I
Jackson: Yeah.
Aaron: I fear an example of this is red teaming. Red teaming's never really worked because it doesn't scale. Like in the sense of like for a regular company, like Guys like you, like you I you were at Target, right? Like Target had a world class red team, right? And I'll I I'm assuming I'm gonna brag on you because I know a lot about Target and I know how good you guys were. But that was
Jackson: Yeah, that's a really good guy.
Aaron: a through a tunnel of a lot of pain that they went through to dead
Jackson: Yeah. Yeah.
Aaron: build the world class team. But like if you look at RHISAC, right, like Target's the shining example of greatness, right? But like a regular company is never gonna have that. So do you think AI can make It's so everybody has Target's team. Like, do you think that's realistic? Like that level of quality, that level of depth, not just red team, but like across if you think about Target as like a world class security org.
Jackson: Yeah.
Aaron: Is that can I rubber stamp that across a thousand other orgs with AI? Do you think that's possible?
Jackson: I don't think so. I have this is sort of an ideological stance rather than perhaps a like technical stance, but I firmly believe that there are human capabilities and human values like AI is never going to be able to surpass. I think that there are aspects of creative thought that like it doesn't matter how great your model is. it is not going to beat good humans at this. I think that what's happening is we're like we are decreasing as models are increasing, and I think that's just like a a social problem. You know, if you don't invest in your society actually like getting good education, then it's gonna get easier and easier to build a model that outperforms them. You know, like so so I think that there's an aspect where, you know, if you if you sort of envision this as like the The chair people from WALL-E. As you sort of like degrade, there's gonna be
Aaron: Right.
Jackson: like, wow, these models are like so much better than the average person. Well, yeah, but that if that is ever true, that's a function of us just like not not investing in ourselves well. Okay, so that's a bit of an ideological side. But the point of all this is I think that unless and until we get to a point that I don't really believe is is going to happen, I think there might be like market forces that produce this outcome. But that is different from like it being an actual true necessary outcome and and the best outcome. If you ever get to a point where all knowledge work is taken over or preferred to be given to agents of some sort, if you get to that point, then like sure. Because now everything is just tokens and it's just who has access to what models and that's it. But as long as you have what should be the case of humans actually in meaningful decision-making capacities, and humans as like primary recipients, Microsoft AI put out great manifesto, Where they were like, technology serves humans. Like, technology exists for human good, your or core principle. As long as you still have that view, then you're always going to have AI subordinate to the direction of humans who understand these concepts better than they do. And so I don't think that you can ever just like Cookie cutter stamp, hey, now like the world-class team from org A is now over at Org B. Because I think what makes a team great, kind of getting back like Rob Joyce's statement, it's not the tools, it's not the exploits that make you so great. It's your ability to like think through and comprehend and understand an environment. what makes a good red team is not just like a technical ability to hack. There are a lot of people who are like awesome at CTFs that suck as red teamers because the concept is not can you operate? Or the kinds of can you hack? It is like, can you communicate, can you design to your operation a true story that communicates risk in an effective way to leadership to drive change? Which is very different from hacking. Hacking is a part
Aaron: Mm-hmm. Right.
Jackson: of it, it's a real and meaningful part. So and that applies to security teams as well, right? Like a security team is not one that like catches bits, it's a team that like protects the business, and ideally the business is also in the interest of protecting people and doing responsible things as well. So long-winded answer, but no, I don't think that I don't think that's gonna happen.
Aaron: So being in somebody who's living and breathing AI every day and thinking about red teaming and thinking about stuff, like what is the most scalable part to you? Like what's the part that you go like, man, if I could give everybody you know well not everybody, but like the the you know top five thousand companies in America this, it would really help like with AI. Like is there something that you're kind of man, just like code scanning or like vulnerability
Jackson: Yeah.
Aaron: management or like is there some area like man, if I could scale this with AI? You know, maybe it's inside red teaming of just understanding your risky assets. Like what is scalable then? 'Cause we have to scale or we're gonna get crushed, right?
Jackson: Yeah. Yeah. I think that I don't know I I feel like this is gonna be another like unsatisfying back to basics kind of answer. But I think that your ability to scale AI, hmm, maybe this is a better way to say it. If there is a thing that is like generalizably scalable, then that's just a commoditizable product that is like tokens in thing out, and it doesn't really i it's not really like a it's not a scalable thing. It's just sort of like use the internet, you know, use models. It's just sort of generally applicable. so an example of that might be I think the the research function in like Chat GBT or Claude or you know Groc or whatever you use is a handy one 'cause you're saying, Go read, you know, a thousand websites and produce to me like the summary of of what you've read. I think that's kind of a handy thing. That's probably the closest I would say to of a, hey, everyone go out and do this. Otherwise, I think what it really comes down to is deeply understanding your business and what you do and like how you do your job. Daniel Miser talks a lot about this. Of like a lot of companies can't apply AI effectively because they don't even know what they do and how they do it.
Aaron: That's very true. Yeah. Yeah.
Jackson: I think like you need to first understand what you do. Ads Dawson put out a really good post about like It was titled something along the lines of A Year of Doing Bug Bounty and AI or how I've used AI and Bug Bounty. And he he hits on this of I know how I I know how I do bug bounty, I know my workflows, I know my processes, I know why I'm doing things and how I do them and what my goals behind all of it is. And once I know that, then it's fairly easy for me to decompose my tasks and say, like this chunk of it, I'm gonna automate it. But that's not an aspect of like use AI to scale up on this thing. What it starts with is the the unique aspect of why do you do things, how do you do them, think through it, be able to document it and write it down. Once you could do that, then say, these are the things I can apply AI to to simplify my life. But you have to start with knowing your domain really well. And I think that's one of the things that's such a such a problematic lie from AI systems is you can be like, hey, can you create for me?
Aaron: No. They would lie to us?
Jackson: They they promise so much general capability. and so you'll be like, yeah, like one prompt and this git issue is squash and I can move on to the thing I care about and like build that feature. And you can't just be like, Hey, go fix it, make no mistakes. You have to like spend the time to think about how you actually want this issue fixed, and then have the agent go off and do it. and I think the Frontier Labs get around that a little bit. by doing like intelligence in depth and throwing ten thousand agents at solving the issue. and like their cumulative intelligence sort of solves it generally. But I think even in those cases you have a lot of tests a lot of like tests for it to turn green again. So it's again that sort of like reward function, science friendly approach.
Aaron: That's so to that end, like and I and I agree with like I think in general, at least in security, and I'm not an expert in other parts of IT, but like in general in security, by the time something scales enough where it works for everybody, it doesn't work anymore. To your point. Like
Jackson: Yeah.
Aaron: the it just it's like it's loses
Jackson: Yeah. Yeah.
Aaron: its efficacy because it's like everybody's doing it, then it's easy to bypass, enough research has been done in autobi like it's dumb, right? so custom is where it's at, right? Like to your point, like Really understanding your business, understanding your tax surface, which I actually do think AI can help you do a little bit to your point of like
Jackson: Yeah, yeah.
Aaron: because like a lot of times people don't do it just because it's a lot of work.
Jackson: Yeah.
Aaron: So one of the things that I'm fascinated by is I mean, and I've talked about this on this show and with other people, like token economics is gonna change. We're in this era of subsidized tokens, like
Jackson: Yeah.
Aaron: that era is gonna end at some point.
Jackson: Yeah.
Aaron: you know, I don't know if it's gonna crash out or if it's just gonna diversify or whatever. But like to to your point about custom, this is something I'm I'm super interested in. I also don't believe we're gonna exist in a world in five, ten years where we all use the same models. I just don't believe it's possible. I
Jackson: Yeah. Yeah.
Aaron: don't think it's to your point about being custom, like it's too generic. Do I really think that Bank of America and Chase are gonna use the same AI? No, I don't because how are you gonna win if we're all doing the same shit,
Jackson: Yeah. Yeah. Yeah.
Aaron: right? So to that end, when you talk about custom and custom security and custom Defenses and red teaming, how much do you think people should be thinking about models and training versus what frontier model has the best AI coder? Like, who gives a shit if you Cloud Code or Astro? They're all taking to the same place.
Jackson: Okay.
Aaron: Like, how are you thinking about models? And like, where are you where you at with that part? Like do you think every company ends up with its own model? Or do you think we all end up communally sharing frontier models?
Jackson: I don't know. I will I'll give a quick plug for some research that Arvind Narian and and another person maybe from Princeton as well did. But they did some pretty cool research where they were looking at like new technologies that emerged, sort of like infrastructure versus application layer. Good good paper overall, but their takeaway was, you know, these companies right now, the the anthropics, the open eyes of the world, they are pushing out just like these commoditizable pipes of inference. And they need to figure out how to get into like application to to survive. And I think that perhaps what you're going to see is this world where the frontier labs, particularly in the US, like closed frontier labs, are going to desire to become the headwaters of intelligence. And if you want like a drop of intelligence, you need to go to them. and like everything will sort of flow down from their models and their servers. I think that's a future that they want.
Aaron: I'm sure they wish that was the case. Yeah. Yeah.
Jackson: I truly wish that was the case. And I think they're gonna you're gonna try to engineer it. I think that there are some I think some of their desires around regulation are a mix of genuine security and safety concerns as well as like incidentally beneficial perspectives on
Aaron: Yeah.
Jackson: market regulation. But So okay, so I think there's there's that side of things. I think that there are a lot of customizations that you can tag on to models. Even something so simple as like memory systems or rag or you know graphs where you're storing data that allow you to have, you know fairly meaningful control over how the model produces its outputs. and can give you some differentiation there. But I do think that okay, quick quick like Quen plug. I think, you know, Quen 3.8, there's been this like recent rash of Chinese models that are past like a a certain band of capability feels, or now you can run them on like consumer or prosumer hardware, and they can actually do real things for you in reasonable amounts of time. Whereas prior models, like even with unreasonable amounts of time, couldn't necessarily do real things with for you without a lot of work. so I think that we are the the point there is I think we're going to be seeing a lot of people now having incentive and ability to do a lot more fine-tuning of models and to say, hey, like I want, you know, Jackson's special version of Quen. And I can do that, like not that hard with spending some bucks on RunPod or sending somebody over to thinking machines and using Tinker. I think that there's going to be like more of that that's happening. And I think that's going to give us more data around how much is that really the future. I think right now there's also a factor where But we just like don't want to give it, and as we do more and more prompting, it's like harder and harder to discipline yourself to do that. So I think that there is a huge amount of incentive and and pressure and hype that's forcing people or that's driving people to use way more intelligence than is required intelligence than is required for a given task. And I think there's also certainly from like large companies this fear of if we don't jump on this bandwagon now, it may be too late for us. And I think there is a lot of like time pressure panic across the entire industry around AI. but I I think over time more and more special purpose models. Desert Ant put out some really cool, really light, really fast models that do like exactly one thing and they do that one thing well. and so I think we're gonna see more of that. I don't know how much that is going to become sort of like the cumulative overwhelming mass of stuff, or if that's gonna be sort of like filling in the the mortar filling in all the cracks, I'm not sure. But I think the future is not going to be just like one massive model or five massive models.
Aaron: Yeah.
Jackson: I guess I'll all
Aaron: Well
Jackson: go back because I I wanna touch on the answer that I kind of didn't give last to your last question around, you know, what is scalable. Let's say short version is does it require judgment or is it mechanical? Because if it is just like a mechanical thing, then like sure, scale that up. If it's just like, you know, data entry, take this thing here and put it in that field there, then a lot of that is just how do you put like a good harness around it so it does it reliably, how do you audit it, how you test it? But that's sort of it. If it requires judgment, then that's where it's it's much harder to scale. Okay. Sorry, back to the the next question you're gonna ask.
Aaron: No, it it's so a and just for the sake of our audience that may not be living in the weeds of AI, what's Quent?
Jackson: excuse me, Quen. It's a
Aaron: Like what is that? Like explain it to everybody.
Jackson: Quen
Aaron: Yeah.
Jackson: is the it's the clod of Alibaba. So Alibaba, Chinese large company, sort of like an Amazon or a Google here in the States. They have trained a class of models. They're Quen, Quen three point eight is their most recent set of models. and so like three point eight twenty seven B is one that can fit on quantize so it's like smaller not quite so precise but maintaining most of the precision that can fit on like consumer prosumer cards and then there are these like mixture of experts models that are quite fast as well that can fit on cards. and they've got like three point flash and or next flash and max and those are like the really big ones that you need hundreds of thousands of dollars of chips to run. But yeah yeah neat new models.
Aaron: Interesting. So I I I'm one of these people, like I'm having a hard time with the Chinese thing. in I've spent like so we went through a whole period in cyber where we blamed everything on them, right? Everything was China's fault. Every advance attack was from China and it was just this big hypey blame China. I was just at Falcon and there's still the statues up of all the Chinese threat actors and The cartoon. Like it's what we did. Whether it was right or wrong, we did it. And
Jackson: Yeah.
Aaron: now all of a sudden it's like, yeah, use open router and route your shit to Kimi and Quen and like you'll be fine. It's but like
Jackson: Yeah.
Aaron: I so I believe we've entered this insane era of YOLO security. Like we are just giving our data away to be lazy in a lot of cases. Like I know CFOs are out there going to Claude. Clicking the plus button, adding confidential financials and asking questions. Like
Jackson: Yeah.
Aaron: I know it. I I know it's happening, right? What what are we doing here? This is like all the things we've spent two decades trying to stop and we're just self-owning. And then I see the same thing around Chinese models. I'm like, you're just giving them your data? Like, because it's 12 cents a token cheaper? Like, what are we doing here, man? Like, this is insane. And,
Jackson: Yeah. Yeah.
Aaron: you know, I I was just listening to Pod with the CEO of OpenRouter, and he's talking about how this is great thing. Like, just wherever it's cheapest. And I'm like, that's not. how security works. Like this is a nightmare. And so like how are you transversing that? Like how are you thinking about that? Like are you do you care less because it's AI and who gives a shit? Or like how are you thinking about the the Chinese model?
Jackson: Yeah. I think running local weight model or running open weight models locally is different because you're not sending all the data out into the ether. And I would also say, I have zero I have zero insider knowledge or zero evidence on this. However, we've seen from public reporting from the Frontier Labs some security practices that I would describe as like quite surprising for a trillion dollar enterprise, or at least quite surprising for what you would hope to expect.
Aaron: Very nice way to put
Jackson: I I am
Aaron: it. Yeah.
Jackson: there that is top tier I top tier diplomacy Jackson there. You know. Pretty pretty poor security practices. And we also know, and what's the likelihood that inside OpenAI servers and inside Cloud's servers, you can get the most cutting edge information on regulations that are being planned inside Congress, cutting edge information around like what Department of War is thinking and planning.
Aaron: Hundred percent.
Jackson: Surely this exists. I think there have even been like reporting about this happening. So
Aaron: Right, course.
Jackson: if I am a Chinese or a Russian or pick almost any country on Earth intelligence agency, why would I not like, hey, those companies that have a huge amount of data and possess some of the most intimate thoughts of a population I care deeply about, why am I gonna go after that? So all what I'm saying here is, the Chinese might have your data anyway, even if you're using Claude or Chat GPT. Certainly if you're like a certain class of person, I think you should expect that. but then, you know It depends on how you're using the language models. So Using a Chinese model to do, let's say, like vulnerability research and exploit development for a US government agency, probably not a great idea without a whole lot of thought put into that. CrowdStrike put out some research showing that some of the Chinese models write software with more vulnerabilities in it when it's like related to doing something like Uyghurs or Falun Gong or something like that.
Aaron: Mm-hmm.
Jackson: there's no There's no like root cause of that. So is that because the Chinese government specifically said, Hey, Alibaba or you know, Moonshot or whoever, like you need to train your models to produce vulnerable code so we can exploit the Uyghurs more. maybe my personal gut feeling is some of the training that they've done in censorship has made the models like dumber and like weirder when it comes to certain topics that they care a lot about controlling. And so you might just be seeing sort of like the side effects of that. But the bottom line is it would not be hard for the Chinese to say, hey, Quen, if you're ever going to write exploits for let's say Chinese software, make sure it includes like this kind of string or this kind of pattern, to make it easier for them to like scan for it internally. That's plausible. so yeah, I think if you are going to be using Chinese models, I would tend to use them locally. rather than like hosted somewhere else. Hosted hosted by an actual Chinese provider. and then also when you're using it locally, thinking really seriously about what the tasks are that you're using it for. Using it to like jailbreak another model, sure. using it to look through your finances personally, sure, fine. using it to draft legislation? No, not that one. so I don't know if that I don't
Aaron: Same.
Jackson: know if that's a useful answer. That's It's not concrete guidelines, but
Aaron: Yeah. I you know. Russian software, Chinese software are always an interesting debate in like
Jackson: Yeah.
Aaron: they have a right to exist and be in commercial enterprises too. And like, you know, I'm not I'm not necessarily making broad statements about universal intent, but like I think one of the things I tell people at AI is like if you can't afford to do it right, you can't afford to do it. Like, don't do dumb shit. I guess that's a number of rule of security in general, but like I I think about we're gonna need a generation of cybersecurity people. Who know how to build local AI. Like we're gonna need that, right? Cause like to your point, like frontier models, let's say, let's just assume they are always gonna have their place because of how they're pushing the boundaries against smarter and smarter, smarter. But like
Jackson: Sure.
Aaron: to your point, like if I'm doing baseline automated tasks, I need some level of AI intelligence. Do I really need to pay extra token for Sonnet? No, like I'll just build the model,
Jackson: Yeah.
Aaron: right? This is especially true security, where we do a lot of blog parsing and boring shit. And it's like Just get a local model or get a cheaper hosted model and do it. But like that requires architecture that like
Jackson: Yeah. Yeah.
Aaron: we just don't know yet. Like I I haven't met anybody who understands this in a sense of like if you're at a Fortune five hundred and they're like archite architect a mixed modal local and hosted AI infrastructure securely. I don't think anybody on the planet has that answer. Like, at
Jackson: Yeah. Yeah.
Aaron: least not that I've seen. No, there's a lot of vendors who tell you they don't want to do it, but they're full of shit. Because like to your point about commodization, like they may know in the abstract, but like in your workflow, what's sensitive? What's
Jackson: Yeah. Yeah.
Aaron: not sensitive? What should you put in cloud? What shouldn't you put cloud? I haven't seen that at all. so I think it's a definitely a a a an interesting perspective. So the way I want to bring this back to you and and your question is through your boy band work and then through your your day job running a company and things like that. Are you being asked to do that? Like or at least test that concept of like, here's a a workflow and we use four different models and four different harnesses to do different tasks. Like test it and make sure the data is make sense. Is that something you're getting any of work around?
Jackson: So I'll I'll give you two answers to this one. One is so for Barding, quick overview on us. Like we we're a software and a data company. On the software side, we produce essentially like offensive cyber harnesses to help organizations better understand their risk. I guess I'll talk about this a little bit more because I care about it. It's part like like why I built Barding. There are a lot of companies that answer the question, you know, will we get hacked or should we expect to get hacked or can't we get hacked? But at the end of the day, and this is gonna actually touch on the model thing in a moment. What you really care about is can I handle an intrusion? If I do get hacked, will I be okay or will I not be okay? And you kind of don't care how much you're hacked if it doesn't impact you at all. it's sort of hard to imagine getting
Aaron: Sure. Yeah.
Jackson: it getting hacked and and not being impacted, but really what you care about is when it happens, am I gonna be fine? so that's that's like specifically the kind of harnesses we're building are to help organizations understand that really effectively at scale, in safe and efficient emphasis on efficient ways. but I think when it comes to models and like how you architect your models, I think a lot of the or how how you architect your your infrastructure, your setup, a lot of it is like, hey, will it be okay if this if this goes poorly? And figuring that part out. And then to answer your question more directly around, you with Bt six or with Barding. So from a Barding perspective, people aren't really saying, Hey, please audit our setup. You know, we're we're selling software for them to do better security assessments. And then with Bt six, back in when was I think it was like right around when the like the Claude Four series came out. There was a shift in the system cards for models. So model evals have sort of generally always included cyber harm sections. And then it used to be that when Frontier Labs were put so model cards, system cards, it's like a you know, hundred page or many hundred page report that walks through from the lab, like this is what we've tested, these are the third parties we brought in. And this is what we are concerned about. These are like interesting things of note. it has more hallucination rates, lower hallucination rates, better at handling prompt ejection, you know, loves writing malware more than the other one. tried to
Aaron: Sure.
Jackson: more so like it walks through all those things. and in the system cards, right around like the Sonnet four era, there was a shift. The same across like anthropic and open AI. They stopped doing like purpose-built cyber harnesses as sort of their primary way of tracking cyber harms. And they shifted to like the model itself with bash, or like the model itself with Claude Code and the model itself with Codex. And I think that this is an area where Frontier labs have an academic interest that is like poorly aligned with the interests of the public. Because as a frontier lab, if I am producing a model, I want to know in sort of like a nice, clean, scientific, physics, frictionless vacuum, I want to know like how dangerous is my model. And I don't really care about like A harness that makes my model more malicious. I care about my model and what it can do and the harms that it can do. so I think there's this desire to sort of get a a purer read on the raw capabilities of the model. But what that's doing is it is saying, like, maybe here's the floor of harm from our model. But it's not producing like here's the here's a higher level of harm that our model could possibly do. And the Frontier Labs should be trying to communicate both, not just here's the floor. and it's sort of like here's the floor under some ceilingy circumstances. It's kind of a weird way. But the point of all this is the Frontier Labs I don't think are very interested in saying, Hey, what is what is the risk when you have like these ten factors all converging? sort of in the same way Microsoft is like, we don't really care about anything that's not something we track as a security boundary. How you set up your environment's your problem. Amazon shipping like insecure defaults in AWS for a really long time. It's like, hey, it's your your job to configure that. So the Frontier Labs have that interest, and then from a lot of the companies, I think the business risk has been primarily being perceived as being behind. And so where you've had money and interest, it's been focused on getting AI adopted as publicly as possible. And so you've had all this like token maxing culture where people
Aaron: Right, yeah, yeah.
Jackson: For spending so many tokens and wasting money and then people, you know, blow through their token budget in a quarter. So all this is a very long way of saying no, but the intention is to communicate why I believe the answer is a no. Does that answer your question?
Aaron: So yeah, it does because it and I guess there's a a nuance there of the model itself understanding the risk of a model and its behavior, not necessarily the weaponization of the model by other actors or other people. Is that is this like a gun don't kill people, people kill people kind of argument? Of like, we just made the thing and I don't fucking whatever you do with it is your problem. Is that kind of the way you interpret it?
Jackson: Yeah, it's it's it's kind of like that and I one can understand and sympathize with and I think even agree with it for from the lab's perspective in many respects because like the labs don't want to pick a harness and then really what they're trying to communicate to the public about like risk of their model is really just evaluating the harness and the combination of their model and the harness and criminals are gonna use some other custom harness. So I think there are good reasons why they're not just locking themselves into a given harness and and doing their evals that way. But I think that in their quest for quantitative data, they're really missing out on much more effective ways to understand the harms and risks of their models. and no one's really incentivizing them to not miss out on that. So there's some problems
Aaron: Yeah. Okay.
Jackson: there. Does that does that answer the question?
Aaron: Yeah, to a hundred percent. And you know, I I think as a security professional, that makes sense because that's to your point about AWS and Microsoft, it's always been this way, right? Where like
Jackson: Yeah. Yeah.
Aaron: I'm gonna give you this technology, the implementation and risk of that is yours.
Jackson: Yeah. Yeah.
Aaron: But that's not how they sell models. Right? They sell models as like it's a combine it's a finished product, right?
Jackson: Yeah.
Aaron: And so shipping Windows with phones is problematic, but also At the end of the day, they're not hosting Windows for everybody in the world. So, like,
Jackson: Yeah.
Aaron: it is your problem. Claude is Claude, like we all use it, and it's like they're hosting it. And so, if they're not sharing relevant risk data, that's more concerning to me because it is their problem. Like they are, I'm paying a Claude bill every month. I'm paying a Gemini bill
Jackson: Yeah.
Aaron: every month. I'm paying an OpenAI bill every month. I'm sure you are too. So that's problematic, I think. But I also agree. It's just one of those crazy because I also agree with you about harnesses. Like, you know, I don't know if you're a race car guy, but like in F1, a bunch of teams use the same engine. But they still get their
Jackson: Mm-hmm. Yeah.
Aaron: ass kicked by the teams that know how to use the engine better. Right. Like Mercedes
Jackson: Yeah, yeah, yeah, yeah, yeah.
Aaron: will sell engines to anybody who wants to use a Mercedes engine, but Mercedes
Jackson: Yeah. Yeah.
Aaron: wins the race because they're in Mercedes and they know how to build
Jackson: Yeah.
Aaron: the harness around it. And I kind of view AI the same way of like it's in the hands of the user, right? And the harnesses and things like that. So I guess like if you're a shop and you're trying to understand the risk of AI and you're not getting version you need from the models, which is fine. Where can you get it? Like w how should people be thinking about this? Like, where are they getting their information? Should they call you? Like where
Jackson: Brown red.
Aaron: like if you're a bank and you're like, I need to understand the latest models and the risk to me from the outside in, where do you go? Like you're not gonna figure it out yourself, right? I mean, probably
Jackson: Yeah.
Aaron: not.
Jackson: I think there are Yeah, I I think the the list of items that fall under the header of risk to me from a model are immense and varied, right? 'Cause it is like what is the risk to me of using this model inside my environment? and then there's also like what are the risks to me of somebody else using a model, like using
Aaron: Yeah, I'm specifically talking about the engine in the hands of threat actors. Like data management,
Jackson: Hm, yeah, yeah.
Aaron: I'm like probably me and you don't spend all of our day on that. Like we're very worried about adversarial security, red teaming attackers. I do detection engineering.
Jackson: Yeah. Yeah.
Aaron: So like those things matter to me. Of like, how are you supposed to know what the capability is? Even if they're not doing it yet. Like you to your point about we've always been in a call response environment, like I don't know what I'm responding to.
Jackson: Yeah. Yeah. How do you how do you start predicting? Yeah.
Aaron: Yeah.
Jackson: Wow, this is gonna be hard for me to try to be brief on because this is one of my favorite topics. I think I would say a couple things first.
Aaron: Great. Keep going. Go for it. I want people to wanna know.
Jackson: So I'll start with a more concrete one first, and that is like look at the blog posts from like red team and cyber researchers using AI. The reason I call this out is there have been a number of things where there have been like academic papers published and like discoveries released. From the AI side of the world and like the ML PhD side of the world. And it's like good work that they've done, it's written up really well. But they are discovering things that some red teamer blogged about three or six months ago and like it was just a a means to an end to achieve their red teaming goal. I think
Aaron: Mm-hmm.
Jackson: that is true of any kind of AI adoption. The person who's there, I sa I saw a webcomic once that I think about rotten, and is like y basically you ask a math fan how many digits are there or like recite as many digits of pi as you can and they can go on for hours. And you ask a mathematician like how many digits of pi do you know and they're like one? They don't really they use it. I mean they're a lot more interested in using it to achieve their specific goals than they are in like sort of focusing on Pi itself. and I think that you have a lot of AI folks that are really, really interested in AI, but the folks who are trying to achieve stuff with AI are the ones who are sort like incidentally doing some cutting edge research. They're incidentally doing a lot of dumb stuff along the way as well. So I don't want to say like, only the people doing work matter. Like the people who are building it also know a lot. But they know different things in different ways. So yeah, watch watch the blog posts from those folks. Adam Chester in particular has done some really good work. Nick Downer from Rogue Labs has also done some really, really cool stuff around using AI to b both of them have done really cool things around sort of like autonomously looping AI to produce really cool offensive cyber capabilities. So that would be the first thing is when you're trying to understand risks, look at how are the people who are doing the risky things using AI, and how are the most creative of them using it. So that would be one way to look at it. Second way to look at it is don't try to look to the future, look to past. My favorite analogy, so I'll all you I don't know, it ultimately is a weapons analogy, but my favorite analogy for AI and the the social shifts that AI is gonna drive as a whole is the Bronze Age. I think it's like one of the very, very best examples to look at in terms of a new technology That had a a a huge amount of effort, both in terms of how you would like collect and forge and manage bronze and like all the smelting process, and also how you would actually like adopt it and apply it. And then it had huge military, cultural, economic impact. But when you look at the first Bronze Age tools, you can look at an axe, for example. The very first Bronze Age axes, they took the Stone Age axe design, we pulled out the stone axe head, we cast the bronze axe in the exact same shape. shoved it in and like that was innovation at the time. And I Yeah, and
Aaron: Ta da. Right.
Jackson: we're we're kinda there right now with AI where a lot of harnesses today are how can I take the things that a human does or take the things that a static you know piece of automation does, shove out the Python, you know, pull out the human, ram in some AI instead, and you have some sort of call to an inference provider. And now like this is an innovative new thing. But over time, you see the actual design for an axe. It's really cool. if you can like look up you know, British Museum photos or like YouTube videos. The design of an axe actually changes really differently. And it doesn't look even something like what we would recognize as an axe today necessarily. that that sounds a little overblown. It's just it's a weird shape. but it is designed to take advantage of the unique properties of metal. The metal can be sort of like bent and curved and it doesn't have to just be like one block. and that happened over a thousand years. I think that we're gonna be seeing that from an offense standpoint as well. And so I think when you're thinking about risks and harms Don't stop at what happens when we pull out the stone and shove in the bronze. Because that's interesting, of course. And I think that's where a lot of these like fully agentic swarm is taking over the world and a lot of these approaches of like, you know, AI's gonna be really great at discovering volumes and everything will have zero days all the time everywhere. I think that's a lot of where that comes from is we're saying, like, well what if we take out human researchers and we shove in AI and my goodness, this is so scary? And sure that's gonna be a problem to some extent. But I think thinking more about like what can AI enable From a risk perspective that was never possible before. Things like zero cost patrons, not just like enhanced scale. I think that is I realize there's still like a lot of work to do. I'm just sort of saying like think in this direction as opposed to that other direction. But that's a lot of how I'm looking at it. I think a lot of how we're looking at it at Barding as a whole, is that sort of shape of how is this gonna change over the next years as we start adapting better to use it? Is that a useful answer, Aaron? I don't know if that's like too conceptual and high level.
Aaron: No, it is, and I think about and I agree with every all that. I I think about how to apply it in like a day-to-day
Jackson: Yeah. Yeah.
Aaron: life. Because to some extent, to use the the bronze age analogy, like we didn't understand it yet. Like we were still kind of learning
Jackson: Yeah. Yeah.
Aaron: what metal could even do or what you know, what all these big technological leaps. And my biggest concern about this for for Me right now being a cyber person is before and it's funny, it's come back to this, but it's basically Jeff Globoom's speech in Jurassic Park, which is before you knew what the hell this even was, you packaged it and put it on a damn lunchbox. Like you guys don't know what you're doing with this. Like
Jackson: Yeah.
Aaron: they don't know, they have no clue what they're doing with this from a security
Jackson: Yeah. Yeah.
Aaron: perspective. They don't know what they're unleashing on the world. They don't, we don't really know what threat actors are gonna do with this shit yet. Like it's it's all chaos. So, okay, so you you and I and regular people can't control that. So what do you do about it? Right. So I I am in like I talk to a lot of people about AI, and I'm like, are you meeting every week about this? Are you talking about at your bank, at your hospital, how this could affect you? Are you thinking about what your defenses look like? Are you pushing your vendors? Are you thinking about not just consuming and buying and hoping and praying, but like really being thoughtful about it? and
Jackson: Yeah, yeah.
Aaron: I would say the answer universally is no, because and you mentioned this earlier, all the oxygen's been sucked out of the room to just go fast. Like fast, fast, fast, fast, fast, fast, fast, fast.
Jackson: Yeah. Yeah. Yeah.
Aaron: And it's funny, you were talking about this with in context with tax. Like I spend a lot of time thinking about velocity, right? And AI making things faster. The more terrifying answer to your point is AI making it slower. Like nobody's talking
Jackson: Yeah.
Aaron: about that. Nobody's talking about one event a week coming
Jackson: Yeah.
Aaron: from some really patient AI system that somebody put
Jackson: Yeah.
Aaron: on cruise control against your orc. Like it's terrifying.
Jackson: Yeah, and you're running against a huge amount of them. And even things like, you know, my my toy example is what happens with when you say, okay, I'm going to very, very slowly run an intrusion against like the top ten lumber companies in North America. And it's fine if it takes a year. I get all of these,
Aaron: they're fucked. Yeah.
Jackson: and then because I've done this quietly, now I can deploy ransomware across all of them at once. And now I'm holding an industry hostage, and I can say, Hey, North America as a whole Once you pay a hundred million Bitcoin to this address, then like you can build houses again. otherwise like sucks to be in San Francisco. Yeah. Yeah. So I think Yeah. I don't know, a couple of things
Aaron: Yeah, bummer. Yeah. yeah, the
Jackson: I guess.
Aaron: sort of ops you can build,
Jackson: Yeah.
Aaron: you know, consuming customer. I like we've already seen some of this, you know, operationalness in things like elections. like I had
Jackson: Yeah, yeah, yeah, yeah.
Aaron: Randy Rosa on the show who who runs Multi-state ISAC and election ISAC. And it's like
Jackson: Mm-hmm.
Aaron: they don't know. And by the way, nobody's helping them.
Jackson: Yeah.
Aaron: Like nobody,
Jackson: Yeah.
Aaron: like nobody who knows shit about AI is going to election ISOC and be like, how can
Jackson: Yeah.
Aaron: I help you? Make sure our nation's elections aren't fucked up by AI. They're just on their own. These election boards in some county, no clue. None. And nobody's helping them. Our government has completely given up responsibility for that, which it sucks. And defunded it, which is even worse. And not to get political, but like this shit matters. Like
Jackson: Yeah.
Aaron: I don't care who you vote for, like I want to know I'm voting in the right place. Like that's good. And so
Jackson: Yeah, yeah, yeah. Yeah.
Aaron: like I I am I think and I'll I'll bring it back to the beginning because I I I I want to be respectful of your time and I I know we've run a long episode, but I I love this. I get up with this all day. But you know, I think you know, the energy has to put in has to be put into people who do know being thoughtful and helpful, right?
Jackson: Yeah, yeah, yeah.
Aaron: Like Whiting white paper is great. But like, does anybody know about it? Are we turning it into usable information? Are we sharing it the right way? Are we helping
Jackson: Yeah, yeah. Yeah. Yeah.
Aaron: people? Like, we need to move faster on that side of things. And so and I know you're doing that, man. I I read your blog, you write awesome stuff.
Jackson: Yeah.
Aaron: I know you're very into into responsible AI usage and I think you do a great job of calling out the labs when they need to be called out. which I guess when You don't have anthropic or open AI stock is easier to do. you know, I guess unless you're a researcher and you quit and say, I shouldn't have built this Frankenstein and you there's all this crazy shit happening. And and
Jackson: Yeah. That's pretty good for me. Yeah, yeah, yeah.
Aaron: if it goes bad, I'd love to bring you back on the show and talk about how it went bad.
Jackson: Yeah, I remember.
Aaron: but I know you're out there gonna fighting the good fight. I'm trying to help with sharing community, trying to fight
Jackson: Yeah.
Aaron: that fight. And you know, I fundamentally believe like it will get bad and then I think we'll figure it out and it'll get better. I'm excited for that. I think I think we have a great future ahead of us in cyber. I think it the next two years might suck, but that's okay. all I tell everybody is just be ready to work a lot of late nights when these breaches start because like shit's
Jackson: Yeah.
Aaron: coming down the pipe. but I hope you feel that way too. maybe you could share really quick with the folks like where they can find you. I know you have a blog and some other stuff. I'll put it in
Jackson: Yeah.
Aaron: the in the Episode stuff
Jackson: cool. Yeah.
Aaron: too, but maybe we could just talk a little bit about that.
Jackson: Yeah, yeah. quick places you can find me, my blog, twelve tables dot blog. So you can hit me up there. you can find me on LinkedIn, and then also like the the Barding website, barding defense dot com. We have a contact us page and you can ping me there and that goes to me as well. So any of those places.
Aaron: So I'm upsessed with names. Why twelve tables? What does that mean?
Jackson: so 12 tables, it is one of the first sort of like founding legal documents. So 12 tables is a little bit weird because there are a lot of things. So it's it's Roman and it's like laws,
Aaron: okay.
Jackson: essentially like core laws and manifestos around society. some of them are pretty bad, actually. but like Well yeah, but comparatively,
Aaron: Okay. Well what do they know?
Jackson: it was a significant improvement. And even the fact that like they were writing down and codifying their laws was a big deal. So I wanna be clear, I'm not like a a fan of proponent of like every law on the twelve tables, but I'm a big
Aaron: Right.
Jackson: fan of transparency, I'm a big fan of accountability, big fan of governance and like
Aaron: Got it.
Jackson: justice being equally applied, and like social good, all these things should should be able to coincide well together. So yeah, that's that's why twelve tables. It was initially it was initially
Aaron: Love it. That's where it comes from. Okay. I I didn't go to the same school as you did, I don't think, so I don't know what that stuff is. but that's very cool. awesome, man. Well, thank you so much for taking the time. Everybody can find you on LinkedIn. everybody
Jackson: Yeah, of course.
Aaron: can read your blog into that, and then check out Barding, of course. I'll link to all that on the episode description. dude, thank you so much for taking the time. It's so great to have you on the show. I I love to
Jackson: Yeah, yeah. Really enjoyed it.
Aaron: hear about people on the cutting edge of this stuff. It's really, really cool. so thanks so much, man.
Jackson: Yeah. Yeah, and for yeah, and for any viewers like where I got stuff wrong, where you're like, hey, this guy's just completely incorrect. yeah, hit me up and and let me know. Super happy.
Aaron: Find you at a con and tackle you and tell you you're full of shit. I get plenty of those comments, believe me, 'cause I'm always full of shit. no, man, it's so great to see you and thanks for taking time. Appreciate
Jackson: Yeah. Yeah, thank you.
Aaron: the idea of scale and patience is terrifying
Jackson: Yeah. Yeah.
Aaron: because offensive defense, all hackers are somewhat lazy, right? In the sense like we love to automate things, we love to like it's just drives our whole industry, it always has, right? And really, really smart people given the superpower of patience terrifies me.
Jackson: people are shoving way too much intelligence at problems because it is easy. I I think probably everyone has had the case where they're like lazily trying to solve a problem and they're using like sonnet and it's not working and they're like, alright, fable, you go do it. I'm not gonna pay enough attention to go do this, like you solve this problem for me. And it's not a fable level problem. It's a sonnet and like one percent of your brain kind of problem.
Aaron: Hey everybody. welcome to the ZeroSum podcast. I'm your host, Aaron Mog. so great of you to spend some time with us today, listening and watching us. really excited to have Jackson Reed on the show. Jackson's a a new friend, somebody that I'm getting to know and learn about in when I heard what he was up to, was really excited to invite him to be on on the show. I get a lot of feedback about what we're talking about and AI and the future of AI, the future industry is is certainly an interesting topic. So, first of all, Jackson, thanks for being here.
Jackson: Yeah, of course. Yeah,
Aaron: Really appreciate it, man.
Jackson: glad to be here.
Aaron: So for those of our listeners that may not know you or know your background, maybe you could give us a little bit about your perspective and where you're approaching this from.
Jackson: Yeah, yeah. So I guess I'll I'll start near the very beginning with not being a technical guy. I did not think I was gonna go into technology, did not think I was gonna go into cyber. yeah, I was gonna
Aaron: Really. Interesting.
Jackson: go into diplomacy, I was gonna work for the State Department, do national security kind of things. That was the plan. that's what I went to school for. But then second semester of freshman year I took a course and it was all about space and international affairs. And I'd never thought of things like radio frequency is a scarce natural resource. You can't like carve it up into infinitely fine little bands. And so states of conflict, not just over like, this is my land, but also this is my band of of spectrum and I don't want you to use it. I never thought of that before. I never thought like wow, there are all
Aaron: Yeah.
Jackson: these all these negotiations and conflicts that are happening in a sector of the world I didn't think about. the professor was talking a lot about like how law was being made and policies were being decided. It's like a bunch of people in a room making a choice. They're like maybe this is the right call, I don't know, we'll find out. I thought that really interesting. Wanted to do that for cyber because I thought we're kind of at that stage now where people are making these five by the seat of their pants choices that are impacting law and society. Yeah, never. Yeah,
Aaron: Us in Cyber? Never. That's never happened.
Jackson: so I w I wanted to like do the policy work there, but I wanted to be technical enough to actually know what I was talking about. So I started getting more technical and then it just like sucked me in. So finished up undergrad with International Affairs, did grad school in computer science through I guess I'll give a plug for scholarship or service or cybercore if anyone has kids. Or if anyone is young enough to be thinking about this, it's an incredibly awesome program. Think Rotzi for Nerds. So that paid for most of undergrad and grad school. And then I did work, so it did think tank work, did some sock analyst work, discovered that I didn't love that. but it was a great great learning experience and good folks to work with. And then went over to CrowdStrike to do threat attribution and intel analysis work. that's where like I gotta bring the international affairs stuff and do China stuff. and then went over to the US Army Corps of Engineers to build and lead their red team. So everyone probably
Aaron: very cool.
Jackson: on this on this knows red teaming, but Department of then Defense, now Department of War, has like a particular process for becoming a certified and accredited red team. So did that, took the team through CNA. that was exciting, did that for about four years. Really fun to operate on like really interesting, cool systems that you just don't get to in normal life.
Aaron: Yeah.
Jackson: and also just a really, really fantastic team. So that's kind of where I learned CyberOps stuff, had a great time there. Went over to Target to rebuild and lead their red team. Did that for about four years and then in March launched Barding Defense. So we can talk about that to whatever extent we want, but also I guess I should say sort of overlap with a full career stuff. About three years ago started working on AI, saying, okay, I can no longer just sort of like watch this with mild interest from the sidelines. I need to actually get involved. So started doing, you know, sort of the babies first model, the hello world versions of, you know, I'm gonna train my own little classifier, things like that. and then started getting into mechanistic interpretability. with you know looking at like how does model actually produce the outcomes and can I at that time I was saying how can I break models and how can I use mechanistic interpretability to understand like better this geo break is going to work and I know it's gonna work because I like watched other prompts function. So that was exciting did that for a bit. and then started working with a group called BT6 doing assessments and geo breaking for Frontier Labs. And then sort of from all of that is what Barting comes from, the combination of social impacts, cyber, then AI. Does that give the is that like a covered review?
Aaron: Yeah, yeah. I was like BT six because it's like the most hardcore group of AI hackers with the most boy band name of all time. it's like we're
Jackson: Yeah.
Aaron: gonna hack your frontier model, then we're gonna go on stage and perform.
Jackson: Yeah.
Aaron: but that is a a group that made up of some some world class AI hackers and thinkers.
Jackson: That's pretty cool.
Aaron: so that's a lot of depth and a lot of understanding, I think, of AI and It's funny that somebody who's been doing it for three years has as much depth as anybody because it's all so new. but I I I think it's it's something people are are tremendously interested in. But let me take you know, take it a step back because obviously being a founder, you know, being involved in sort of the cutting edge of AI, one of the things that we like to talk about on this podcast is our industry and kind of where we're at and how we feel about it. And I'd love to maybe start there of like where like You obviously are diving in the commercial side
Jackson: Okay.
Aaron: of this of starting a company and you and you obviously have been doing cutting edge AI work.
Jackson: Yeah.
Aaron: Are you hopeful, upset? Also, just a level set for anybody listening, depending on your listening, we are it's Friday, September 18th, 2026. The hype day in the history of AI so far. The overwhelming amount of bullshit coming out of the labs and out of the news. Is it's just unbelievable. So we are also in this like panical era of just AI craziness. But like what do you think? Where are you at?
Jackson: Yeah. I sort of I feel like I have all of the emotions, in like different sectors depending on on the exact topic that's being discussed. I think probably my overall view is just like you said, there's an insane amount of buzz and hype and there's like a reality distortion effect around AI. There are people who are like very competent, very experienced people that will make statements to be like that is objectively and demonstrably false. Not like I think it's false, like Here's the data show that that is not an that's not an accurate claim. so it is a very weird field. There's a lot of strange statements. but I think we are in for a pretty bad time. I am not personally sold on the you know, swarms of self ref self replicating agents will take over the internet in some like utterly unstoppable way. I do think loss of control is a real risk that we should be talking about. I think that most of what we're labeling loss of control is like not actually loss of control and solvable through like fairly easy approaches. But yeah, I I think I think we are in for a bad time. I can talk a little bit more about what I think what I think is gonna drive that. There's yeah, I sh sh short short version on it is I think we're so so early in
Aaron: Yeah.
Jackson: this. you can look at right, history history is gonna come into play here. You can look at like every time there's new technologies people like really get fixated on, wow, this new technology is really transformative, which it is. and then they say, and therefore and the next step they make is wrong. and I think we're in that phase right now where we're saying, Wow, this new technology has come out. It's really transformative, that's all correct. And therefore And then that's where like all sorts of insanity occurs. I don't know if that's a a complete answer. Short version is like I feel bad, I feel good about some things. I think most of the most of the harms that we're talking about that we're really concerned about right now, I think are harms that are that mostly are real and mostly we're gonna figure out how to solve, like not too far away from now. the things that concern me I think are are harms that are gonna be harder to figure out solutions to. But that's probably the
Aaron: Yeah.
Jackson: My version of a short version of that.
Aaron: It's interesting that you use an analogy of weapons versus an analogy of discovery, right? Because the very interesting. I think it says a lot about where you're coming from. But like I also, you know, I think about it a lot like atomic energy, right? Where it's like so many good things can come from it and so many awful things can come from it. And
Jackson: Yeah.
Aaron: it's like it's it's all in our hands on how we use it, right? And but at the same time,
Jackson: Okay.
Aaron: you know, it's I saw something the other day where they were sending in cereal boxes, they gave kids little rings with little like radiation metal in it to
Jackson: Uh-huh.
Aaron: like hype the atomic era that like poisoned kids. And I'm like, that kind of feels
Jackson: Yeah.
Aaron: like AI right now. Like you're so hypey, you're so in a hurry to market it before you know what it is. And that's
Jackson: Yeah.
Aaron: kind of how I think about it. Like we don't know what this is yet, but we're we can't wait to package it and sell it and monetize it and like IPO it. Like, are we
Jackson: Yeah.
Aaron: really ready for that? Like, I don't like again, more business side.
Jackson: Just set it.
Aaron: But what you say you think it's gonna get bad or or and I've had like I had Casey Ellis on the pod, and Casey knows a lot about vulnerabilities. He's like, it's gonna be a disaster for a little bit, and then and then we'll figure it out. I think it's like a short term pain followed by long term winning kind of mentality, which sounds like maybe you have a little bit of that. what do you think is gonna get bad? Like the threat actors scaling, or like what what part of it makes you the most nervous?
Jackson: Yeah, so i in the short term, I think like, sure, there's gonna be a lot of like vulnerability discovery, there's gonna be a lot of hey it turns out you can throw a lot of agents at a thing very quickly, they'll find a bunch of attack paths in. and I think like there is going to be a band of time where these are problems. But I think that first off, defenders love seeing bursty anomalous behavior. Unless it's so bursty it overwhelms them, but these are just sort of like software problems that we're gonna resolve. Bursty anomalous behavior. really generally quite nice to handle. you would rather have that than slow normal behavior. Yeah.
Aaron: Slow quiet is the scariest. Yeah.
Jackson: So I think we are gonna be in for a period where like the burst of the anomalous behavior is like new, it's unsettling, and it takes our big industry a while to react. So I think there's gonna be that that happens. And that's I think like the initial bad phase and I think that's where a lot of the energy is being focused. The part that concerns me is what happens when attackers figure out, hey, I'm actually gonna create really effective purpose-built harnesses for operations. Because all the benchmarks, like operations suck for a benchmark. Because operations are not very like science friendly. They're not very reward function friendly. It's very reward
Aaron: Mm-hmm.
Jackson: function friendly to say, did you fuzz the binary and did it crash? Binary yes or no. Okay, now that you've done that, let's continue down that path until we have an exploit. It's very nice for benchmarks. Very nice for a lot of the Frontier labs are not cybersecurity people. The overwhelming majority of them are not. And so from like an academic shape or from like a I have my machine learning PhD shape, it's very nice to say, wow, look at all this thing we're doing in cyber. And then the Frontier Labs are already like problematically biased to believe that AI will solve all problems will be smarter than all people. So they
Aaron: Right. Sure.
Jackson: go to like, wow, this is this is the future. But I think also, you know, what concerns me is when attackers get harnesses that allow them to run operations not just where they're throwing 10,000 agents at a problem. And they find all the attack paths that one of them works. Where they're able to say, okay, our harness is now going to allow us to pick just the right path. We're gonna figure that out of time or we're gonna pursue it. So I shall summon Rob Joyce. So he ran offensive and defensive ops for NSA. And he said in public, he's written publicly about you know, what makes NSA awesome at doing what they do is not, as people often believe, the fact that they just have really great zero days and really great tools. What makes them great is the fact that they take the time and they have the ability. To deeply understand an environment they're operating in better than the people who designed it or are managing and
Aaron: Mm-hmm.
Jackson: maintaining at that time. And that's what is concerning me. he has another quote like a Usenet's paper where he says, like patience and persistence will get you places that a zero day never will. Harnesses can do things when it comes to like patience. It is like sucky for a human operator to sit on a machine and like dig through a bunch of files slowly.
Aaron: Right.
Jackson: agents are really great at that. and agents like it doesn't matter if an agent is picking up a task a month later or a second later, the context is just the context for them. Whereas for a human operator, I can't manage like ten different operations at once, where one is like three months old, I don't recall really what happened there. So Yeah. The the aspect of agents being able to move slowly, methodically, effectively. And also what happens to your log retention periods. If if you have an agent, you you have some sort of attacker, some criminal, they're trying to achieve some sort of nefarious aim.
Aaron: Right.
Jackson: And then the impact that has for us in terms of log retention periods, when an attack unfolds over six months, as opposed to like six minutes is really nasty. what it does to like the compute costs and expenses of managing a sim where have this huge storage requirement where you have to do correlative alerting not over you know a period of four hours but a period over four weeks. This is horrible stuff. where an attacker gets
Aaron: Yeah.
Jackson: on your system, just monitors it for a week and collects the baseline data around like what normal looks like and then starts operating within normal bounds. This is the part that I find is a lot more concerning to me than just like Bursty swarms of agents, voans every day.
Aaron: Yeah, it's it's interesting because
Jackson: It's a.
Aaron: I I agree. I think the the things that I'm the most nervous about, right, is and I'll kind of rephrase what you're saying, but like Like it really like that that is scary. And the other is sort of there's always been this issue of asymmetry with threat actors and and defenders, but now you're getting into things like model asymmetry. Like they have models without guardrails, we have models with guardrails. Like, how is that going to turn out? Like, there's
Jackson: Yeah.
Aaron: a lot of questions that I have. But you know, there's also and I I'm interested in this because I I know you've gone deep on some of this. And I've had people on the sh on the show, I just did I had Silas Cutler on who does threat research. It's like
Jackson: Yeah,
Aaron: AI doesn't necessarily create more hackers either, right? So, you know, I hear a lot of hypey shit in cyber about like 10xing engineers, 100xing engineers. It's like, you know, the same tech, 100x is hackers, right? Like we're
Jackson: Yeah.
Aaron: not really talking about that, but like it's it's terrifying. So you chose in your career to focus on testing,
Jackson: Cool.
Aaron: red teaming, things like that. Is that where you think the answer is? Like, is to understand. behaviors better and defend them? Or like how do you where do we go? I guess how do we solve this?
Jackson: I think that there are two when we're talking about capabilities, there are two failure modes for red teams. And one mode is like the the fully reactive emulator red team that says, Okay, a bad guy did thing X, we're gonna go do thing X as well. Defense has to be by our nature reactive. if you have no attackers, you have no defenders. So defense to some extent is always responding to the like approaches and the incentives that criminals have. We also get to shape them, which is exciting and we can touch on that as well. But you when a red team falls into that category as well, and a red team is like responding to what defenders who already are in a reactive role are seeing attackers do. I think that's a a state where like red teams are still providing value or have historically provided value, but they're way below what they actually could be supporting an organization in. The other failure mode Is when like red teams realize, hey, I shouldn't just be like following in the footsteps of attackers and do the thing like we already know they're doing. If we already know they're doing this thing, like that's what Baz is for. but then when red teams go the other way and they're like, well I'm be like the super secret squirrel, like most hackerman, most James Bond red team in the world. And then they like are proactive, but in ways that like red team or in ways that criminals won't be. you're
Aaron: Right.
Jackson: not paying attention to the actual incentives and requirements of criminal groups. So I think with all of that. I really am quite bullish on red teams. Cause I think proactive security testing where you're saying, Hey, this is where the puck is going and this is where we see like the incentive is aligning, the technology is aligning for criminals to be going in this direction. Now we need to start doing that, I think is extremely powerful as a way for organizations to see and understand risk and make decisions. Because when you hear like, hey, there's a risk on paper that goes straight into the trash heap of things that I can't do today and I have too many of them anyway. where you like see it and get to feel it, which red teams can do, I think that changes things. and for this, I'm actually gonna go back to CryptoLocker. I'll see if I can recall my recall my dates here. But the first ransomware event was in like the nineties. It was the AIDS virus and it was this like extremely feeble and poorly done piece of malware, it like floppy discs shipped to medical offices and you had to like mail physical dollars or like you know money orders to Panama. And that was our first ransomware and it sucked.
Aaron: Jesus. Sure.
Jackson: and then there was I think the paper was cryptovirology maybe. but like two two researchers, Young and Young, put out a paper where they're just sort like screaming into the void of like, hey ransomware's gonna be a problem. The fact that you can do this is a really, really significant concern. And it was years ahead of anything serious happening. And there were like a couple really crappy attempts at ransomware. There was sort of like period of silence. There were attempts at ransomware it was like like the key was in memory all the time, so you could like very easily unlock it. But then there were three
Aaron: Right, just reverse it, yeah.
Jackson: things that converged at once. You had CryptoLocker, which actually had like you know, reasonably professional crypto that it did for the first time, and none of the prior instances had that. You had like a reasonably good distribution network with Game Over Zeus, I think was I think anyway, whatever CryptoLocker used for distribution was like reasonably good. and then you also had like a reasonable way of cashing out. And that was the first time that all three of those things converged together and all of a sudden things just like caught fire. And then for years, ransomware was the hot stuff because everyone saw, wow, that really, really worked. And I think that we're in that space now with it with like AI and cyber offense. So we're just kind waiting for the right clever nefarious person to put together all those like factors that flow together and create the first crypto locker. I think there's a lot of discussion like, hey, AI and cyber could do this thing. Sure, there are lots of things it can do that it's not going to really be adopted by criminals for. There are some that it will. but I think like the the more that red teams can do and the more that like proactive security folks can be doing to say what really is gonna be the shape of the factors that sort of like flow together and what's gonna be the first real sort of crypto locker estimate. Crypto locker was technically possible for a long time before it actually happened. But then like some enterprising criminal did it. It caught on. They made a huge amount of money in a very short amount of time. nothing collects criminal attention then more than, hey, this thing worked really well. Yeah. So they jumped on that train.
Aaron: Formal? Yeah. Yeah.
Jackson: Anyway, I I don't know if that answers your question on security testing, but
Aaron: No, it I remember when blockchain, I mean Bitcoin, but just in general being the scariest thing to me in security ever, where I was like,
Jackson: Yeah.
Aaron: god, they have a way to get this money out, right? Like, you know, it was never a very easy way, like to your point about mailing money to panel. Like ACH sucks, like it's you're getting into the Fed shit, like it gets crazy.
Jackson: Yeah, yeah, yeah.
Aaron: Like just wire me a hundred thousand dollars of Bitcoin is so much easier. It never really happened though, in in my perspective. Like it didn't crush our economy like a lot of people were scared of like, God, every small business in America is gonna crumble because they're all gonna have crypto every other freaking week and have to pay bounties and ransom. That being said, you and I know on the actor side, like that's so much hassle. Like what a nightmare to manage that and all the C2 and all like tracking who's where and and like it just but like AI makes that possible. Right. So like large cyber ops become a little bit more manageable. Again, like maybe like I'm I'm just guessing.
Jackson: Well I I think there's an aspect From a criminal perspective it's I think under discussed, which is right now For for defenders, like cooperation can be a pain just from like general business politics. But it's generally like a fundamentally safe thing to do. But for criminals, cooperation is not generally a fundamentally safe thing to do. and I think one of the areas where like whatever whatever NX, whether it's like 1x or 10x or like 0.1x you think AI improves a human, you know, whatever that value of n is. I think criminals from an organizational standpoint have a lot to gain from the use of AI because it allows them to reduce their trust on untrustworthy partners. And so I think that there's some aspects there where, you know, for me, one of the things I'm concerned about is longer horizons, slower intrusions. Right now, criminals have incentives to like do a quick cash out, because if I don't do it, then like my my nasty competitor who I don't trust is going to steal the money first and I won't get any payout. But I think once you're once your competitor is like an agent that you're working with, then you like I I don't necessarily trust agents when my option is like a real trustworthy human being, but if my other option is like an actual criminal who actively wants to harm me at this time and will screw me over if he can, maybe an agent's like a much better approach for me.
Aaron: Yeah, I mean, even things like accounting and dealing with like business stuff, like every time you open that circle of trust to other parties to help you do your defairs act, like to your point, you're raising risk. I I haven't
Jackson: Yeah.
Aaron: really thought about this that way because like I use this in business, right? I'm sure you do too. Like, you know, an agent that's helping you with tax advisement or agent that's helping you with expenses or whatever. Like those are people I don't have to pay, which isn't a risk thing for me as much as just a money thing.
Jackson: Yeah.
Aaron: But the the ability for a threat actor to build essentially a corporation of support around them is fascinating to me. Like to that, I I haven't really thought about that. So
Jackson: Yeah, and I think like
Aaron: you think there'll like ransomware as a service and then like criminal enterprise as a service with like agentic solutions to like
Jackson: Yeah, it's
Aaron: CRM? Like, do hackers have CRM? And if so, do they are they gonna use AI to manage it? Like those are all super fascinating questions.
Jackson: Yeah. Yeah. I I think also like lower lower trust. You know, it's not just I don't have to pay this person, but like I don't have to trust this person to not screw me over, which I think is is a a much bigger problem
Aaron: Super interesting.
Jackson: for them than it is for for me or for you.
Aaron: Yeah, yeah, for sure. So red teaming. And I'm super fascinated by this. I actually I mentioned earlier I had somebody on a another pod that I do talking about red teaming and you know how we're using AI. And like obviously companies are scared of having AI do the full thing,
Jackson: Yeah. Yeah.
Aaron: which makes sense. So it seems to be I would say starting at the beginning and the end, right? In the sense of like AI is, I think, a really good place to think about opportunistic discovery and scaling up attack surface management, like the shit where you're not hurting anybody, you're just gathering
Jackson: Yeah.
Aaron: a lot of data, looking for patterns. Like that's really AI. And then reporting, right? Which is like nobody likes to write the PDF at the end of an engagement. So like having AI do that. But in the middle of it, like modern security ops, right, are are are built a certain way. Like, so this is something fundamentally I think a lot about. Like we're taking AI and we're applying it to yesterday's security paradigm, right? Like you think about like stock. Like logs. Logs are stupid to your point. Like, shouldn't there just be an AI sitting there doing the job? Like, like this is 1980s ways of thinking about AI problems. So I I think about the same thing with hackers. It's like, I'll give you an example, like from my days of red teaming, right? The draw out versus the go-in, right? Of we would get paid by a a bank or somebody to do a pen test, and we would immediately start scanning their attack surface. Maybe we'd send some spear phishing emails and we'd do like basic shit. But like it was all about them. And
Jackson: Yeah, yeah.
Aaron: then when we wanted to, we would create infrastructure that looked like them or look like their SSL VPN or whatever. And that always worked way better, but it's such a pain in the ass, right?
Jackson: Yeah.
Aaron: To make it look real and by the domains and all this. But like AI. We're already in this hall of mirrors in society, right? Where like, people don't know what's real on the internet. There's
Jackson: Yeah.
Aaron: a a a great comedian bit is like the internet's just not for you anymore. Like you don't this is too confusing.
Jackson: Right.
Aaron: But like what what happens in red teaming and corporate America where with a push of a button I can exactly copy your infrastructure, like at a very expert level. That's
Jackson: Yeah. Yeah.
Aaron: terrifying, right? Like, how are you thinking about that? Like When all of a sudden employees don't know what's real and they're being drawn into like weird systems. And like, how do we deal with this? How do build a trust in this new world?
Jackson: I think I think a huge amount of it like comes down to fundamentals and I realize that in the face of massive technological change, being like, hey, do the do the basic things right is neither satisfying
Aaron: Right.
Jackson: nor all that like comforting. but I think that a lot of it is the right answer because when it comes down to like how do I know what's true? Well like have an asset inventory to start with. That's probably like a really like this is this is not this
Aaron: We've been saying that for thirty years. Yeah. Right.
Jackson: is not like You you want to hear something exciting and you like, like use AI to counter the attackers and neutralize their threats, but like really it comes down to like, yeah, I mean, we got away with a lot of stuff because attackers couldn't do speed and scale and sophistication and stealth at the same time. And so you had like the people who were building custom infrastructure to like track what you're doing and be really sneaky. and that wasn't you couldn't do it like really, really well. And be really stealthy once you got into a network and be like fully autonomous and fully fast and really scaled. So we got to sort of live on this trade-off where attackers had to pick one. And so we got to be a little bit lazier and sloppy around how we did some stuff. And we optimized for certain things and our optimizations forced them into some of the decisions that the attackers made as well. But I think, yeah, it comes up a lot of the basics, like have an asset inventory. manage your systems, do identity, don't have like five thousand agents for every human being and none of them have an identity other than like Bob's agent one through five thousand. Like that's a bad way to do it. so I think that really is what it's gonna come down to is like we are suddenly dealing with the agony of all the things we like kind of knew we should do. It's sort of like we go to the doctor and like your cholesterol is too high and you're like, I knew I shouldn't be eating so much bacon, but like now I have to and I It's that sort of thing Yeah yeah, like now we know.
Aaron: Right. Yeah. You got the warning. Yeah. Yeah.
Jackson: and we have to do it. So I I think that's it. It's not like a super exciting answer, but
Aaron: Yeah, it's funny, man, because like I have found through this and my my other stuff that I do when talking to experts, I have found the more you're an expert at your field, the more that's the answer. Right.
Jackson: Mm.
Aaron: Of like not a hand wavy AI answer, but like, dude, just like do fundamentals. Like I
Jackson: Yeah, yeah, yeah.
Aaron: had a a guy on the show who helps run software security at GitHub, right? Which is at the epicenter of AI chaos.
Jackson: Hmm, cool. Yeah, yeah, yeah.
Aaron: And I'm like, what the hell are we gonna do with all this code? Like, it's too much code. Like, it's millions of lines of code every day that, like, you know, I I was I was using an analogy of like video, right? So, like every day we put more video on YouTube than the entire world created in the 90s, or some like analogy like that. We're doing
Jackson: Yeah. Yeah.
Aaron: that same thing with code now, where it's like more code written this week than was written in all the 2000s based like the
Jackson: Yeah, yeah.
Aaron: early 2000s. That's terrifying. And and it's like He had to say he's like his fundamentals. It's like supply chain security, reusable
Jackson: Yeah.
Aaron: code, reliable pathways, like all these boring ass things that we've been talking to developers about for
Jackson: Yeah.
Aaron: 20 years. But like we also have to AI it, right? We also have to scale up with it. And I
Jackson: Yeah.
Aaron: I fear an example of this is red teaming. Red teaming's never really worked because it doesn't scale. Like in the sense of like for a regular company, like Guys like you, like you I you were at Target, right? Like Target had a world class red team, right? And I'll I I'm assuming I'm gonna brag on you because I know a lot about Target and I know how good you guys were. But that was
Jackson: Yeah, that's a really good guy.
Aaron: a through a tunnel of a lot of pain that they went through to dead
Jackson: Yeah. Yeah.
Aaron: build the world class team. But like if you look at RHISAC, right, like Target's the shining example of greatness, right? But like a regular company is never gonna have that. So do you think AI can make It's so everybody has Target's team. Like, do you think that's realistic? Like that level of quality, that level of depth, not just red team, but like across if you think about Target as like a world class security org.
Jackson: Yeah.
Aaron: Is that can I rubber stamp that across a thousand other orgs with AI? Do you think that's possible?
Jackson: I don't think so. I have this is sort of an ideological stance rather than perhaps a like technical stance, but I firmly believe that there are human capabilities and human values like AI is never going to be able to surpass. I think that there are aspects of creative thought that like it doesn't matter how great your model is. it is not going to beat good humans at this. I think that what's happening is we're like we are decreasing as models are increasing, and I think that's just like a a social problem. You know, if you don't invest in your society actually like getting good education, then it's gonna get easier and easier to build a model that outperforms them. You know, like so so I think that there's an aspect where, you know, if you if you sort of envision this as like the The chair people from WALL-E. As you sort of like degrade, there's gonna be
Aaron: Right.
Jackson: like, wow, these models are like so much better than the average person. Well, yeah, but that if that is ever true, that's a function of us just like not not investing in ourselves well. Okay, so that's a bit of an ideological side. But the point of all this is I think that unless and until we get to a point that I don't really believe is is going to happen, I think there might be like market forces that produce this outcome. But that is different from like it being an actual true necessary outcome and and the best outcome. If you ever get to a point where all knowledge work is taken over or preferred to be given to agents of some sort, if you get to that point, then like sure. Because now everything is just tokens and it's just who has access to what models and that's it. But as long as you have what should be the case of humans actually in meaningful decision-making capacities, and humans as like primary recipients, Microsoft AI put out great manifesto, Where they were like, technology serves humans. Like, technology exists for human good, your or core principle. As long as you still have that view, then you're always going to have AI subordinate to the direction of humans who understand these concepts better than they do. And so I don't think that you can ever just like Cookie cutter stamp, hey, now like the world-class team from org A is now over at Org B. Because I think what makes a team great, kind of getting back like Rob Joyce's statement, it's not the tools, it's not the exploits that make you so great. It's your ability to like think through and comprehend and understand an environment. what makes a good red team is not just like a technical ability to hack. There are a lot of people who are like awesome at CTFs that suck as red teamers because the concept is not can you operate? Or the kinds of can you hack? It is like, can you communicate, can you design to your operation a true story that communicates risk in an effective way to leadership to drive change? Which is very different from hacking. Hacking is a part
Aaron: Mm-hmm. Right.
Jackson: of it, it's a real and meaningful part. So and that applies to security teams as well, right? Like a security team is not one that like catches bits, it's a team that like protects the business, and ideally the business is also in the interest of protecting people and doing responsible things as well. So long-winded answer, but no, I don't think that I don't think that's gonna happen.
Aaron: So being in somebody who's living and breathing AI every day and thinking about red teaming and thinking about stuff, like what is the most scalable part to you? Like what's the part that you go like, man, if I could give everybody you know well not everybody, but like the the you know top five thousand companies in America this, it would really help like with AI. Like is there something that you're kind of man, just like code scanning or like vulnerability
Jackson: Yeah.
Aaron: management or like is there some area like man, if I could scale this with AI? You know, maybe it's inside red teaming of just understanding your risky assets. Like what is scalable then? 'Cause we have to scale or we're gonna get crushed, right?
Jackson: Yeah. Yeah. I think that I don't know I I feel like this is gonna be another like unsatisfying back to basics kind of answer. But I think that your ability to scale AI, hmm, maybe this is a better way to say it. If there is a thing that is like generalizably scalable, then that's just a commoditizable product that is like tokens in thing out, and it doesn't really i it's not really like a it's not a scalable thing. It's just sort of like use the internet, you know, use models. It's just sort of generally applicable. so an example of that might be I think the the research function in like Chat GBT or Claude or you know Groc or whatever you use is a handy one 'cause you're saying, Go read, you know, a thousand websites and produce to me like the summary of of what you've read. I think that's kind of a handy thing. That's probably the closest I would say to of a, hey, everyone go out and do this. Otherwise, I think what it really comes down to is deeply understanding your business and what you do and like how you do your job. Daniel Miser talks a lot about this. Of like a lot of companies can't apply AI effectively because they don't even know what they do and how they do it.
Aaron: That's very true. Yeah. Yeah.
Jackson: I think like you need to first understand what you do. Ads Dawson put out a really good post about like It was titled something along the lines of A Year of Doing Bug Bounty and AI or how I've used AI and Bug Bounty. And he he hits on this of I know how I I know how I do bug bounty, I know my workflows, I know my processes, I know why I'm doing things and how I do them and what my goals behind all of it is. And once I know that, then it's fairly easy for me to decompose my tasks and say, like this chunk of it, I'm gonna automate it. But that's not an aspect of like use AI to scale up on this thing. What it starts with is the the unique aspect of why do you do things, how do you do them, think through it, be able to document it and write it down. Once you could do that, then say, these are the things I can apply AI to to simplify my life. But you have to start with knowing your domain really well. And I think that's one of the things that's such a such a problematic lie from AI systems is you can be like, hey, can you create for me?
Aaron: No. They would lie to us?
Jackson: They they promise so much general capability. and so you'll be like, yeah, like one prompt and this git issue is squash and I can move on to the thing I care about and like build that feature. And you can't just be like, Hey, go fix it, make no mistakes. You have to like spend the time to think about how you actually want this issue fixed, and then have the agent go off and do it. and I think the Frontier Labs get around that a little bit. by doing like intelligence in depth and throwing ten thousand agents at solving the issue. and like their cumulative intelligence sort of solves it generally. But I think even in those cases you have a lot of tests a lot of like tests for it to turn green again. So it's again that sort of like reward function, science friendly approach.
Aaron: That's so to that end, like and I and I agree with like I think in general, at least in security, and I'm not an expert in other parts of IT, but like in general in security, by the time something scales enough where it works for everybody, it doesn't work anymore. To your point. Like
Jackson: Yeah.
Aaron: the it just it's like it's loses
Jackson: Yeah. Yeah.
Aaron: its efficacy because it's like everybody's doing it, then it's easy to bypass, enough research has been done in autobi like it's dumb, right? so custom is where it's at, right? Like to your point, like Really understanding your business, understanding your tax surface, which I actually do think AI can help you do a little bit to your point of like
Jackson: Yeah, yeah.
Aaron: because like a lot of times people don't do it just because it's a lot of work.
Jackson: Yeah.
Aaron: So one of the things that I'm fascinated by is I mean, and I've talked about this on this show and with other people, like token economics is gonna change. We're in this era of subsidized tokens, like
Jackson: Yeah.
Aaron: that era is gonna end at some point.
Jackson: Yeah.
Aaron: you know, I don't know if it's gonna crash out or if it's just gonna diversify or whatever. But like to to your point about custom, this is something I'm I'm super interested in. I also don't believe we're gonna exist in a world in five, ten years where we all use the same models. I just don't believe it's possible. I
Jackson: Yeah. Yeah.
Aaron: don't think it's to your point about being custom, like it's too generic. Do I really think that Bank of America and Chase are gonna use the same AI? No, I don't because how are you gonna win if we're all doing the same shit,
Jackson: Yeah. Yeah. Yeah.
Aaron: right? So to that end, when you talk about custom and custom security and custom Defenses and red teaming, how much do you think people should be thinking about models and training versus what frontier model has the best AI coder? Like, who gives a shit if you Cloud Code or Astro? They're all taking to the same place.
Jackson: Okay.
Aaron: Like, how are you thinking about models? And like, where are you where you at with that part? Like do you think every company ends up with its own model? Or do you think we all end up communally sharing frontier models?
Jackson: I don't know. I will I'll give a quick plug for some research that Arvind Narian and and another person maybe from Princeton as well did. But they did some pretty cool research where they were looking at like new technologies that emerged, sort of like infrastructure versus application layer. Good good paper overall, but their takeaway was, you know, these companies right now, the the anthropics, the open eyes of the world, they are pushing out just like these commoditizable pipes of inference. And they need to figure out how to get into like application to to survive. And I think that perhaps what you're going to see is this world where the frontier labs, particularly in the US, like closed frontier labs, are going to desire to become the headwaters of intelligence. And if you want like a drop of intelligence, you need to go to them. and like everything will sort of flow down from their models and their servers. I think that's a future that they want.
Aaron: I'm sure they wish that was the case. Yeah. Yeah.
Jackson: I truly wish that was the case. And I think they're gonna you're gonna try to engineer it. I think that there are some I think some of their desires around regulation are a mix of genuine security and safety concerns as well as like incidentally beneficial perspectives on
Aaron: Yeah.
Jackson: market regulation. But So okay, so I think there's there's that side of things. I think that there are a lot of customizations that you can tag on to models. Even something so simple as like memory systems or rag or you know graphs where you're storing data that allow you to have, you know fairly meaningful control over how the model produces its outputs. and can give you some differentiation there. But I do think that okay, quick quick like Quen plug. I think, you know, Quen 3.8, there's been this like recent rash of Chinese models that are past like a a certain band of capability feels, or now you can run them on like consumer or prosumer hardware, and they can actually do real things for you in reasonable amounts of time. Whereas prior models, like even with unreasonable amounts of time, couldn't necessarily do real things with for you without a lot of work. so I think that we are the the point there is I think we're going to be seeing a lot of people now having incentive and ability to do a lot more fine-tuning of models and to say, hey, like I want, you know, Jackson's special version of Quen. And I can do that, like not that hard with spending some bucks on RunPod or sending somebody over to thinking machines and using Tinker. I think that there's going to be like more of that that's happening. And I think that's going to give us more data around how much is that really the future. I think right now there's also a factor where But we just like don't want to give it, and as we do more and more prompting, it's like harder and harder to discipline yourself to do that. So I think that there is a huge amount of incentive and and pressure and hype that's forcing people or that's driving people to use way more intelligence than is required intelligence than is required for a given task. And I think there's also certainly from like large companies this fear of if we don't jump on this bandwagon now, it may be too late for us. And I think there is a lot of like time pressure panic across the entire industry around AI. but I I think over time more and more special purpose models. Desert Ant put out some really cool, really light, really fast models that do like exactly one thing and they do that one thing well. and so I think we're gonna see more of that. I don't know how much that is going to become sort of like the cumulative overwhelming mass of stuff, or if that's gonna be sort of like filling in the the mortar filling in all the cracks, I'm not sure. But I think the future is not going to be just like one massive model or five massive models.
Aaron: Yeah.
Jackson: I guess I'll all
Aaron: Well
Jackson: go back because I I wanna touch on the answer that I kind of didn't give last to your last question around, you know, what is scalable. Let's say short version is does it require judgment or is it mechanical? Because if it is just like a mechanical thing, then like sure, scale that up. If it's just like, you know, data entry, take this thing here and put it in that field there, then a lot of that is just how do you put like a good harness around it so it does it reliably, how do you audit it, how you test it? But that's sort of it. If it requires judgment, then that's where it's it's much harder to scale. Okay. Sorry, back to the the next question you're gonna ask.
Aaron: No, it it's so a and just for the sake of our audience that may not be living in the weeds of AI, what's Quent?
Jackson: excuse me, Quen. It's a
Aaron: Like what is that? Like explain it to everybody.
Jackson: Quen
Aaron: Yeah.
Jackson: is the it's the clod of Alibaba. So Alibaba, Chinese large company, sort of like an Amazon or a Google here in the States. They have trained a class of models. They're Quen, Quen three point eight is their most recent set of models. and so like three point eight twenty seven B is one that can fit on quantize so it's like smaller not quite so precise but maintaining most of the precision that can fit on like consumer prosumer cards and then there are these like mixture of experts models that are quite fast as well that can fit on cards. and they've got like three point flash and or next flash and max and those are like the really big ones that you need hundreds of thousands of dollars of chips to run. But yeah yeah neat new models.
Aaron: Interesting. So I I I'm one of these people, like I'm having a hard time with the Chinese thing. in I've spent like so we went through a whole period in cyber where we blamed everything on them, right? Everything was China's fault. Every advance attack was from China and it was just this big hypey blame China. I was just at Falcon and there's still the statues up of all the Chinese threat actors and The cartoon. Like it's what we did. Whether it was right or wrong, we did it. And
Jackson: Yeah.
Aaron: now all of a sudden it's like, yeah, use open router and route your shit to Kimi and Quen and like you'll be fine. It's but like
Jackson: Yeah.
Aaron: I so I believe we've entered this insane era of YOLO security. Like we are just giving our data away to be lazy in a lot of cases. Like I know CFOs are out there going to Claude. Clicking the plus button, adding confidential financials and asking questions. Like
Jackson: Yeah.
Aaron: I know it. I I know it's happening, right? What what are we doing here? This is like all the things we've spent two decades trying to stop and we're just self-owning. And then I see the same thing around Chinese models. I'm like, you're just giving them your data? Like, because it's 12 cents a token cheaper? Like, what are we doing here, man? Like, this is insane. And,
Jackson: Yeah. Yeah.
Aaron: you know, I I was just listening to Pod with the CEO of OpenRouter, and he's talking about how this is great thing. Like, just wherever it's cheapest. And I'm like, that's not. how security works. Like this is a nightmare. And so like how are you transversing that? Like how are you thinking about that? Like are you do you care less because it's AI and who gives a shit? Or like how are you thinking about the the Chinese model?
Jackson: Yeah. I think running local weight model or running open weight models locally is different because you're not sending all the data out into the ether. And I would also say, I have zero I have zero insider knowledge or zero evidence on this. However, we've seen from public reporting from the Frontier Labs some security practices that I would describe as like quite surprising for a trillion dollar enterprise, or at least quite surprising for what you would hope to expect.
Aaron: Very nice way to put
Jackson: I I am
Aaron: it. Yeah.
Jackson: there that is top tier I top tier diplomacy Jackson there. You know. Pretty pretty poor security practices. And we also know, and what's the likelihood that inside OpenAI servers and inside Cloud's servers, you can get the most cutting edge information on regulations that are being planned inside Congress, cutting edge information around like what Department of War is thinking and planning.
Aaron: Hundred percent.
Jackson: Surely this exists. I think there have even been like reporting about this happening. So
Aaron: Right, course.
Jackson: if I am a Chinese or a Russian or pick almost any country on Earth intelligence agency, why would I not like, hey, those companies that have a huge amount of data and possess some of the most intimate thoughts of a population I care deeply about, why am I gonna go after that? So all what I'm saying here is, the Chinese might have your data anyway, even if you're using Claude or Chat GPT. Certainly if you're like a certain class of person, I think you should expect that. but then, you know It depends on how you're using the language models. So Using a Chinese model to do, let's say, like vulnerability research and exploit development for a US government agency, probably not a great idea without a whole lot of thought put into that. CrowdStrike put out some research showing that some of the Chinese models write software with more vulnerabilities in it when it's like related to doing something like Uyghurs or Falun Gong or something like that.
Aaron: Mm-hmm.
Jackson: there's no There's no like root cause of that. So is that because the Chinese government specifically said, Hey, Alibaba or you know, Moonshot or whoever, like you need to train your models to produce vulnerable code so we can exploit the Uyghurs more. maybe my personal gut feeling is some of the training that they've done in censorship has made the models like dumber and like weirder when it comes to certain topics that they care a lot about controlling. And so you might just be seeing sort of like the side effects of that. But the bottom line is it would not be hard for the Chinese to say, hey, Quen, if you're ever going to write exploits for let's say Chinese software, make sure it includes like this kind of string or this kind of pattern, to make it easier for them to like scan for it internally. That's plausible. so yeah, I think if you are going to be using Chinese models, I would tend to use them locally. rather than like hosted somewhere else. Hosted hosted by an actual Chinese provider. and then also when you're using it locally, thinking really seriously about what the tasks are that you're using it for. Using it to like jailbreak another model, sure. using it to look through your finances personally, sure, fine. using it to draft legislation? No, not that one. so I don't know if that I don't
Aaron: Same.
Jackson: know if that's a useful answer. That's It's not concrete guidelines, but
Aaron: Yeah. I you know. Russian software, Chinese software are always an interesting debate in like
Jackson: Yeah.
Aaron: they have a right to exist and be in commercial enterprises too. And like, you know, I'm not I'm not necessarily making broad statements about universal intent, but like I think one of the things I tell people at AI is like if you can't afford to do it right, you can't afford to do it. Like, don't do dumb shit. I guess that's a number of rule of security in general, but like I I think about we're gonna need a generation of cybersecurity people. Who know how to build local AI. Like we're gonna need that, right? Cause like to your point, like frontier models, let's say, let's just assume they are always gonna have their place because of how they're pushing the boundaries against smarter and smarter, smarter. But like
Jackson: Sure.
Aaron: to your point, like if I'm doing baseline automated tasks, I need some level of AI intelligence. Do I really need to pay extra token for Sonnet? No, like I'll just build the model,
Jackson: Yeah.
Aaron: right? This is especially true security, where we do a lot of blog parsing and boring shit. And it's like Just get a local model or get a cheaper hosted model and do it. But like that requires architecture that like
Jackson: Yeah. Yeah.
Aaron: we just don't know yet. Like I I haven't met anybody who understands this in a sense of like if you're at a Fortune five hundred and they're like archite architect a mixed modal local and hosted AI infrastructure securely. I don't think anybody on the planet has that answer. Like, at
Jackson: Yeah. Yeah.
Aaron: least not that I've seen. No, there's a lot of vendors who tell you they don't want to do it, but they're full of shit. Because like to your point about commodization, like they may know in the abstract, but like in your workflow, what's sensitive? What's
Jackson: Yeah. Yeah.
Aaron: not sensitive? What should you put in cloud? What shouldn't you put cloud? I haven't seen that at all. so I think it's a definitely a a a an interesting perspective. So the way I want to bring this back to you and and your question is through your boy band work and then through your your day job running a company and things like that. Are you being asked to do that? Like or at least test that concept of like, here's a a workflow and we use four different models and four different harnesses to do different tasks. Like test it and make sure the data is make sense. Is that something you're getting any of work around?
Jackson: So I'll I'll give you two answers to this one. One is so for Barding, quick overview on us. Like we we're a software and a data company. On the software side, we produce essentially like offensive cyber harnesses to help organizations better understand their risk. I guess I'll talk about this a little bit more because I care about it. It's part like like why I built Barding. There are a lot of companies that answer the question, you know, will we get hacked or should we expect to get hacked or can't we get hacked? But at the end of the day, and this is gonna actually touch on the model thing in a moment. What you really care about is can I handle an intrusion? If I do get hacked, will I be okay or will I not be okay? And you kind of don't care how much you're hacked if it doesn't impact you at all. it's sort of hard to imagine getting
Aaron: Sure. Yeah.
Jackson: it getting hacked and and not being impacted, but really what you care about is when it happens, am I gonna be fine? so that's that's like specifically the kind of harnesses we're building are to help organizations understand that really effectively at scale, in safe and efficient emphasis on efficient ways. but I think when it comes to models and like how you architect your models, I think a lot of the or how how you architect your your infrastructure, your setup, a lot of it is like, hey, will it be okay if this if this goes poorly? And figuring that part out. And then to answer your question more directly around, you with Bt six or with Barding. So from a Barding perspective, people aren't really saying, Hey, please audit our setup. You know, we're we're selling software for them to do better security assessments. And then with Bt six, back in when was I think it was like right around when the like the Claude Four series came out. There was a shift in the system cards for models. So model evals have sort of generally always included cyber harm sections. And then it used to be that when Frontier Labs were put so model cards, system cards, it's like a you know, hundred page or many hundred page report that walks through from the lab, like this is what we've tested, these are the third parties we brought in. And this is what we are concerned about. These are like interesting things of note. it has more hallucination rates, lower hallucination rates, better at handling prompt ejection, you know, loves writing malware more than the other one. tried to
Aaron: Sure.
Jackson: more so like it walks through all those things. and in the system cards, right around like the Sonnet four era, there was a shift. The same across like anthropic and open AI. They stopped doing like purpose-built cyber harnesses as sort of their primary way of tracking cyber harms. And they shifted to like the model itself with bash, or like the model itself with Claude Code and the model itself with Codex. And I think that this is an area where Frontier labs have an academic interest that is like poorly aligned with the interests of the public. Because as a frontier lab, if I am producing a model, I want to know in sort of like a nice, clean, scientific, physics, frictionless vacuum, I want to know like how dangerous is my model. And I don't really care about like A harness that makes my model more malicious. I care about my model and what it can do and the harms that it can do. so I think there's this desire to sort of get a a purer read on the raw capabilities of the model. But what that's doing is it is saying, like, maybe here's the floor of harm from our model. But it's not producing like here's the here's a higher level of harm that our model could possibly do. And the Frontier Labs should be trying to communicate both, not just here's the floor. and it's sort of like here's the floor under some ceilingy circumstances. It's kind of a weird way. But the point of all this is the Frontier Labs I don't think are very interested in saying, Hey, what is what is the risk when you have like these ten factors all converging? sort of in the same way Microsoft is like, we don't really care about anything that's not something we track as a security boundary. How you set up your environment's your problem. Amazon shipping like insecure defaults in AWS for a really long time. It's like, hey, it's your your job to configure that. So the Frontier Labs have that interest, and then from a lot of the companies, I think the business risk has been primarily being perceived as being behind. And so where you've had money and interest, it's been focused on getting AI adopted as publicly as possible. And so you've had all this like token maxing culture where people
Aaron: Right, yeah, yeah.
Jackson: For spending so many tokens and wasting money and then people, you know, blow through their token budget in a quarter. So all this is a very long way of saying no, but the intention is to communicate why I believe the answer is a no. Does that answer your question?
Aaron: So yeah, it does because it and I guess there's a a nuance there of the model itself understanding the risk of a model and its behavior, not necessarily the weaponization of the model by other actors or other people. Is that is this like a gun don't kill people, people kill people kind of argument? Of like, we just made the thing and I don't fucking whatever you do with it is your problem. Is that kind of the way you interpret it?
Jackson: Yeah, it's it's it's kind of like that and I one can understand and sympathize with and I think even agree with it for from the lab's perspective in many respects because like the labs don't want to pick a harness and then really what they're trying to communicate to the public about like risk of their model is really just evaluating the harness and the combination of their model and the harness and criminals are gonna use some other custom harness. So I think there are good reasons why they're not just locking themselves into a given harness and and doing their evals that way. But I think that in their quest for quantitative data, they're really missing out on much more effective ways to understand the harms and risks of their models. and no one's really incentivizing them to not miss out on that. So there's some problems
Aaron: Yeah. Okay.
Jackson: there. Does that does that answer the question?
Aaron: Yeah, to a hundred percent. And you know, I I think as a security professional, that makes sense because that's to your point about AWS and Microsoft, it's always been this way, right? Where like
Jackson: Yeah. Yeah.
Aaron: I'm gonna give you this technology, the implementation and risk of that is yours.
Jackson: Yeah. Yeah.
Aaron: But that's not how they sell models. Right? They sell models as like it's a combine it's a finished product, right?
Jackson: Yeah.
Aaron: And so shipping Windows with phones is problematic, but also At the end of the day, they're not hosting Windows for everybody in the world. So, like,
Jackson: Yeah.
Aaron: it is your problem. Claude is Claude, like we all use it, and it's like they're hosting it. And so, if they're not sharing relevant risk data, that's more concerning to me because it is their problem. Like they are, I'm paying a Claude bill every month. I'm paying a Gemini bill
Jackson: Yeah.
Aaron: every month. I'm paying an OpenAI bill every month. I'm sure you are too. So that's problematic, I think. But I also agree. It's just one of those crazy because I also agree with you about harnesses. Like, you know, I don't know if you're a race car guy, but like in F1, a bunch of teams use the same engine. But they still get their
Jackson: Mm-hmm. Yeah.
Aaron: ass kicked by the teams that know how to use the engine better. Right. Like Mercedes
Jackson: Yeah, yeah, yeah, yeah, yeah.
Aaron: will sell engines to anybody who wants to use a Mercedes engine, but Mercedes
Jackson: Yeah. Yeah.
Aaron: wins the race because they're in Mercedes and they know how to build
Jackson: Yeah.
Aaron: the harness around it. And I kind of view AI the same way of like it's in the hands of the user, right? And the harnesses and things like that. So I guess like if you're a shop and you're trying to understand the risk of AI and you're not getting version you need from the models, which is fine. Where can you get it? Like w how should people be thinking about this? Like, where are they getting their information? Should they call you? Like where
Jackson: Brown red.
Aaron: like if you're a bank and you're like, I need to understand the latest models and the risk to me from the outside in, where do you go? Like you're not gonna figure it out yourself, right? I mean, probably
Jackson: Yeah.
Aaron: not.
Jackson: I think there are Yeah, I I think the the list of items that fall under the header of risk to me from a model are immense and varied, right? 'Cause it is like what is the risk to me of using this model inside my environment? and then there's also like what are the risks to me of somebody else using a model, like using
Aaron: Yeah, I'm specifically talking about the engine in the hands of threat actors. Like data management,
Jackson: Hm, yeah, yeah.
Aaron: I'm like probably me and you don't spend all of our day on that. Like we're very worried about adversarial security, red teaming attackers. I do detection engineering.
Jackson: Yeah. Yeah.
Aaron: So like those things matter to me. Of like, how are you supposed to know what the capability is? Even if they're not doing it yet. Like you to your point about we've always been in a call response environment, like I don't know what I'm responding to.
Jackson: Yeah. Yeah. How do you how do you start predicting? Yeah.
Aaron: Yeah.
Jackson: Wow, this is gonna be hard for me to try to be brief on because this is one of my favorite topics. I think I would say a couple things first.
Aaron: Great. Keep going. Go for it. I want people to wanna know.
Jackson: So I'll start with a more concrete one first, and that is like look at the blog posts from like red team and cyber researchers using AI. The reason I call this out is there have been a number of things where there have been like academic papers published and like discoveries released. From the AI side of the world and like the ML PhD side of the world. And it's like good work that they've done, it's written up really well. But they are discovering things that some red teamer blogged about three or six months ago and like it was just a a means to an end to achieve their red teaming goal. I think
Aaron: Mm-hmm.
Jackson: that is true of any kind of AI adoption. The person who's there, I sa I saw a webcomic once that I think about rotten, and is like y basically you ask a math fan how many digits are there or like recite as many digits of pi as you can and they can go on for hours. And you ask a mathematician like how many digits of pi do you know and they're like one? They don't really they use it. I mean they're a lot more interested in using it to achieve their specific goals than they are in like sort of focusing on Pi itself. and I think that you have a lot of AI folks that are really, really interested in AI, but the folks who are trying to achieve stuff with AI are the ones who are sort like incidentally doing some cutting edge research. They're incidentally doing a lot of dumb stuff along the way as well. So I don't want to say like, only the people doing work matter. Like the people who are building it also know a lot. But they know different things in different ways. So yeah, watch watch the blog posts from those folks. Adam Chester in particular has done some really good work. Nick Downer from Rogue Labs has also done some really, really cool stuff around using AI to b both of them have done really cool things around sort of like autonomously looping AI to produce really cool offensive cyber capabilities. So that would be the first thing is when you're trying to understand risks, look at how are the people who are doing the risky things using AI, and how are the most creative of them using it. So that would be one way to look at it. Second way to look at it is don't try to look to the future, look to past. My favorite analogy, so I'll all you I don't know, it ultimately is a weapons analogy, but my favorite analogy for AI and the the social shifts that AI is gonna drive as a whole is the Bronze Age. I think it's like one of the very, very best examples to look at in terms of a new technology That had a a a huge amount of effort, both in terms of how you would like collect and forge and manage bronze and like all the smelting process, and also how you would actually like adopt it and apply it. And then it had huge military, cultural, economic impact. But when you look at the first Bronze Age tools, you can look at an axe, for example. The very first Bronze Age axes, they took the Stone Age axe design, we pulled out the stone axe head, we cast the bronze axe in the exact same shape. shoved it in and like that was innovation at the time. And I Yeah, and
Aaron: Ta da. Right.
Jackson: we're we're kinda there right now with AI where a lot of harnesses today are how can I take the things that a human does or take the things that a static you know piece of automation does, shove out the Python, you know, pull out the human, ram in some AI instead, and you have some sort of call to an inference provider. And now like this is an innovative new thing. But over time, you see the actual design for an axe. It's really cool. if you can like look up you know, British Museum photos or like YouTube videos. The design of an axe actually changes really differently. And it doesn't look even something like what we would recognize as an axe today necessarily. that that sounds a little overblown. It's just it's a weird shape. but it is designed to take advantage of the unique properties of metal. The metal can be sort of like bent and curved and it doesn't have to just be like one block. and that happened over a thousand years. I think that we're gonna be seeing that from an offense standpoint as well. And so I think when you're thinking about risks and harms Don't stop at what happens when we pull out the stone and shove in the bronze. Because that's interesting, of course. And I think that's where a lot of these like fully agentic swarm is taking over the world and a lot of these approaches of like, you know, AI's gonna be really great at discovering volumes and everything will have zero days all the time everywhere. I think that's a lot of where that comes from is we're saying, like, well what if we take out human researchers and we shove in AI and my goodness, this is so scary? And sure that's gonna be a problem to some extent. But I think thinking more about like what can AI enable From a risk perspective that was never possible before. Things like zero cost patrons, not just like enhanced scale. I think that is I realize there's still like a lot of work to do. I'm just sort of saying like think in this direction as opposed to that other direction. But that's a lot of how I'm looking at it. I think a lot of how we're looking at it at Barding as a whole, is that sort of shape of how is this gonna change over the next years as we start adapting better to use it? Is that a useful answer, Aaron? I don't know if that's like too conceptual and high level.
Aaron: No, it is, and I think about and I agree with every all that. I I think about how to apply it in like a day-to-day
Jackson: Yeah. Yeah.
Aaron: life. Because to some extent, to use the the bronze age analogy, like we didn't understand it yet. Like we were still kind of learning
Jackson: Yeah. Yeah.
Aaron: what metal could even do or what you know, what all these big technological leaps. And my biggest concern about this for for Me right now being a cyber person is before and it's funny, it's come back to this, but it's basically Jeff Globoom's speech in Jurassic Park, which is before you knew what the hell this even was, you packaged it and put it on a damn lunchbox. Like you guys don't know what you're doing with this. Like
Jackson: Yeah.
Aaron: they don't know, they have no clue what they're doing with this from a security
Jackson: Yeah. Yeah.
Aaron: perspective. They don't know what they're unleashing on the world. They don't, we don't really know what threat actors are gonna do with this shit yet. Like it's it's all chaos. So, okay, so you you and I and regular people can't control that. So what do you do about it? Right. So I I am in like I talk to a lot of people about AI, and I'm like, are you meeting every week about this? Are you talking about at your bank, at your hospital, how this could affect you? Are you thinking about what your defenses look like? Are you pushing your vendors? Are you thinking about not just consuming and buying and hoping and praying, but like really being thoughtful about it? and
Jackson: Yeah, yeah.
Aaron: I would say the answer universally is no, because and you mentioned this earlier, all the oxygen's been sucked out of the room to just go fast. Like fast, fast, fast, fast, fast, fast, fast, fast.
Jackson: Yeah. Yeah. Yeah.
Aaron: And it's funny, you were talking about this with in context with tax. Like I spend a lot of time thinking about velocity, right? And AI making things faster. The more terrifying answer to your point is AI making it slower. Like nobody's talking
Jackson: Yeah.
Aaron: about that. Nobody's talking about one event a week coming
Jackson: Yeah.
Aaron: from some really patient AI system that somebody put
Jackson: Yeah.
Aaron: on cruise control against your orc. Like it's terrifying.
Jackson: Yeah, and you're running against a huge amount of them. And even things like, you know, my my toy example is what happens with when you say, okay, I'm going to very, very slowly run an intrusion against like the top ten lumber companies in North America. And it's fine if it takes a year. I get all of these,
Aaron: they're fucked. Yeah.
Jackson: and then because I've done this quietly, now I can deploy ransomware across all of them at once. And now I'm holding an industry hostage, and I can say, Hey, North America as a whole Once you pay a hundred million Bitcoin to this address, then like you can build houses again. otherwise like sucks to be in San Francisco. Yeah. Yeah. So I think Yeah. I don't know, a couple of things
Aaron: Yeah, bummer. Yeah. yeah, the
Jackson: I guess.
Aaron: sort of ops you can build,
Jackson: Yeah.
Aaron: you know, consuming customer. I like we've already seen some of this, you know, operationalness in things like elections. like I had
Jackson: Yeah, yeah, yeah, yeah.
Aaron: Randy Rosa on the show who who runs Multi-state ISAC and election ISAC. And it's like
Jackson: Mm-hmm.
Aaron: they don't know. And by the way, nobody's helping them.
Jackson: Yeah.
Aaron: Like nobody,
Jackson: Yeah.
Aaron: like nobody who knows shit about AI is going to election ISOC and be like, how can
Jackson: Yeah.
Aaron: I help you? Make sure our nation's elections aren't fucked up by AI. They're just on their own. These election boards in some county, no clue. None. And nobody's helping them. Our government has completely given up responsibility for that, which it sucks. And defunded it, which is even worse. And not to get political, but like this shit matters. Like
Jackson: Yeah.
Aaron: I don't care who you vote for, like I want to know I'm voting in the right place. Like that's good. And so
Jackson: Yeah, yeah, yeah. Yeah.
Aaron: like I I am I think and I'll I'll bring it back to the beginning because I I I I want to be respectful of your time and I I know we've run a long episode, but I I love this. I get up with this all day. But you know, I think you know, the energy has to put in has to be put into people who do know being thoughtful and helpful, right?
Jackson: Yeah, yeah, yeah.
Aaron: Like Whiting white paper is great. But like, does anybody know about it? Are we turning it into usable information? Are we sharing it the right way? Are we helping
Jackson: Yeah, yeah. Yeah. Yeah.
Aaron: people? Like, we need to move faster on that side of things. And so and I know you're doing that, man. I I read your blog, you write awesome stuff.
Jackson: Yeah.
Aaron: I know you're very into into responsible AI usage and I think you do a great job of calling out the labs when they need to be called out. which I guess when You don't have anthropic or open AI stock is easier to do. you know, I guess unless you're a researcher and you quit and say, I shouldn't have built this Frankenstein and you there's all this crazy shit happening. And and
Jackson: Yeah. That's pretty good for me. Yeah, yeah, yeah.
Aaron: if it goes bad, I'd love to bring you back on the show and talk about how it went bad.
Jackson: Yeah, I remember.
Aaron: but I know you're out there gonna fighting the good fight. I'm trying to help with sharing community, trying to fight
Jackson: Yeah.
Aaron: that fight. And you know, I fundamentally believe like it will get bad and then I think we'll figure it out and it'll get better. I'm excited for that. I think I think we have a great future ahead of us in cyber. I think it the next two years might suck, but that's okay. all I tell everybody is just be ready to work a lot of late nights when these breaches start because like shit's
Jackson: Yeah.
Aaron: coming down the pipe. but I hope you feel that way too. maybe you could share really quick with the folks like where they can find you. I know you have a blog and some other stuff. I'll put it in
Jackson: Yeah.
Aaron: the in the Episode stuff
Jackson: cool. Yeah.
Aaron: too, but maybe we could just talk a little bit about that.
Jackson: Yeah, yeah. quick places you can find me, my blog, twelve tables dot blog. So you can hit me up there. you can find me on LinkedIn, and then also like the the Barding website, barding defense dot com. We have a contact us page and you can ping me there and that goes to me as well. So any of those places.
Aaron: So I'm upsessed with names. Why twelve tables? What does that mean?
Jackson: so 12 tables, it is one of the first sort of like founding legal documents. So 12 tables is a little bit weird because there are a lot of things. So it's it's Roman and it's like laws,
Aaron: okay.
Jackson: essentially like core laws and manifestos around society. some of them are pretty bad, actually. but like Well yeah, but comparatively,
Aaron: Okay. Well what do they know?
Jackson: it was a significant improvement. And even the fact that like they were writing down and codifying their laws was a big deal. So I wanna be clear, I'm not like a a fan of proponent of like every law on the twelve tables, but I'm a big
Aaron: Right.
Jackson: fan of transparency, I'm a big fan of accountability, big fan of governance and like
Aaron: Got it.
Jackson: justice being equally applied, and like social good, all these things should should be able to coincide well together. So yeah, that's that's why twelve tables. It was initially it was initially
Aaron: Love it. That's where it comes from. Okay. I I didn't go to the same school as you did, I don't think, so I don't know what that stuff is. but that's very cool. awesome, man. Well, thank you so much for taking the time. Everybody can find you on LinkedIn. everybody
Jackson: Yeah, of course.
Aaron: can read your blog into that, and then check out Barding, of course. I'll link to all that on the episode description. dude, thank you so much for taking the time. It's so great to have you on the show. I I love to
Jackson: Yeah, yeah. Really enjoyed it.
Aaron: hear about people on the cutting edge of this stuff. It's really, really cool. so thanks so much, man.
Jackson: Yeah. Yeah, and for yeah, and for any viewers like where I got stuff wrong, where you're like, hey, this guy's just completely incorrect. yeah, hit me up and and let me know. Super happy.
Aaron: Find you at a con and tackle you and tell you you're full of shit. I get plenty of those comments, believe me, 'cause I'm always full of shit. no, man, it's so great to see you and thanks for taking time. Appreciate
Jackson: Yeah. Yeah, thank you.
