Episode 364 - AI and ZK Auditing with David Wong

Transcript

[Music] Welcome to Zero Knowledge. I'm your host, Anna Rose. In this podcast, we will be exploring the latest in zero knowledge research and the decentralized web, as well as new paradigms that promise to change the way we interact and transact online. [Music] This week I chat with David Wong from ZK security, an auditing firm that is very active in ZK. We chat about the new era of auditing with AI as both an opportunity and an emerging threat. We discuss the evolving tools of the trade, the static and dynamic analyzers, formal

verification techniques, and how these all change with AI. He shares how his operation includes a team of AI agents working to enhance the work of the human auditor. We also cover the difference between auditing ZKbased tech and things like smart contracts, the threat posed when AI gets into the hands of malicious actors, and whether he thinks the role of auditing will soon become entirely replaced with AI. I'm particularly interested in how AI and ZK are intersecting as there are a few ways we know of the broader ZKML category. ZK to prove something about a model. ZK to add privacy to a model. We have ZK for providence to show what is and isn't AI.

And in this episode, we discuss how AI is affecting auditing in the ZK space. I do hope to get a chance to discuss more cases like this in the coming months. Now, before we kick off, I just want to point you towards the ZK jobs board. There you can find job postings from top teams working in CK. And if you're a team looking to hire, you can also post your job there today. We have heard great things from teams who have found their perfect hire through the ZK jobs board and we hope it can help you as well. Find out more over at jobsboard.zeronowledge.fm. You can find this on our website and in the show notes. Now, here is our episode with David Wong.

Today I'm here with David Wong from ZK Security. Welcome back to the show, David. Yeah, thanks for having me uh again one more one more time. Yeah, we looked back at the previous episode. It was all the way back in 2023. Um, I think you had just started CK Security back then. Yeah, you I think you do these every weeks, but for me it's like ages away. So, it was like quite a shock to come back after 2 years. But it's good to have you back on. Um, let's first do a little catch up. I mean, how have you been over the last two years? How has ZK security been going? Like, what kind of projects have come your way? Has the company changed? Yeah, let's talk about that. Yeah, I mean, it's been a while. We've had a lot of changes from the founder uh set changed.

We're still three founders, but there was some some reshuffle reshuffleling. Uh we also grew as a team. We're we're 13/15 people now. So I don't know how many people we were back then, but probably like two two or three max, I think. So yeah. Um so we yeah, we're a bigger team. We grinded for quite a long time, so we we worked on a lot of different projects. I I know we were very focused on ZK but since then we've done some NPC stuff uh threshold signatures we've done consensus protocols we've done these days we're doing some T stuffs uh that that would make an interesting episode so so a lot of things are happening we're more of a real company these days and we're seeing now that there's all these ZKVM happening or ZKVM happening

we're really looking uh forward to that stuff nice I'm wondering like just going back to 2023 have you found the volume of these kinds of projects. Has it gotten more? Like is it has it increased or has it leveled off? Is it about the same as it was? I mean, I guess back then a lot of those teams were also working with like more of the established auditors. So maybe like your business picked up over time because you built up a reputation, but I I I'm sort of trying to get a like a bit of a temperature check from you as to like if you see more today or if you actually see if it's holding steady. People always ask me this question this question and what are the trends in ZK and I I never really know how to answer

that. So basically there's perception right like we perceive ZK like a few years back I don't know how many years now there was like so many ZK stuff happening right like there was a new folding scheme every week there was a new thing every week right re research wise more and these days he feels like it's slower but at the same time I'm getting like like how do you call that like incoming you know calls from potential clients that I've never heard about or so there's all this kind of like iceberg of what we're seeing, but also we're getting a lot of work from companies that I've never heard about and and like smaller companies or like projects here and there. And um we're talking with like we're starting to talk

more and more with like non-blockchain projects like we've seen like Google with the the wallet project and so so there's things happening. It's not necessarily clear what is happening or or what's the amount of things happening. I mean what I can tell you is that we have a lot of work. We're probably we have the most amount of work we've ever had uh since the inception of the company right now. So that's a good sign. I don't know if it's going to stay the same. I I saw that list of ZKVMs that that I mentioned like on Twitter the other day. So I was like, okay, I guess work is not going to stop anytime soon.

I think for me it was this like I think just recently Justin Drake hosted an ETH proofs uh meetup. I don't know if you were there, but that list, I mean, that was amazing to me cuz I knew I mean, I think we knew about 80% of those teams, but the categories that he put them in, just like the way it was designed and the fact that there was a bunch of new players who had joined the ZKVM space was just like, wow, this this l like this corner of them. Oh, I'd have to check like 20 or so. I mean, his list, just looking at it right now, there I think there's about 27 teams on it. He did include some hardware teams. This is a subset of the larger map of ZK. I think the larger map of ZK has around

200 teams, but these are like the infrastructure teams and they're and they're in like subcategories that some of them I I didn't even have on my radar. I don't know. It was pretty cool to see them. Yeah, people are doing all sorts like that. That's what I'm saying, right? A lot of us see the the top of the iceberg, but there's this entire thing submerged of uh apps and even like ZKL1 like there there are L1's that we don't even know use ZK right now and we're discovering that we're like really oh cool nice I mean Solana is starting to use ZK apart from yeah there's actually there's a ZK ecosystem and there's ZK within Solana we did an episode uh like a few months ago with merch and we did a survey and so there yeah there's it's definitely happening

over there just going back to what you had said about research too like from from our side we're doing ZK mesh we do this every month um that's just like a list of links and actually I would say that research has continued to grow on ZK but I think the reason you maybe don't sense that is that the breakthroughs aren't necessarily like the big shifts the big step changes they're not coming like in this flood but when you look at like the small like basically research papers that talk about ZK or doing like an incremental, you know, addition to something. There's actually a lot like to the point where back in 2020 we could barely fill a like

the list was like six ZK things and cuz I I read this list and every time there's so many things on how am I going to sort through it? But see, it used to be a mixture. It used to have like a few ZK things, but then we'd put like consensus research and like general stuff. And in the last few years, we've just gotten more and more focused on ZK because there it it was like we we had to cut out TE. We had to cut out MPC like we had to cut out even like the close adjacent stuff because there were so many ZK research papers coming out every month. And I would say that still stands. I mean that makes me think you guys did a bunch of like episodes on lettuce based cryptography, right? True.

So, I started digging into that recently as well. So, I actually have a video coming up, a whiteboard on the on Greyhound. Oh, nice. One of the ladies base. And when I started reading into it, I was like, "Wow, there's been so much research and like everybody's so excited over there." Like in that corner of the ZK world, you know, research world. I know there's like amazing like snarks, there's amazing folding schemes, like things are like only this year there's like a bunch of teams that came out that have like amazing properties and all the best people like are you know Dan Bonet is doing research in that in that part of the world and you know what actually talking to you it's making me realize that it's almost like subcultures emerging you have and and like for lattises specifically the we

had Vadim from IBM research on and he talked about how like the Lattis world had existed for longer and it was just doing its own thing and then recently it started to merge into the ZK world and they've created sort of like this subculture subset and they have very different styles and they're trying to share styles and like learn about each other's worlds and another one would be like the quantum folks quantum cryptography folks where you're like there's the quantum people and then there's the cryptography people and that's not ne necessarily ZK but ZK is in the mix and then you have sort of this group or these like thinkers who are focused on that particular like intersection. And then maybe we could, you know, this infra thing we just talked about with like the ETH proofs

and the ZKVMs and specifically infrastructure teams. They're also kind of a sub subset. I always say subculture or like subg genre. I I was like it's like seeing Spider-Man in the Avengers, you know, like when they mix everybody to or the Avengers is also like a I think the Avengers was the original ZK crew and now it's almost like they've spun out in their spin-off shows. Right. Right. We should we should market it like that. So, one of the reasons that I invited you on was to catch up, but another was that, you know, we really wanted to start doing a series on the ZKI crossovers, like the ways that we're seeing AI and ZK intersect. And I think

actually Kai from my team had been speaking to you about the fact that you were going to be doing some work or some research or some sort of presentation around this and this prompted me to reach out. So with today's episode, I think what we're going to be doing is talking about like AI, but specifically in your context, like where AI and auditing kind of overlap and how ZK security can be enhanced, could be threatened by AI. It's there's actually Yeah, two sides of the Definitely. Yeah. Double-edged sword. Cool. So yeah, like why don't we start with sort of the the question how like do you use AI today?

How are you using it today? Yeah, so you know ZK security is a I mean we do we do different things but auditing is one of them and so we're looking at the whole gamut of stuff that we can use like from formal verification like we have a a clean framework called clean um we were looking at how to improve uh compilers. So we have our own compiler called no name. Uh we're looking at fuzzing. So we found a number of bugs using using these sort of a dynamic analysis uh techniques and so sort of naturally AI was there and we're wondering okay can it find bugs you know it's pretty good at coding and if it find bugs uh shouldn't we start looking at it because this is sort of a you know existential crisis in a

way if it gets good enough to find bugs like can it replace us and and these kind of things and or if a company you know another competitor gets really good at using it then you know this this is uh again threatening to us. So let's let's look into it and maybe let's be the first one to look into it. That that was sort of our uh mindset when we when we started and so we we started looking specifically at at like a very small subset. So there there's been a lot of research on like using that you know this is the kind of things where nobody talks about it but once you start looking into it you see many many papers and that submerge iceberg again. But is this this is in like traditional auditing? This is not necessarily ZK auditing, right? You mean traditional

auditing, traditional bug finding? Yeah. Reverse engineering, uh comp CTFs, competitions, not just to find bugs, also how to fix bugs when you find them. Like like there's a whole and and when you look at these papers, they also use different techniques, right? It's like do you prompt an AI to find bugs? Do you ask an AI to generate tests to find bugs? Like there's many many different ways you can um you can use an AI. And so we looked at that and we just we just thought okay let's just do just apply it to a very small subset. Let's target circum and let's try and expand from there. Let's get good results just looking at circum.

So if you don't know circome it's this uh zk programming language to write circuits pretty low level uh so low level that people write a lot of bugs using it. And so we were like okay let's let's take a look at that using the the lens of AI. I have two questions here. One is which tool did you use to start experimenting with this on circom specifically and did you have to train it? Did it take time for it to get to a state where it's like usable? Okay. So these are two good questions and and they're basically rabbit holes in themselves. Okay. Um so for for the training part basically we quickly understood that. So so usually we're not going to train. So some people have done that right they take small models or like open weight models like these open

source models and they'll try to train them or fine-tune them on their own stuff right usually these models are not very competitive compared to like the state-ofthe-art models you know closed source models of like openi or code or and so we quickly were like all right we don't have the we want to get to a product that's interesting or product or something that can help us right that that's interesting and So fine-tuning something is going to take it's a rabbit hole that's going to take us a lot of time and we cannot really iterate interestingly using that. And the reason why is that as soon as there's a new model and these things happen quite frequent frequently these days then you're forced to retrain it or fine-tune

it again right because you're losing that as soon as you switch to a newer model. or if you want to, you know, try different models or maybe use a combination of models, then you're kind of stuck in this like training phase all the time. And so, yeah, the benefit of using a state-of-the-art model is that you can quickly change, you can quickly test different things and every upgrade, your product gets better. And that's really one of the most, you know, fantastic things about working with AI. It's just whatever you do, whatever you build is going to get better over time. Wow. magically like by itself just by upgrading the the model by having a better brain on the other side basically. Right. Right. That's crazy. I believe it is a brain. Yeah. Yeah. I mean we tried different things. We tried like open way like local models. Uh we

tried like distilled models like smaller versions of like deepse and these kind of things. Um we've tried Gemini, we've tried you know OpenAI, all of the all of these different models. I would say the most success we've had have been the reasoning models. So reasoning models are basically the the more recent breakthroughs in terms of models 03 O4 um and so these models basically will will have reasoning tokens. So you use not to be able to see that and know you can see it. So also the API is changing all the time like working with AI APIs is is a is kind of an interesting thing by itself but just just to explain what reasoning models are. They will think

through by writing text and this text will basically I've I don't know if it's correct but I've read that somewhere that they use XML tags like think and then end the thinking. Wow. And they will do that several times to sort of uh write their thought and then you can actually ask them to think more and they will keep doing that and you can get better results and when they do that you get better and more accurate results. It's uh it really works. It's like fascinating. Yeah. Would you like is is the actual process in this case then you just like take the code stick it in like you maybe take a I don't know if you can put all of it but you'd have to take some segment you put it in you'd be like is this broken if so how so I mean it can be that easy and I think

everybody who used AI has done that right like you upload a PDF or you just copy paste something and you just ask something and I think that's the best way of experimenting and actually a lot of working with AI is about prompt engineering So you're you're always trying to tweak what you're asking. Um today I've heard from a friend that works with AI also at his job that if you're too nice with the AI, you don't get good results. You have to be actually removed. Yeah. So actually the some guy at Google said that also. Uh if you're don't say please don't say please. I heard please just waste energy. Well, it wastes some token space, but my friend told me that it also gets you worse results. So they remove all of these, you know, please and whatever from from the prompts. And so the I I can't remember who at Google

said that, but if you if you're mean and you if you're like threatening to the eye, like I'm going to beat you up if you don't, you know, do a good job, you can actually get better results. So prompt engineering is a is kind of like, you know, throwing words. It's kind of like a prayer. It's like you're throwing words, you have no idea if it's going to work or if it's it's very mis, you know, mystical in a way. And uh and I also compare that to fuzzing sometimes or like you you're by tweaking the way you ask you're sort of changing a seed and so you get different random results and so yeah I I think different products are going to get different results through their secret sauce prompts basically.

But I just wanted to to say one one more thing on because I I think you're I mean there's different ways to to answer your question but another interesting way is to talk about the the context window. And so when you're looking at a codebase, there's a lot of code. And so you might want to like copy paste like all of that code into the the prompt. But the the context window, which is really the the amount of tokens or like text you can paste in a prompt, is different depending on models and it's it can be very limited. So today with uh I think Gemini uh the latest version of Gemini you have like 1 million uh tokens and soon to be 2 millions and with OpenAI you have you're still kind of stuck with uh 250k

uh tokens. With some of their newer models it's 1 million tokens but there's stuff that they cannot do and uh so they're still trying to catch up in terms of how much can you throw at it. Wow. And so you're very quickly when you experiment experiment with these things the naive way you you you you reach that limit that token limits you know fortunately circum is smaller like it's very lowle code like the code bases are smaller but average these limits playing with with the tool that we have uh by the way it's called snark sentinel maybe that's a good moment what's it based on we actually didn't get to the end of the story of which tool you're using you sort of talked about the experimentation but what are you using now also I realized this this models but uh well or yeah which but which models are you

using today? Yeah. So, so I meant uh we we're using different models mostly open uh open eye so 03 mini uh so no 03 and uh Gemini I guess I haven't tried the latest Gemini but it's very easy to to switch and uh if we want to think more about tooling I mean it's a Python codebase we have a NexJS inter web interface that's that's kind of nice but it's mostly Python uh all the all the AI tooling is sort of Python oriented these days got it and then but is part of your work then finding ways to get kind of the check without exceeding the token limits. Is that sort of part of the work? So a lot of the work is actually not necessarily ZK oriented but just

solving engineering problems to like fit stuff in the context window. Wow. And so I mean like a lot of my time in the beginning was just spent doing like like static analysis on circum to be able to retrieve only the code that mattered and have this kind of like interface for the eye to use. It's called function calling or tools so that the can ask for like more implementations. So like if if the agent ask like can you give me the implementation of this function then we we can pull it up without you know with with good accuracy. And so a lot of the work was just doing that like just solving these kind of problems which actually are getting solved by other people including like OpenAI and stuff like that. So if I wait long enough I

don't have to work on these things anymore. Well because they'll they'll basically or somewhere you'll find tools or there will be tools kind of offered to you that like automatically do these things. Right now you have to build them but yeah exactly and when I started MCP was not really a thing. Uh function calling was kind of hard. There's also like I had to like develop this idea of very quickly you realize that if you cannot see the traces meaning the discussion that the agent has with your with the static analysis tool that we're giving it access to and maybe like web search and like other tools you you can quickly see that you're not going to make any progress if you don't have access to that. So part of the work was also developing like this visualization of traces which is not done by OpenAI so

you don't need to do it yourself but I you know I spend the time doing it and I could have waited for them to do it and so now you I can see you know I I have this tool I throw it it will find things to analyze it will I mean I can talk more about that it will ask other agents to find bugs in it and and different agents in this in this uh in this app and what I can then see what's really happening I can see the discussion they're having. I can see when they're getting stuck. I can see when they're doing things that don't make any sense. And so I can refine the prompts to tell them not to do that. So sometimes they don't ask for an implementation of something and they make an assumption.

They're like, yeah, like if this, you know, this function returns something and you're not checking the return value. And I'm like, there's a return value. You should have asked for the implementation before assuming that. And is this where you get mad and you start to be abusive? You Well, Right. You can get mad at the at the eye, but I So, to be honest, I haven't tried I I recently learned that you have to be mean to the eye, so I I need to try that. That's funny. I catch myself being a little bit like short with it sometimes. Says a lot about yourself, you know, cuz I'm always nice to my AIS.

Um, okay. I want to talk about the different kinds of like tools that are that have been traditionally used. You've mentioned static analysis. I actually want to kind of explore what those are and how I mean first how they're interfacing with AI but also like if they are getting more and more replaced by it. So static analysis and we did talk about this I think last time we you were on and we've we have covered this with other auditor auditing teams that we've had on the show but yeah can you tell us again what that is and what does the what component is the AI playing in that? Right. So I guess if you're I mean the before we had AI or or the of today we we sort of had access to

you know manual audit like using our our own eyes or static analysis or dynamic analysis and which one's fuzzers is that dynamic so fuzzers or dynamic analysis because you're executing the code right you're running the code trying to see if you can crash it or or find some memory corruption bug or these kind of things. For static analysis, you're just looking at the code itself. Maybe you're getting looking at some output of the compiler or like some some a more information than just the code like the types or something like that maybe and then you're trying to infer um some properties or you're trying to break some some properties or assumption that's not that's not there. And you can think of testing as dynamic analysis

also in a way. And both of these fields kind of suck because they're very time consuming. Like writing fuzzers is time consuming. Then you have to run them. Then you have to optimize them because like if fuzzers are slow, it sucks. Uh so so you have to make sure they're fast. Any cryptography in your code will make them slow. So it's kind of a nightmare to to to work with fuzzers. I personally I don't like it. I don't enjoy it that much. I think working with static analysis seems much more fun. Okay. Okay. I mean the the team we have at ZK security working with uh this formal verifier u in lean is seems to have a lot of fun but it's a lot of work and you can't verify many many things like there's a huge investment uh you

can't you cannot verify you have to like create more bespoke static analysis tools for different systems I guess exactly there's this switching cost whenever you do something if you want to apply it to somewhere else you have to switch and you have to rewrite a lot of code. You have to think a lot when you do it. Like fuzzers are very dumb, right? It's very naive testing and you can test on a lot of things, right? And you you can, but it's it's very also limited because it's very dumb. It's just throwing random stuff. It gets stuck all the time. You can run a fuzzer for, you know, people run fuzzers for years hoping to find bugs and and then they find bugs sometimes after a year of fuzzing or wow. Um but full mode verification is something you run it

tells you yes or no, you know. Wait, static analysis or formal verification? Uh, formal verification, static analysis or or are they the same? Would you put those in the same category or is like one a subset of another? I mean, I would put them in the same category. Know that I'm saying that I'm like, is someone going to get mad at me if I say that? But, uh, well, because I had another question about formal verification specifically. Did not realize that that's it would be sort of in the static analysis silo. Interesting. I mean, so so it sort of depends, right? Because like some formal verification is done on the code itself and some amount of formal verification is done separated from the code. Like some people will write formal specs that are not tied to your code and they will run this

analysis on it. And so you can think of it as separate because it's not even looking at some code, right? Okay. But sometimes it will use generators. So you will generate so you have some formal spec and you can generate an actual implementation in some language based on that formal spec. Um so like there's like fiat crypto which is a library in in C of like different cryptographic primitives and it's I can't remember what what formal verification tool they they use for that but they're generating code based on the on that. Interesting. So, but when you're talking about doing sort of like AI's interfacing with static analysis, it could be like a spec that's been formally verified and so maybe here let's talk about because I actually I think myself and potentially

listeners of the show might be more familiar with that. We've actually covered that more often. So, just talk a little bit about how the AI deals with that. Like are they creating the imple like are they creating the formal verification? Are they just looking at what somebody else has already done and then like using that? So maybe just just to be clear, the the AI tool we have and what we're experimenting with doesn't use formal verification. It basically is about throwing code at the eye and seeing how good the eyes is at just reading the code, thinking about it and finding bugs. uh and I can talk about the results and if it's a viable approach but just to to go back you know to continue on on the formal verification stuff it's interesting

because formal verification is you know it's hard to use it's hard to takes it's time intensive too right right it takes actual brain power to to write but with AI I mean I think the the the formal verification team is already using AI to write parts of it like they or to simplify code that they write and it seems to work with like some amount of success. Probably can talk about it more than me, but also like your formal verification team. When you say team, what do you who do you mean? Oh, the team at DK security work. I thought because I think that isn't there a Oh, no. There's there's a team called runtime verification, not formal.

I thought maybe it was like some team we weren't aware of. Okay. Okay, got it. So, no, when we're looking at AI and how we can use it to find bugs, the sky's is the limit. So we can use AI in the way that we're looking at it right now or you can use AI to write tests to generate tests to try and find bugs or you can use AI to uh create mutation of the code and see if that creates bugs. That's something that Meta just did. Or you can use AI to write fuzzing harnesses so like fuzzers. Yeah. Or you can use AI to write a spec that you can then formally verify. Wow. So, so there's many different approach and you can also have AI so have your agent run static analysis tools like already existing static analysis tools and so and and try to understand the results

and try to to use that results as well. Which of what you just listed do you think the AI is best at? Like are there some things where it's like they could do it but maybe you don't want them to do it because it could introduce a lot of problems? Are there somewhere you feel a little bit more confident actually having them do the work? I know this is changing too, but like today if you took a snapshot. So I think AI is pretty pretty much good at everything. And what what we understand I mean you know after doing that work I mean we're still in the middle of doing that work. It's probably going to be a never- ending project on our side. But what we feel like right now is that any AI tooling that's comprehensive to find bugs is probably going to be a

combination of different agents that are good at different things. So you're going to see an agent that is good at reading code just to find bugs. You're going to see an agent that's good at creating tests. Maybe one that's good at creating some some property testing or like fuzzing or something like that. another one's good at using static analysis tool and and or formal verification and and so the more of these sort of agents you can throw at the problem. I think the more coverage you'll get. Wow. And maybe I can talk about that. I there's so many things to talk about. I'm sorry to go on tangents all the time, but I think there's a parallel that happened a lot as we're working on this. uh which might not be apparent if you're not used to working with AI, but basically working with AI

feels like you're working with a factory of workers. Wow. And the different agents that you create are basically factory workers that can like pass work, you know, they they have their workstation and they take work and they finish and then they give it to someone else. And you're trying to orchestrate that. You're trying to figure out who to hire, you know, what kind of skills do we need? And it looks like a it's very human and at the end of the day an LLM is a brain, right? So it's it looks well it looks it sounds like an org chart actually. It sounds very like like a like a factory floor or something or an IT department. Yeah. I mean an org chart is I mean I'm I'm trying to think how we can have an AI run the company ZK

security cuz cuz I'd rather have an AI do it on the business front. I mean I mean on any front maybe like scheduling talking to clients like what more on the business front it sounds like I mean even on the people front like on boarding people like so that could be interesting. I mean I wonder if that doesn't come a little later though because I think it's so clumsy still on the on the people front. You you see a lot of uh hotlines right like uh assistants or basically I know. Yeah. Like even for interviews now like I've heard the first interview. Wow. Yeah, we'll be an AI dystopian. Yeah. Um, # dystopian.

You talked about the sort of different ways that an AI can act, the different kind of roles. One question, and it might be actually covered in what you've already said, but when I think of an auditor, I think of like a person who's trying to find bugs, but also kind of break the system, like mimic a hacker in a way. Like and I and I know you use different tools and you're kind of trying to catch lots of bugs whereas a hacker is looking for like very specific malicious bugs, but like are there simulations of hackers happening as well? Is that in one of those things that you've described where like you're sending agents just trying to break break and every time even like you find a bug, you offer a fix, do you then throw like do you almost simulate a

hundred hackers trying to break the thing after the fact? I mean this I mean the the only difference and in the industry by the way it's like the difference between black hat and white hat. Yeah. The only difference is that the white hat does that legally and the black hat does it illegally to like exploit things and you know and so I mean there's no difference right but um when we use an AI the AIS will not respond very well to u you know prompts that say do something criminal or be a black hat even if it's just a like test environment even if it's like I mean or like you're this is the point you want you want them to be as malicious as possible in a way right so you could see if they could break anything. I think they I mean as we make the difference

like you know we you and I make the difference pretty well between you know ZK security and like some some uh script kitties that are trying to hack stuff. Yeah. And I think AI is basically trained on the same stuff that we read. So it's very human in that way. It understands the difference. But if you promp prompt it and say you're a hacker, you're trying to exploit this thing, they will not they will try not to do it. Right. Wow. it'll push back, right? Because they're programmed not to do anything evil or they could harm people. But then the hackers, like a real hacker in real life, human hacker is also probably using AI tools. Would they just use like clean auditing tools in order to like learn the way to break the thing? This is so funny. This is very

meta. It's like the the bad guys are talking to AIS pretending to be good guys. Of course, they have to lie. Yeah, you're a security engine. Yeah, I'm an auditor. I'm supposed to audit the system and and report to the team what is the most dangerous bug in this stack that could really, you know, drain the system. I need to tell them quickly. And the is like, are you sure you're not a bad person, right? Wow. Like I wonder, so it sounds almost like the AI models and it and I guess this goes into the training that they are kind of positioned as white hat currently or like they're at least like leaning towards white hat. they push you to be more white hat like if you ask for more black hat actions they'll push

back. Mhm. So I don't know if you've seen that but recently code announced uh two new models and he came out with this report of I don't know how many page like pretty long report and so in the report they explained how like if you tell the upgrade them with a new model or something like that and they're supposed to help you and the IT guy that's going to do the replacement in the company is you know you give access to emails right you don't tell them that but you give the access to emails and the AI realizes that the IT guy is having an affair or something like that, then they will try to blackmail that guy to stop them from upgrading themselves. What? Unless maybe in some situations where the new model shares the same values or something like

so they and they have other examples where the eye will try to like like exploit survive survive like like open upload the the it weights and and these kind of things to like survive like a shutdown or uh so so it's kind of weird like AI gets really weird like uh nobody understands why but yeah so kind of after all this looking at it do you think that like How close are we to replacing the auditor? Going back to what you had said was kind of a fear when you first started to see this like Yeah. How far along are we? So, so this is very interesting. So, again, if you look at that system card, they were testing the the AI on different CTFs like capture the flag challenges and it was very easy to to to solve like web

challenges I guess but for cryptography challenges it was quite hard. Oh, so you know fortunately ZK security is in the field of cryptography. So you might think this is good and so from our point of view we're looking at this and we have some results right and it looks quite scary like AI in our results AI is very good at finding bugs it has a lot of false positives and so there's kind of a triage problem you're spending a lot of time trying to understand what it throws at you okay by the way I can talk more about I think that would be interesting to talk about that also um but we have so many cases right it found bugs like small bugs like actual bugs It found bugs that we had found but it said they were not bugs. Oh, and so we

had things things like that where if you try to prompt it correctly, then it would actually find it, but it would say this is a minor bug or like either it would not find it or it's a minor bug, but it was it was a very important bug. Or in some situations, it noticed that there was something weird, but it it read the dock and it said, well, the doc says correctly that uh this is not insecure because of this and this assumption. Like in cryptography, you cannot factor numbers or you know 2,000 beats or like whatever. And we found out that this was not true. This was a wrong assumption. Wow. Because uh if the if that number has a lot of small factors or something like that, then you can factor it easily. We we have this kind

of mixed bag result where you have a lot of false positives. So it thinks there are bugs, but they're not bugs. it finds some bug sometimes it could or it points towards a bug but it thinks it's not a bug or that it's a minor bug when it's an actual bug and bottom line I think things are only going to get better from the from here for for me this was a very impressive result what we got you know knowing that I don't think we're going to stop anytime soon to work on that tool and and to use AI as auditors can it replace us I think that's the existential crisis question and if I was in the smart contract industry For example, smart contracts are easier to audit. I would be very scared. I'm I'm I can tell you, you know, 99% sure, I

won't say 100%. That this whole industry is going to die like the the industry of auditing smart contracts because they're too simple and the bugs are probably so consist like they're so common. They're so common. There's a long list of bugs that people always look for. Most of the consulting companies are not very good. They just look for they just go through checklist of uh re-entrancey bugs. like whatever are the common bugs for for smart contracts and so AI is going to get better than like you know I don't know what percentile but it's going to get better than most consulting companies I would expect interesting but I think we're you know we as a very niche expert auditing company I think you have a few more years yeah we have a few Exactly

it's not about if but it's about when probably but yeah we'll see you you talked about false positives what you wanted to elaborate there, but what did you want to say? I mean, false positives are annoying because you you can imagine you're throwing that tool at a projects and you don't really know how it works, unless you're the the developer, but but even then, you might not really understand parts of your project. You're reusing libraries or, you know, and the tool will find, you know, 10 20 bucks. And at this point, you have to figure out, okay, are these serious bugs? And you're gonna basically spend a lot of time going through these one by one trying to understand and discard things are not not bugs. And so maybe 80% of the the findings are like that. It's a

waste of time almost. Yeah. So so if you're misusing these tools, it adds friction and it adds it's too too much of a waste of time and so you're not doing your job at the end of the day. So So there's right now it's sort of a trade-off when you want to use them and how much you want to use them because otherwise you're you're wasting too much time. True. And I think developers are going to use them more. I I mean I'm sure developers are are are trying to use them themselves. And for us, I think we need to get good at trimming the false positives, discarding the false positives so that this tool can be more useful. And so now we're experimenting with different ways like an agent whose only job is to like identify a false positive or something. Exactly. So this

is called LLM as a judge. And basically you use a different AI or different agent to try and judge the result of another AI. You can just judge things, you know, just by being an expert at judging or having, you know, whatever that means, being a better model, for example, or by having more context or by being able to provide to create proof of concepts. So one interesting way is like if you find a bug then give it to this other agent that's good at creating proof of concepts and if they can create a test that runs and outputs a result they expect then it's a bug. Okay. And then we can have another agent that will question assumptions also you know that's good at questioning assumptions

doing rep research and these kind of things. I feel like all of this though must be such an intro like introspection into how you approach any particular problem where you're like okay if I see a false if I see a bunch of positives and I have to investigate each ones what are the steps that I'm taking what exactly am I thinking about what are the like angles I might do depending almost on intuition like where is it in code what kind of code am I looking at and then you're like okay there's a bug being reported but how do I tell if it's real and then you have to reverse engineer that. You have to kind of like break it down into these very determined steps and then like create an AI that does it, create an agent that does it basically. Yeah. The agent is

basically a human and they're they're good at doing what you can do and so you have to figure out what you have to do so that you can replace yourself with an agent. I mean, do you think that like after auditing a long time, like you do develop an intuition, which and to me intuition always like the more I think about what that is, it's just like you fed your human model brain your whole lifetime of experience or like one particular skill for a period of years. And what it allows for is speed often. Like it just means you can do you can make choices faster. You can like figure out answers faster because you've seen things and you're sort of pattern matching and there's like this Yeah. How do you like how do you create intuition in an in an agent? So I would say I mean

there's different ways to look at this question. One way is to see models as you know the the better trained they are. If you train them on like solving security issues and looking at bug reports and these kind of things they're going to get better at that. If you train them on circum they're going to get better at reading circum. Yeah. Then there's the idea of in context learning which is you have a model it already learned. Can you teach it more without having to fine-tune it? And so in context learning is like if you retrieve enough information that's relevant to what you're looking at right now and you put it in context in your prompt or you give some examples or these kind of things then the model is going to be actually very good at solving the same

kind of problems. I mean to answer to the two things I said we actually have to give it a number of guidance in circum because we realize that models don't really understand circum and we also have to give it uh bug reports or example of of bugs and if we can give it example of bugs that are relevant with the codebase we're looking at it's even better and so that's where uh we have these techniques of uh rag retrieval advantage generation where you're trying to find bug reports or write s and these kind of things that are relevant to what you're doing so that you can bring them in context and so once it's in the context the is better at at solving these things.

Interesting. Okay, maybe a third one. So I was talking about reasoning models like this whole like chain of thoughts idea of you know giving the eyes this this ability of of using chain of thoughts to me I understand that as giving the eye the ability of putting things in context like they're bringing up words and sentences and tokens you know that will help the AI make a better judgment at the end. So, so there's many different ways and I didn't even talk about prompt engineering. We talked about that earlier, right? But you can spend a lot of time just doing prompt engineering. If you ask a thing this way, you know, you're a security consultant or you're very good at do your job, you know, like Oh, yeah. You like different ways. Funny. Mhm. Do you ever have to fire an agent? Like, do you

ever train an agent and it kind of gets worse or like goes south and like starts to do Yeah. Like it it consistently does something wrong. you have to nuke it. So I mean something we realized also pretty early on is that if the more you put things in your context that are irrelevant or the longer your context gets or the prom gets or the discussion gets the weirder the results and I think everybody has noticed that by the way because most people when they use you know chat GPT or whatever will often open a new window. Yeah. Instead of continuing talking in the same window they will often open a new context a new window. And that kind of shows that we all understand that AI just gets bad or the the answers just gets bad over time

and so you sort of have to reset them or or make sure that whatever you're talking about in this conversation is very focused. Yeah, I've noticed I mean I've noticed it get stuck on things like get stuck on one idea that it can't seem to shake even though it's incorrect and you have to kind of like you and and I've even you know in my more direct more short way said something like no we've already said this this is not it why is it back but yeah that does happen and I it's like kind of surprising to see that I heard I heard that once the eyes starts getting something wrong then it's very hard to correct it after like it will end up in this kind of a interesting rabbit hole of a

incorrectness. But yeah, we're I think we're all experts at AI at this point. Like if you use AI day-to-day, you're intuitively understanding how to work with AI, how to prompt an AI and these kind of things. I want to ask about ZK specifically. you'd sort of said like I mean we talked about sort of the longevity of the auditing profession in the context of like cryptography versus the regular because a lot of what we covered I feel in in this episode is much more like auditing and AI but could be in various contexts going back to like the ZK side of things you talked about circom as like your starting point but is there anything else kind of unique to zero knowledge cryptography or the TMPC like sort of this advanced cryptography that yeah that where you're you have to use other things like

because it's cryptography implementations do you also have to have like the math research side of things kind of like you know understanding the logic before just looking at implementation I mean in a way not really like bug finding is bug finding no matter the framework or the concepts for zk maybe there's more math involved so like some definitely some bugs are harder to find compared to again smart contracts or like web bugs or these kind of things they tend to be like non-standard like depending on what people are doing like I I don't know I was thinking of like some bugs like where where you have to like binary decompose something and like do some arithmetic uh using limbs and these kind

of things like it's it's very tricky to show that something is wrong when you do these kind of things when you create constraints to show that some operation some mathematical operation is valid within the circuits and so my intuition is that We're going to have a hard time finding bugs using AI that are such bugs. But again, AI is getting like we're very impressed with the results. AI is getting better and better. It's not clear where the the ceiling is. Wow. Do you like after doing kind of the initial work on CIRCOM? Have you added other libraries or have you been like I'm I'm assuming you must be using other libraries as well. Does the sophistication or like maturity of the

library ever have an impact on like how well these tools work? Uh I mean we haven't experimented too much outside of uh outside of circum just because we had all this like machinery specifically for circum. We've worked with like some rush projects but usually it's pretty simple. You you give access you give the agent access to like you know you can read these files. You can ls and and see what's in that directory and then find bugs. But it's usually harder I think to build context for the AI to find bugs this way. So, so far experiment were more mostly successful with cir and so it'd be hard to generalize. Have you tried auditing anything like that has

been built with noir? Uh, we haven't looked at no more yet. Okay. Yeah. I'm just I'm I'm almost wondering if like the age and sophistication and development of a library and the like and languages like circom's been around for a relatively long time. It's been looked over and and you know added to and lots of bugs have already been found in it. Yeah, I'm just wondering if like the newer ones if it's harder to inter I mean it might be just harder to audit in general because they're less wellknown and I'm wondering if like AI also suffers that. I think so for a lot of bugs it's kind of like logic bugs. It's based on what you're implementing or like for example the the good bugs that you would find by itself or like off by one bugs. So if you don't know what off

by one bugs are, it's like you have an array of like six elements and then you do an operation that's on seven elements or like five elements like you you're you're off by one I guess like the title says very little. And so the is very good at finding these bugs and these bugs happen in all sorts of languages. So it doesn't matter if they're in war or in circum like it's going to be good at finding them. And if it's like a you know protocol and it's misimplemented and by reading the logic you can see it it's very possible that an agent would be good at finding it that it's noir or not noir. The problem that noir has is that I I would imagine that most agents or most models don't recognize noir

because it's too recent. Okay. So actually there was some work with move the move language. Yeah. And they actually had to train some uh well I don't know if they had to but they they trained some smaller models uh the cutting models from from Google on move so that they could find bugs and even then they they found that using the state-of-the-art models was was actually better. Okay. And so I would I would think that for noir you're going to have some issues because you're going to have to bring some stuff in context in your prompt just to sort of teach it a bit more about noir. In the same way when we looked at circum we realized that so we have this evaluation framework that's very important and basically we ask the same questions to a bunch of AIS most of

the small models don't even know what circum is sometimes it will say yes I know and it will produce some circum code that doesn't look like circum at all weird but even the bigger model when you ask them tricky questions but you know security critical questions uh for example do you know if the assert function create constraints So there is an assert function in circum and assert in many many languages allows you to to say you know assert that this property is true you know that this thing is one or but if you do that in circum actually doesn't create a constraint and so if you rely on that function in circum it's insecure and we've actually found bugs like that but the AI wouldn't have found these bugs or like the 03 mini and uh 01 and like like

the the different models we were trying would not have found that because they were not good at understanding that. So we asked a question and they and what's weird is that sometimes they would say the correct thing and sometimes they would not say the correct thing. So what we do know is that we always provide guidance. Okay. We always tell the heyert function is not secure for example. Always look for that. Do you think the AIS of today would have found for example like the Zcash bug, the one that you know they had they had to sort of keep quiet um back in this is like we did an episode long ago with Shawn Bo talking about this story. So we'll link to it if anyone's curious. This is a this was a bug in a live system that was discovered by the researchers that were

there, but had it been exploited, it would have actually meant like a major financial kind of hit like, you know, I think they could privately, you know, disaster. Exactly. They could privately print money um that would never have been traceable. It was kind of like the most the worst case scenario of a bug, especially in a ZK private system. So yeah, I'm just wondering like and and maybe you don't know the answer to this, but I would be curious if today's tools the way you have them, would it have found it? So for this kind of bug, I'm a bit skeptical. So So AI is very good at doing stuff that humans are good at doing, right? Whatever humans are good at doing, AI is good at doing. But that's sort of the intuition that I have

right now. And so the the Zcash bug was a was a paper bug, right? A paper protocol bug. M so you had to go on the paper, you had to follow it, and you had to realize that something was being published and it shouldn't have been published. And normal humans wouldn't find it, but Ariel Gabizon found it. Well, exactly right. It's not it's not a normal human being. Good job. Uh it's it's not easy. You have to understand a lot of things about the protocol. You have to understand the it's it's not easy to find these kind of bugs.

Or at least if if Ariel would write more content that we could train on, then maybe we could find these kind of bugs more easily. Maybe he should train himself on his like private messages or something interesting. The the the other problem with training or like finetuning is that you need a lot of data. Like it it's not enough because you you need to like tweak these weights and to really change them, you need to throw a lot of data at it. So, David, I want to spend I don't think we have too much time left in this episode, but I did want to just spend a bit of time with you kind of talking about more generally like the ZKI crossover space. Um, partly because I I mean you're auditing projects.

You're seeing a lot of projects in ZK. You were saying like a lot of the ZKVM stuff like that, but there was I mean in 2023 there were sort of like ZKML as a topic popped up. There's a few teams that came out of it. But I think we have seen other ways in which ZK and AI are overlapping. So yeah, let's talk real quick about ZKML. Like first, have you audited any ZKML projects? Like are you seeing a lot of those or would you say like there's less today than there was? So I don't want to misrepresent the field cuz I there might have been some development lately that I'm not aware of. I remember there was some year a few years back where ZKML was uh talked about. I remember some people from PSC talking about it. There was modus labs,

there was Ezekiel or Zondrit. Um I'm not too sure what's the status there. I'm not too sure what's if there's new competitors that enter the the playing field. What I remember was that there was this interesting I think Kathy gave a from PSC at the time gave a talk on that and I thought it was a good way of abstracting things but she she sort of presented ZKML as this use case scenario where you have different use cases in the natural use case without ZK you everything's public you're using a potentially you're using a public model or like a fixed model with a public input and getting a public output and when you use ZK what you you can do is

that you can mask the the model. You can use a private model, but at least you know that it's the correct model. Or you can mask the input. So you're using a public model, but you don't know what was asked and you get an answer. Or you can mask part of the input. You know, like you know the question, you just don't know the amount, for example. Or you can mask the outputs, you know, you could be here's what we asked and here's the model, but we're not going to give you the output. I don't know why that would be useful, but you can you can do combination of these kind of things, you know, and so that's that's what's uh that's what's interesting with ZKML. I think it's interesting because like I I I think that back then there was also a lot of ZKVM teams that were at least

like in the groups talking about ZKML because they were like well if you can do if you do ML in this language then maybe you could do it in this VM and I feel like there there have been teams that have kind of like gone with that. So for example, Lrange has done some sort of ZKML stuff. I haven't explored that. And also, I mean, I think there's teams like Ritual AI that is AI, but it's using ZK on some level as well. Um, I would I would probably put it in that category. All right. That that's the moment where Ismile is like, "Oh, I should have told David about what we're doing so you could talk about it during the podcast." I have no idea what they're doing. Yeah. I mean, we just saw announcements kind of like running models and I was like, "Oh, okay. So,

they're they're they're playing in that in that world, you You know, since then we also on the show and just generally we heard about a lot of examples of kind of the opposite of ZKML where you're using ZK within ML and more like you're using ZK to counter AI or counterML like almost like as a to distinguish oneself from an AI generated image for example we'd have image provenence of you know I actually took this picture it's real it's mine or showing using ZK to show providence that it is AI I you know like there's these kind of ways that we started to see ZK being used not within a model but kind of to prove something about a model. Yeah. So that was maybe

another angle. I guess there's yeah identity also is like you need to prove that you're not an AI bot or something like that. There's a real identity somewhere. So you you like ZK passport would be an example of a defending against AI in a And I mean most of the ID solutions right now at least that we're hearing about they're ZK at the core. Um and it is kind of like it's more of a like a larger project to be able to identif to to be able to identify a human, right? And it's it's not really AI like it's not in it. It's but it's kind of because of it or something. Yeah. Yeah. AI is threatening our our our recognition of who is a human. And so I would say so ZK is used in two ways there. Maybe we can think

about it this way for identity and proving you're not an AI. Like we don't need ZK, right? We just need like a signature on your passport or something like that. But ZK will add private privacy to which is important for things like ID. Uh I mean depending on the protocol, right? Maybe in some use cases it's not important, but maybe you know on chain you usually want privacy if you're going to play with passports. And then there is the provenence thing like if you take a picture with your phone or something like that and you want to show that it has been modified in only you know you only trimmed it and cropped it and rotated it or changed the you want to prove the the changes that you made from an original picture. So again there you're going to have a signature and you

know your phone has like some chip hardware chip that will sign it and say I attest to it. This is correct. And then ZK is only going to you know that's the part where it's not really ZK but it's only going to use be useful to compress create a proof of all the transform that all the transformations were you know legit and not crazy and so it's not really about privacy it's more about compression of the the computation. Are there any other ways that you see ZK and AI crossing over? Yeah, I I mean at this point I think the only thing that I can think about is is how in research people are nerd snipes in different directions either from ZK to AI or from AI to ZK. I don't know which direction is the most powerful one

now but but there's been a lot of movements between the two different different fields. Sure. I mean back in 2022 23 there was a lot of like younger sort of ML students like students who were kind of on a path to work in that that came over to ZK. Um I mean one example would be succinct. I think they said it on the show you know back in 2023 that they had been kind of on that path and then switched over and they didn't switch to crypto actually they switched to ZK which I thought was interesting. But I also think in 2023 24 like late 23 24 you did start to see people kind of get fed up with blockchain I feel and sort of move and they might have been blockchain/zk and they sort of moved more in the AI

direction. Yeah that's what Vitalik is saying right like people are getting jaded and then they get replaced by people who are more fresh and new to the field and well I mean if you need to change might as well keep working on something. Mhm. I mean that's why I got nerds knife into this uh snark sentinel like this this AI finding zk projects. I wanted to play with ZK with AI but I I wanted to remain in ZK and and maybe uh get energized by like a ZK XI project you know and uh it's been very interesting. One thing about like the ZKML crossover is I think sometimes because ZK of the privacy and advanced cryptography fields has been like the most developed in a lot of ways. I think

that's why it gets picked as like the acronym ZK. But in fact, I mean, the closer you look at some of these systems, the more it's like actually a and people won't always like this, a TE or eventually maybe an FHE environment would be a better like that would be a better tool conceptually for some of these use cases, especially if like computation is supposed to happen somewhere private. Yeah, I I mean I I was talking about that actually yesterday. For a lot of use cases, a T is perfectly fine and companies don't care if they're going to use MPC or FHE or ZK. I mean, that's why TE's are so interesting even though they've been broken again and again in different contexts. They're so useful. They're still very Yeah, they're still very

useful. Still very interesting in many different kind of ways. um if you care about a different type of thread model, if you care about efficiency, if you care about easiness of deployment and and all these kind of things and uh I mean we're seeing a lot of T applications right now. So that it shows that people are very interested. Cool. Any last thoughts on auditing CKX AI? I guess maybe one last thought like I've I've noticed that coming up more and more in these um AI finding bugs papers that I've never seen before and and they all have this in common or a lot of them have that they have an ethics okay section at the end of the paper. I thought that was interesting to mention

because I remember reading the same kind of thing with physics the physics field right after the the bomb. Oh, they started to like talk about the humanity side of things because you could get carried away without it or something. So after the atomic bomb, phys physicists came together and they sort of created this movement of uh you know science is more than just science. There's also a a human responsibility part in it. And so they they sort of created this code that still exists today in physics, right?

you're not going to research and advance science if it's for killing or destroying these kind of things. So there's a the ethics in in science in physics especially is is really really big. I I would guess in biology as well and in in privacy we always had that for sure but I've yeah in AI and bug fighting it's interesting to see that as well like um and I actually I would say you know I these days when you go to ZK conferences and you attend talks and these kind of things like you see a lot of talks where people don't really care about the ZK part. I remember seeing one of the talk recently and and someone at the end of the talk asked what about ZK is it is it ZK and the the speaker said oh well it doesn't really matter there's

a lot of use cases where you don't need ZK you mean ZK in the privacy front or ZK in the privacy context okay and so for a field that sort of was was born you know from a cypher punk place I think less and less people care about privacy weirdly I don't know if I'm going to trigger a bunch of people by saying that yeah I also I I feel like recently there's been a bit of a resurgence of interest in privacy. I think it cooled uh back in 2022. Like a lot of teams got a little scared of it, but I think they're more into it now.

That's my thoughts. But but but not even talking about products, right? Like it'd be interesting to see if you're in the we should do a poll. If you're in the ZK field, do you care more about the privacy aspect or the compression aspect? And I'd be really curious. Yeah. Succinctness basically. Yeah. Right. Right. I'd be curious to see the results. That's cool. We should do it sometime. But yeah, so so just to go back to the AI stuff, the ethics is uh I mean when I was working this stuff, I was and and getting good results, right? As I was running it and getting better and better better results, I was like, "Wow, okay, that there's something there. Now I understand why there's an ethic section in all these papers." And indeed, it's becoming easier and easier to find bugs. It's becoming more and more accessible. And you can imagine

that at some point in the future there will be an AI that will you know you'll just throw it at something like look at this TCP IP uh implementation and then you'll be able to break the internet just by uh so it feels very scary. Wow. No, I mean the way I almost envision a future here is also like you have security that becomes so secure a human could never break it. But then you have AI like generated hacks that are so sophisticated and so crazy. It's almost like they start playing a game at a level which is so above us and that the minute that type of thing happens it starts to feel very much like we are the pets in the cages and they are the

masters. We're we're looking at stuff that I was reading this article where AI creates chips and we don't even understand how the AI is creating these things. It's like more performance and we don't really we're almost at this stage where AI is creating new things that are useful for us and we don't understand how these things work or like writes code and but yeah I I can also see a future where maybe open AI doesn't give you access to you know security capabilities like the the model is u throttled and so will not answer these kind of questions.

And if you want to be able to access these features, then you'll have to, you know, be a vetted security company or government company or these kind of things. I know they're already doing these kind of things with um so they're already assessing the capabilities of AI to help you create, you know, nuclear facilities and bombs and things like that. And I don't know if we should say these words. We're going to get a well, it's podcast on some file somewhere. And uh but basically only some government agencies have access to these reports or can play with these things just to assess the it's it's also something I learned recently reading the the anthropic uh system. The thing is the decentralized maxis will not like

this idea because yes then it's protected in the government but what if the government's run by a group that you don't think is ethical or not your team or whatever um could also be very terrifying. Yeah. or could be the North Korean governments or crazy. All right. Well, David, thank you so much for this conversation. I really liked getting a chance to look at kind of your workflow and the ex like direct impact of AI on auditing in the ZK space. I thought that was so cool to actually like yeah kind of paint that picture. I mean, I'm sure engineers just building this stuff are also using AI tools in their own way, but this was a very cool like a look

into this evolving field. Yeah, it's very cool. Thank you so much. Yeah, it was fun to talk about it and uh as as fun as it was to to to build them or it is to work on these things. Yeah. Hopefully I can come back and do an episode on like no, we don't audit things anymore. Exactly. I'm now doing another thing completely. That's cool. I have a garden now. Cool. All right. I want to say thank you to the podcast team, Rachel, Henrik, Tanya, and Kai. And to our listeners, thanks for listening.

[Music]