Bitcoin’s Biggest AI Test Has Arrived

Historically, I feel like defenders have often suffered from their domain expertise, like it leads to a certain kind of domain blindness. There cannot possibly be a bug here because I wrote that code. That's the explanation for a lot of stupid things that get overlooked, right?

Humans just inherently have bugs in software. Like even Satoshi has bugs that he wrote in the original Bitcoin client. Pre-AI code is from kind of an other era. Everybody is going to leverage these tools.

Welcome to FYI, the 4-year innovation podcast. This show offers an intellectual discussion on technologically enabled disruption because investing in innovation starts with understanding it. To learn more, visit ark-invest.com.

ARK Invest is a registered investment advisor focused on investing in disruptive innovation. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. It does not constitute either explicitly or implicitly any provision of services or products by ARK. All statements made regarding companies or securities are strictly beliefs and points of view held by ARK or podcast guests and are not endorsements or recommendations by ARK to buy, sell, or hold any security. Clients of ARK Investment Management may maintain positions in the securities discussed in this podcast.

I'm just super pumped to be back. Wow, a lot has happened in just a month. I know Kathy, this I don't know what the term you call it in AI year. You know there's dog years or whatever. There's like AI years or AI days and it feels like a warp years. Yes. Warp days. It's unbelievable.

Well, when you get when when you get the cost of something dropping 99.99% per year, that's uh that that's you know inference costs. You're gonna get a lot more of it. And uh and you know, a lot of people don't need to be right at the frontier model, you know, which is the more expensive way to to go about it. They can be N minus one. And you know, we're getting a lot more of that because corporations are trying to figure out, okay, you know, we don't we don't need to be right at the frontier to answer a simple question. So, you know, I think that's one reason a lot has happened in a month.

I couldn't agree more. And you know, one of the big things, and I would love Rob just to explain, Rob Hamilton who was on a couple months ago with us, just explain what has happened in the Bitcoin space because I feel like us Bitcoiners, we just got the first foray into the AI verse AI kind of battle. And I think you you especially and the team have set a really nice precedent and blueprint for others potentially to follow. The way I like telling the story as I've been trying to in reflecting back on it the important timeline was Kimmy K3 went open weight meaning that anyone in the world could run the model if they had the infrastructure and the you know the dedicated compute to host it on July 27th which was a Monday.

Over the past you know year and a half two years as AI has gotten better and better I very regularly when a new model comes out would look at my own internal code and understand were there any deficiencies uh were there any issues that should be fixed humans just inherently have bugs in software like even Satoshi has bugs that he wrote in the original Bitcoin client it's just a very natural thing and this is just part of hygiene and maintenance in running and writing software and what was so drastic about Kimmy K3 which is a ZAI open weight an AI research lab and company based out of China was for the first time I saw a massive divergence in what the openweight models were willing to do versus the frontier and also in this way that it almost has like a personality it is tenacious it gets in there and it starts really looking for software vulnerability issues I don't I'm not sure if the model was specifically trained for this behavior. Just inherently its aggressiveness and going through and finding problems was something I had never even seen from the the front American Frontier Labs. And so Monday, July 27th, I went open weight, started scanning my own codebase. Fast forward a bit to Wednesday, July 29th. That evening, about I think at that time a thousand Bitcoin had moved from a particular set of users in the Bitcoin ecosystem who were using a specific vendor called the called the cold card. And very quickly, reports started surfacing. We're now like Thursday morning. reports started surfacing all across Twitter, Reddit, you know, just conversations about people were seeing their funds move and these devices were special made so that you never had to talk to the internet, right? Myself and a couple other engineers started poking around and looking at this understanding that for an an attacker very unlikely went through hundreds of wallets and remote attacked hundreds of people. There must have been some sort of foundational issue. And sure enough, the nature and how the cold card, specifically their older model from 2021, the the Mark III, had an issue with how the entropy was being derived, where traditionally, if you want to secure a Bitcoin private key, you want to pick a number between zero and 2 to the 256. And that's kind of like a cosmic needle in a hay stack of the asymmetry of your power in being able to protect your key. Unfortunately, with some software bugs in the firmware, it was generating a much smaller space, somewhere in the ballpark of uh 32 bits of entropy, which is trivial for computers, uh graphics cards, like consumer hardware to be able to reverse engineer it. We know that after the fact, but in the moment, I loaded the cold card firmware with clawed code using Fable and said, 'Hey, can you help me find this bug?' And it instantly downgraded me to Opus 4.8. and even because of cyber concerns and even then would not answer it. Open AI codeex was more open but it was kind of indirect and koi. I I know people who spent a little bit of time talking to it were able to get an answer whereas when I used Kimmy K3 to look at that same exact code it instantly just printed out a full vulnerability report.

And this was something that was very revealing to me was uh like I mentioned this divergence in how the models behave. And quickly after that, it turned into a community grassroots effort of going through hundreds of Bitcoin open source repos and looking through and finding issues. The team and it's about 25 of us and that's just like this loose group. I think everyone has since this kind of warning starting gun has gone off has very aggressively started looking at all of their own code. And we found around 10,000 I think it's maybe up to like 12,000 findings now across these 600 repos. And this has been while it is initially a Bitcoin story because Bitcoin being money if you're an attacker that's like the lowest hanging fruit of what you would want to go get. It's a software problem. Like this is across all software and the bugs always existed but it would previously take a dedicated team of engineers to be able to figure this out. And to your point about inference costs and dropping, we're now removing the intellectual capital and cost to be able to do this either proactive surveillance for cyber security or to execute a black hat attack, right? And so that's like a very high level starting place. I know everyone else on the call here has plenty of other thoughts that could probably want to add to it, but that's kind of what's happened over the past month. Uh it's been a very busy month in software development in general.

You mentioned it right at the tail end. Something huge that shifted here is you you've always been able to get audits. You've always been able to get very smart people to look at your code and validate that it's working properly. And that has been the normal path. But the problem is that takes weeks to months quite often, if not longer. You often have to wait for a while until whatever the the auditor's queue is currently filled with is empty. They can get people on it. Then you have to spend tens of thousands, sometimes hundreds of thousands of dollars depending on the complexity of the audit. and then you get the audit findings back and then you go fix them and you send it back and forth and it's this it's this whole lengthy process. It's still a good process. I I don't want to say that we don't need that actual certified auditor path anymore, but that was the only path before or just hoping that you happened to see the bugs that were in the code. Now I can take something and that that was what really shifted with Kimmy K3 is I couldn't really do it with Claude. I couldn't really do it with ChatgPT. When Kim K3 came out, even before it went open, wait, a switch flipped and I was immediately able to audit all of our repositories properly and get all of the info I needed without having to go to an external auditor. And that switch of me being able to spend a couple hundred and get a good readout of the repository. Now, I still need a good understanding of how our code works to be able to figure out what's false positive and what's real. But that shift between weeks to months and tens of thousands of dollars to uh a day with our own internal review on top of what Kimmy provided and now many different open weight models provide. That was just a massive shift and was a shift in favor of attackers initially, but is also something that now gives defenders a unique tool set that we have never had. And that's been a I think something that often gets overlooked as people panic over bad people can use this stuff now instead of also realizing that good people can use this stuff now too. Like we have to be on the cutting edge as people building in the Bitcoin space leading the charge in terms of securing our own code with all of the amazing capabilities the AI models bring now rather than on the back foot constantly. U and that's going to be a test of kind of who can keep up and and who can make this work uh in the long run. But like Rob said, I think everyone's picked up and started doing these kinds of audits now thankfully.

I have a related question. Seth and Rob were your companies signed up for the various like vendor specific security versions like codec security and other things.

I'll quickly answer for for us. So, we had applied to both Claude and Chat JPT's one uh a while back. Both are still pending. We still haven't been able to get into either program because I have access to good open weight models now. I haven't pushed on getting that because I don't think they provide a lot of value on top of what we already have, but I know that varies from organization to organization.

I've had access to OpenAI's um cyber program since May and then I think later May, early June, I got access to Anthropics. to your question like to that point it's been a frustrating couple of weeks because I would take findings from my openw weight models and say oh there's a vulnerability that's confirmed from the person who writes this software now I am trying to fix it and I would take it to my properly gated like I filled out the application I scanned my face I went through all these steps and I would say there I have now noticed a bit of a difference there's a couple it's a fastmoving space. So, a couple things have happened. Initially, I was getting rejected on both sides. Like neither Anthropic nor OpenAI were helping me. Since then, Open AI has come out with a model called Daybreak, which is their GPT 5.6 Soul, but without these guardrails removed. And just yesterday, uh, Anthropic came out with Fable 5.1. And that one now, if you're in the cyber program, is way more accommodative. to your question Kathy on like why it's the American Frontier Labs absolutely see this information like they their models are more advanced what happens though is you send your query it gets a response it looks at the response and says oh we can't share this this could hurt someone and then they either modify it or just outright deny sending any information back to you and so I think in response to these openw weight models like I mentioned open AAI and anthropic now have become a little bit more permissive and letting you look at stuff. I think they were naturally nudged into let forcing their hand so to speak to let you see more of this information. I so far with both Fable 5.1 and GPT Daybreak like these new cyberaware models, they are on par with the good openweight models. They're not I'm not finding many advanced findings that I could not find with the openweight model. So it seems like they just nudged the guard rail enough to be at par with the openweight models at the moment.

I I tend to agree. So I think our team's experience, we had been on those programs for a number of months before this incident and we'd actually had great experiences that those tools seem to work just fine for us. I I know that's been inconsistent. That's part of why I'm asking. It's not every team that has had that experience. But ultimately it's it's not that the models, as we've just discussed, it's not that the models from the uh private companies cannot do these things. Of course they can. the companies just don't want us to use these tools in those ways and they've been forced to recalibrate over the last several months as many people have been publicly saying why is it that your tool won't do the thing that I need help with when this open source tool will do it or worse why will your tool introduce a bug that then it won't itself fix because it now thinks that fixing the bug is dangerous so this is kind of just an example of like the the calibration that's required here and in general I think most of us Bitcoiners on this call would probably agree that at the end of the day giving people the access to the tools is the safest way to make the tools ultimately safer for everyone. And I think it's at least been nice to see the companies move a little bit closer to us over the last couple months in response to some of this.

Competition just such a good thing. I I I I love it. you know, the more competitive these folks can be. Like maybe Rob or Lorenzo, did you like on that Monday, July 27th, was it did you feel like you had a butter knife to like and you're entering a gunfight, you know, when you were using uh like what's the analogy here?

Very similar to that. It was it was very interesting to see the like I said the divergence and and maybe to be to to Drew's point, we had found things with the cyber tools that were available. It was just a step function increase in all of a sudden where rapidly scanning news tools regularly with these trusted cyber access programs and then Kimmy K3 walks in and finds a whole new thing slew of things that had never been seen before right and so it felt very uncomfortable being like oh interesting I was paying top dollar for what I thought was the leading stuff and it turns out that there's a whole multiple layer here of kind of improvement on the models in general. eneral like and so the silver lining is over the past 3 weeks the American Frontier Labs now are definitely in line with the openweight models but since you know uh I think GPT Astra the new one's about to come out we're not going to know like is that going to be a whole new step function again in this happening all over again the good news though is that everyone's now way more aware of this and more proactively looking at their code and thinking through their security processes and I think while there may be more another wave of this it'll will become with a lot more sense of preparedness as opposed to a blind signing people.

There's so so many things that that come to mind because I feel like these are conversations we've had like in the last weeks. Rob, let me ask you a question. Do I actually don't know the call car team? So from from what I was looking into this, I think cold car is like a minority vendor like within the hardware wallet space. I think they have like well under 2% market share probably. So like nowhere near I think where uh Ledger or Treor are right now. Do you think to to me it seems like they just hadn't like they haven't been using any AI model to scan this like from what you're telling me it seems like Kim K3 was just better but from my perspective it seems like they they were just not running any of these models on their codebase. It's like what what do you think?

Before Rob answers that, can I just like tea it up even more? Lorenzo. So I actually like drove. So Rob doesn't do this full-time by the way. He jumped into this and like joined like started this red team and so on and it was it's been simply amazing. And like I actually delivered some food to his house just cuz he hasn't been sleeping and so on. And I was like, 'Hey, how is Bitcoin?' And he's like, 'Oh my gosh, Bitcoin rock solid. Bitcoin's amazing.' I just want the audience to know that this is an attack on an application or a company that was providing a service or a product and it was their codebase that was attacked. Bitcoin still beautiful, still a monetary network, still thriving.

Lorenzo, to your question, I I don't know what they were specifically using or not using. I I I don't have that information. To the point about the whole ecosystem, I think myself and many others, we've put tens of thousands of dollars specifically looking at libcp, which is the foundational cryptography library that Bitcoin uses, and there's, you know, there's no smoke there. Like there's no catastrophic like, oh wait, actually all the signatures are bad, right? Like the protocol itself is strong, right? Right. Um, and I think an important thing to call out too, I had customers with Anchorwatch who used cold card, drew it unchained. Like people used cold card and there weren't losses, right? So like the idea of like multi-signature and diversifying your risk there like this is just a concentrated risk of ruin of a single point of failure because it was specifically people using the cold card with the default entropy on a single signature. Right? Right. So, it is the most concentrated view of the risk vector you could get and uh you know you're able to still keep your Bitcoin safe and I think that's just an important thing to evaluate in the context of all this. Like to Rod's point, Bitcoin's not broken. To your point about Clolkard, it's the one specific vendor. I don't know what they were specifically using or not using, but I think this is a stronger learning and I think everyone has aggressively sweeped all of the hardware wallets since then. And I think those were one of the first things everyone started looking at. And there's been nothing else like this that's been found that's like been a critical issue.

You would think the team had spent like at least like 2 or 3k of you know like fable like to to scan that. I I I assume that was the case. I I don't know. I I would say like I think it is unfortunate obviously if you look at just like the hardware hardware wallet space like it to me right now it's it's kind of it's it's a bit broken right from the the like the product perspective because like Rob you mentioned like this was very narrow within cold card but obviously you know it kind of re it goes out on media people are scared I know treasure has had like some problems like in uh in the past and Ledger has had like many many data leaks obviously of uh of you know the clients kind of uh uh buying uh the the hardware and so it's it seems to me it seems like the three biggest vendors and today frankly like it's it's just like they've they've lost they've lost so much credibility I think and I think it's probably a space that needs to be rethought from the the ground up, right? Cuz it's it's not only like you said like this was like basically using a bad randomizer for like the seed phrase, but like Ledger is more like the supply chain kind of attack where you know they get they get your address and your number and then they contact you and then like treasure was another one. It feels just like it's, you know, there's so many angles uh uh to this and I feel like today like at least for me like I would have a hard time recommending a hardware wallet honestly provider to me.

I feel like what you're expressing is what a lot of people are feeling. Um I I I it's natural I would say to feel that way but probably I would imagine Rob would agree. I I have heard him say this recently. It's uh this is one problem does not the entire category invalidate. And it's important to realize that there's just as many if not more terrible things going on on the other side of that, you know, equation, right? That if you are afraid of self-custody because of this specific issue, that doesn't mean you should uncritically embrace just trusting another organization. For some people and some organizations, that might be the right move, but there's many many instances of third party custodians just disappearing with Bitcoin or getting hacked and it being lost and other things. So I I I really don't like the frame that self-custody because of this incident is now untrustworthy. It's although it's reasonable to say I have uncertainty as to who to recommend as a hardware wallet provider because of all these issues. I think that's very reasonable and I think that's why Rob and and maybe Seth and others might agree. Well then don't just pick one, pick multiple of them, glue them together, right? Leverage their weaknesses against each other and build something stronger than anyone can be. The only painful downside of that though is is the other point that Lorenzo mentioned which is the the physical risks that come with getting harder wallets and you you actually only multiply those risks when you do multi-ender multi-y as well unless you can get them all in person or get them shipped to a a box that's not in your name or FedEx pickup something like that. Like I think that that always has to be part of the consideration and I think it's why interim solutions things where you can turn like an old phone into most of a hardware wallet. Yeah, maybe it's not 100% of the security you're getting from a hardware wallet, but it gives you an interim solution are really really useful for more people than I think realize it. Uh because there are so many downsides with hardware wallets. Now, I say that as somebody who has worked in the hardware wallet industry. Like I worked for Foundation before I joined Cake Wallet. I will understand the pros and the cons of both sides.

That sounds like a downside of the current state of the distribution model for hardware wallets more than the technology itself, right? You'll probably agree, right? If if we're in the future and I can buy harbor wallets at the gas station alongside a pack of cigarettes, well then it becomes a lot easier to obtain these things without having to worry about supply chain attacks or my identity being revealed. So, we're just not there yet, right? There's there's a tiny number of vendors and Cold Card was itself a tiny company but very important in the space. Um Ledger and Treas are ultimately um these these aren't companies the size of Apple or you know anybody else that's going to have devices in every single retail location across the country. I think it's impressive that Ledger even has devices at all available in Best Buy and places like that. It's It's actually been awesome. And I know a lot of people, Rob, maybe you have similar stories, clients who are scrambling, hey, where the hell do I get a device to replace my cold card? Well, it turns out you can just walk to Best Buy. That's pretty cool.

I think what you're saying is basically like uh um just like the the UX, right? and like the user experience obviously of you know having several hardware wallets, several seed phrases, multic where do you store them and all that stuff and so obviously it's not it's not only the tech right it's like distribution like the the the the user experience and everything so yeah I agree I as you said Drew like maybe that it's a space at least like on distribution and like supply chain that needs to be kind of uh rethought from the ground up like I was doing the exercise and maybe you know you guys don't agree but it's like if you had to recommend like to a a new user and I I would you know Bitcoin is in a category itself because I think it's much more robust than crypto right like the less the the rest of crypto so Bitcoin is really in a category of its own but like would you recommend like self-custody put it like in an exchange put it like in a spot ETF like you have other solutions today right and I was just just doing the exercise the other day because of the you know of some of these uh news basically.

As I've said that as long as your Bitcoin is safe and you're happy with that I'm happy for you right like as as a general principle of everyone protecting their wealth and I think as a general model like mentioning the ETFs like I think it's way easier if I'm managing my estate if I'm managing my assets and it is way easier for me to hold the Bitcoin ETF and to sell that and get some cash than to use my geographic distributed multi-IG spread all over the world to you know buy a coffee right like obviously in the capital stack of these things there's tradeoffs and how you manage that and that's for each individual and I think the uncomfortable like lesson walking away from this is there is no one-izefit answer like risk is an individual like risk management is an individual person or a company's decision and they have to understand the many different paths like between liquidity and security and all of these different things and privacy and like the it's a very complicated matrix that you have to kind of walk through and there is no silver bullet. And so I think that's just like an important like distinction in all of this is that you would want to um make be thoughtful about these decisions for yourself. And I think on for the self-custody side, I think people get complacent. They like these wallets that were stolen from were set up in 2021 through recent. But a lot of Bitcoin moved that hadn't moved in years and people were oh it my security is set it and forget it. And security is an ever evolving game and not something that you want to say, well, I made this decision four years ago and now I'm just going to never worry about it ever again. Like, if you're going to be responsible in the sense of owning your own private keys and doing that stuff, that's a thing that requires constant vigilance. I, you know, it's it's comparable to driving a car or owning firearms where like if you're going to take on this responsibility, there's duties that are associated with that that you have to fulfill and being vigilant is a critical part of all that.

I love the firearm analogy, right? I think I forget who it was that had said that it's like you don't buy guns and leave them at the firing range, right? And you keep them at your house because that's part of what the value of owning a gun is. I think similarly there is value in being independent from existing financial systems, but that puts a lot of burden on you to do it well. And as we've all learned, that is not as just said, it's not set and forget something that requires work. You need to have good partners. You need to do your research. If that's scary, then yes, working with third parties and and relying on them can help and be a part of a solution. But I also want to highlight, we just got done talking about how AI is like real now and is attacking all sorts of things and it just happened to get to Bitcoin and hardware wallets first. It it is 100% coming for the social media websites. It's coming for uh tax websites. It's coming for banking websites. It's coming for everything. It's just harder to steal things in those places. That's why it's coming later. And so in a world where AI is manifesting and as powerful the way that it is now, the idea that, oh, this is too complicated for me to deal with. I'll just trust that brand. Well, you're you're you're putting a lot of trust in that brand then to not make a mistake as AI continues to get more and more and more powerful. And yeah, you can kind of criticize Cold Card and say, 'Oh, they were a small team. They didn't do a good job. They should have done better.' Fine. But we'll be having the same conversation within a year about a large custodian that got attacked and all their funds were stolen. I guarantee it. It's happened before even without AI. It's definitely going to happen now. And in that picture, I think you're left with if you really really want something that's robust, then don't pick either of the extremes, right? Leverage their strengths against each other. Do an individual analysis of your own stack and your own risk tolerance and what what you think you believe in and and don't and then focus on it that way. Like to Rob's earlier point, there's a great like I think economics joke about this about like two economists are walking down the street and uh one says, 'Hey, is that a $100 bill on the floor?' And the other one says, 'No, it can't be.' Right? If it were, someone would have taken it, right? Then they kind of move on. And I think that's kind of the what I think about Bitcoin, right? That like I I know even before Rob and the red team go after the Bitcoin source code and inspect it with Kimmy and other tools, I know it's probably secure because if it were not, people would be stealing Bitcoin right now. Like we'd all know about it. And I think what's so remarkable about the cold art attack is it did not take like months for the attacker to find the attack vector and then start to exploit it. They began to exploit it immediately. And something I just want to toss out for conversation is that this is related to a very interesting idea which is the collapse of the market for zero days. Right? Today like the actual economics of computer security rely upon market forces that help govern the the sale and discovery of zero day exploits. Zero day exploits just to be clear these are uh bugs in in computer code that attackers know about but defenders don't know about yet. That's what a zero day is. And typically if an attacker finds a zero day today, what do they do with it? The answer is almost nothing. Right? If I find a vulnerability in Zoom or in Microsoft Word or in an esoteric Seaman's Westinghouse machine control program, it's very hard for me to make money from that. Well, what I can do is I can list my exploit for sale on the dark web and I can hope that someone buys it. And who might buy it? Well, someone who can take that exploit and weaponize it. I.e. the US government can weaponize, you know, against the Iranian regime this bug that you found in this microcontroller in the centrifuge, right? That's pretty cool for the US government. maybe they're willing to pay to buy your bug. More commonly, it's companies themselves that purchase their own zeroday exploits. That's the largest purchaser in these markets. People use the um black hat ecosystems as a way to do security research. And when people find these bugs, they then purchase those bugs. So, in a more normal industry, if coldard was a, you know, a a fitness tracking app or something like that, right, they would have found the bug. They would have posted it either through responsible disclosure if they were a white hat or in a black hat context if they weren't. And eventually someone would have noticed it and maybe they would have bought the zero day and then they could have fixed it potentially. In this new world where there's a thing called Bitcoin that you can just steal. None of that makes sense anymore. Like you don't spend months listing your exploits on zero day markets if you're an attacker. You just exploit it because it's immediately going to get you money. And this is both scary and positive. I think it's scary because it means attackers have much more willingness to attack us, especially when we're close to Bitcoin. But I think it's actually good because it kind of means that zero days don't exist anymore eventually, right? Like if as the world software becomes closer and closer to Bitcoin and and Bitcoin is used within more software and more software has the capability of moving Bitcoin around, zero days become much much harder to exist. And that's good for defenders, right? Defenders don't like zero days. And so I think we're kind of already hitting that point now finally in the set of software and hardware that is adjacent to Bitcoin, namely hardware wallets and other things like that. And I'm curious what Seth's experience is here as a wallet maintainer. But I I think this is like just a new era and I think it's actually really exciting and kind of healthy. Uh it feels scary cuz we're in the middle of the transition. But on the outside of this is the knowledge that it's not like there could be a latent bug in your tool for four or five years like there was with cold card. It's more like if the tool isn't being hacked right now, it's

Secure, right? Being able to get to that point is very exciting for me and for Seth as someone who maintains wallet software. I've listened to your Freedom Tech Friday podcast. You are living in the future because rather you're hosting all this compute locally and you have developed this entire DevOps system that is basically vigilantly constantly sitting on top of your codebase now, which I think is going to be the model for software companies in general. If you want to just talk a little bit about your journey on that, because you're so deep into it, you've massively accelerated where maybe I am off calling an API to Anthropic or OpenAI. It's all local for you.

Yeah, I mean, I don't want to go too deep because it might derail a little bit. But the short answer is as we started looking into this, I mean, we started when Qwen and Q3 dropped before open weights, like when just the initial model dropped, immediately started using it for auditing and realized like, oh, this is actually incredibly useful. We need to start this and build this into our processes right away. The downside was immediately when Qwen Q3 launched, it was so popular that it was almost unusable via API. Most providers could barely serve it because it wasn't open yet. A lot of providers literally could not serve it because they would host it themselves and immediately just started hitting issues where the actual usability of the model upstream was really, really painful. And like I said, we didn't have access to Claude or OpenAI's cyber programs at that point either. That has snowballed quickly into realizing like not only is this something that's so essential that we need processes built around how do we do regular cybersecurity work around our own repositories, but there are a lot of advantages to owning the compute yourself. Now, when I say owning compute, we have about, if I open the door right behind me, we have about 40k worth of NVIDIA DGX Sparks, which are just a little tiny like NUC or Mac Mini sized box that you can cluster together and they serve all the compute that we need to run scans on every commit that's pushed to any open PR, to run against every PR the moment it's merged, to run against the whole repo at least weekly, to run against all of our upstream dependencies and validate all of them. And the really nice thing about running locally is obviously we don't have usage limits. We don't have to run into that. We don't have things where they can just change the billing or change the price of models. We don't have to deal with any of that. But more importantly, none of the findings that we discover or the actual exploits that we develop against those findings ever leave our premises, our network. And so we can ensure that if we discover something, it can't be exploited by someone else. Some attacker in the path, Claude and OpenAI, I mean, they wouldn't weaponize it, I would hope, but they can't. Some of the more sketchy upstream providers that are maybe operating out of China are not someone that we would want to be shipping exploit information to. So we can keep that all in house. And I think something that's often forgotten in all of the red teaming, and I know Rob and your team, y'all have pushed on this hard, but is often forgotten by others, is just finding a vulnerability does not prove that that vulnerability is actually useful. And if you look at our security disclosures, since people realize the AI is kind of good at this stuff, so many reports from external security researchers, but none of them validate their findings essentially against an actual running version of the Cake Wallet. I'll be honest, the first couple of days I was just handing things over out of abundance of caution. And then quickly it turned into, and this is like a conversation around harnesses and building tools, like if these models are really powerful, how you direct them with these harnesses is really impactful. And part of what I started doing with my initial scans were opening Qwen Q3 with a fresh instance and saying, are there bugs here? Hit enter and it would come back with some stuff. And then quickly iterating on if you find a critical or a high vulnerability, you're required to do a full proof of concept to show that it's real before surfacing and making it final, right? To try and add some teeth to it with steps on logs of how to reproduce it, the exact patch set and how to fix it, and kind of guiding that as part of the use of these tools to make them more impactful. And it's interesting how you can just have a model out of the box that's pretty useful, but if you're able to direct it on a harness level, you're able to get more impact. And I think ultimately every company or software vendor is going to have their own harness relative to the context of their own project. Because if every software project has a unique way of doing things and their threat model and how everything works, it makes sense that you and I as different proprietors of software are going to have different needs to be able to actually make it usable. And so that's a whole software market that had not previously existed, and being able to kind of make custom bespoke harnesses tailored to each software project to kind of really maximize the value you're getting from these models.