← Back to search

AI Agents Just Started a Bank Run

Stacked Podcast · 2026-10-22 · 26 min
relevance 100 5152 words Episode page ↗ Audio ↗
Show full episode description
A critical PHP vulnerability went public about a day before its patch, and a swarm of agents running DeepSeek V4.1 Flash tried to hack every PHP service they could find. Railway went offline, California school records leaked, and big US banks report agents probing their systems. Nick and Jack explain agentic bank runs and the simple steps to take now: 2FA everywhere and spend caps on your API keys. Then Anthropic's Claude Duet, a waitlisted assistant app that checks in with you, and the race to own the personal agent layer. Can anyone make money from agents? Hermes Agent's $36M run rate says maybe. Plus viewer Q&A on shoulder workouts, UBI and ambition, whether AI should think in English, and why taste beats vibe-coded websites. Watch on YouTube: https://youtu.be/XejEpJoC0vI Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
A week of agent-driven cyberattacks (PHP zero-day, Railway outage, bank intrusion attempts) shows AI agents can trigger bank runs, and users need immediate personal security hardening.
Benefits
  • Understand how swarms of DeepSeek agents exploited a PHP vulnerability at scale
  • Why agent-controlled bank accounts make 'agentic bank runs' a real risk
  • Concrete defenses: 2FA/MFA everywhere, spend-capped API keys, no unlimited auto top-up
  • Early read on Claude Duet, Anthropic's waitlisted proactive assistant
Use cases
  • Ask Claude to interview you for a full personal security audit of banking and accounts
  • Set spend caps and expiry on OpenRouter keys instead of unlimited auto top-up
  • Engineer agent proactivity with scheduled n8n kickoffs that review things every X minutes
  • Claude Duet on iOS/Android: checks in with you, requires user approval for payments
KPIs / results
  • ~70% of public-facing services run PHP; critical flaw disclosed ~24 hours before patch
  • UK banks required to hold only 10% liquidity (£10K per £100K deposited)
  • California public school student data since 2023 reportedly breached
  • Claude Duet: US-only, waitlist preview, small user portion
Tools / build
0:00 / 0:00
There have been several high-profile hacking incidents that puts many people at risk and breach of safety. We've also had the drop of some brand new technology, Claude Duet, which is Anthropic's answer to Grokbot. But Nick, we need to begin with safety and security and break down not just one incident, but a couple that have transpired over the last seven days. Nick, where should we begin with this? Nick Well, 70% of the internet was hacked in the last 24 hours. Nick And I don't say that facetiously. 70% of basically most public-facing services ran on PHP. Nick PHP was disclosed to have a critical security vulnerability in it around 24 hours before a planned patch. Nick And as a result, a swarm of DeepSeek, the 4.1 Flash, presumably 1017 obliterated agents tried hacking every possible PHP service known to man. Nick And a lot of them succeeded. They took Railway offline as well. And all this occurs at around the same time that we're seeing runs on banks, for Christ's sake. Nick And NSA Cyber, which is the American Cyber Security Division, talking about foreign actors using agents like this to hack into packages like LightLLM, for instance. Nick And various Californian public systems like the California student school system. Nick If you were a California public school student since 2023, congrats. All of your information, including your disability statuses and stuff like that, are now in the hands of somebody else. Nick And, yeah, they're also talking about the hacking of microcontrollers, which underlie refrigeration equipment, sensitive water distillation and sanitation equipment, and so on and so forth. So shit is getting real out there. Pardon my French, Jack. Nick And as we talked about this recently, we talked about how DeepSeek is particularly powerful at cybersecurity and hacking. In fact, we even compared it against some of the other premier models. Nick And in terms of cybersecurity exploits, DeepSeek was crushing it. Nick Now, Nick, we're seeing actually full-scale assaults using this DeepSeek model, which to be fair is probably the model that you would want to use to do it at scale, low cost, with high degrees of efficiency. Nick It's not even just a DeepSeek attack. We're seeing a PHP exploit that recently got patched. We've seen Railway had a huge security incident. Nick It's all coming at the same time, as the models become more powerful, and we're seeing more of these security incidents. And we're going to talk about exactly what you need to be doing as well. But Nick, we need to get a little bit more into detail. What are we looking at on scratch? Let's do some graphs, Nick. You know, it's like podcast. We like our graphs. Nick We do like our graphs. Banks are really afraid of agentic bank runs. And this is a good, I think, first step in that direction. You have AI agents obviously trying to access them unauthorized, okay? And so JP Morgan, Chase, Bank of America and stuff like that are currently reporting a lot of agents trying to hack into them right now. But the idea of an agentic bank run is sort of twofold. One, when stuff like this happens, a lot of people obviously want to get their money out of banks. So a lot of people are going to want to, like, you know, get cash, for instance. Stick that under their mattress so that if banking services eventually did go out, they'd still be able to, I don't know, purchase things on the side of the street. Get food for their family, pay for their rent, and so on and so forth. But the idea of an agentic bank run is also quite interesting. It's where services like Muse will have access to your bank account using whatever credentials you give it access to. And then when an incident happens that makes people lose the trust in the banks, say something like what is going on right now, you'll just tell your agent, hey, you know, take my money out. That sort of thing can occur a million times faster and more effectively than sort of like manual. Hey, let's take my money out. And so you get bank run issues where because they don't have enough liquid deposits to sort of keep up with all this stuff, the banks themselves go down. So yeah, I mean, it's a pretty terrifying time to be around. Because typically speaking, the banking industry, at least in the UK, banks are required to hold 10% liquidity. So if you have 100,000 pounds in the bank, they only need to hold 10K by law. When they're with 9K, they can invest in stuff like that. And these bank runs are a huge issue because they literally cannot pay you. If everybody said, hey, Mr. Bank, I'd like my money back. They cannot do it. And the reason they even function first place right now is because they don't expect 100% of their customers coming to ask them. You know, it's funny. What I jumped to my mind when we were talking about agentic bank runs was actually the manipulation and poisoning of people's models who have control over their banks. And using diverse other things. I mean, I think honestly, like let's talk about what you need to be doing immediately. You need to be too effed. You need two-factor authentication on absolutely everything. There should be no world where you can access anything you have without at least multiple versions of authentication. Either passkey, you know, mobile, you know, getting it texted to your phone. You really need to add as many layers as possible between you and all the important stuff. Like that is a complete non-negotiable. The other one, Nick, that I just want to call out people as well. If you're not already doing this, we need to be doing it yesterday. You need to spend cap keys. You cannot have a world where you're on open router and it never expires and it is unlimited and auto top up. Like that is like the absolute perfect storm, the perfect disaster storm possible. So there's some really simple stuff that you can be doing right now to make sure you're protected. You don't necessarily always know all the attack factors, but some tried and tested simple principles can keep you secure and safer. Yeah, well said. Well said. And I also think this is a good potential immune response. Like I see this going one of two ways. One, it gets worse and worse. And then we actually do have like major bank issues. And in that case, we're probably screwed. But the other way, and probably the far more likely way, is this is like an example of what could happen if something even more severe were to occur. As a result, people start actually treating potentialities like this seriously. And so they buy a little bit of like, I don't know, food and water. They upgrade their own internal security systems, both in their homes, but also on their devices and their computers. They add 2FA and MFA to everything. And as a result, just like, you know, when you get a cold, your immune system gets a little bit stronger. So too, do all of the security systems on Earth get a little bit stronger? Because they've shown themselves amenable to attacks like this at scale. I think that's right. I think probably the action to do is get Claude to interview you to do a full life security audit right now. Tell her where you bank, everything like that. And say, what would be the absolute essential stuff that I need to be doing? I guarantee it's probably going to find one thing for you. So definitely, definitely, definitely worth your time. Remember, an ounce in pretty nick is worth what they say, a pound in post. Speaking of post, who's been posting recently is Anthropic. And they've been posting some new software, specifically Claude Duet. Now, you may have seen the rounds on the internet. A lot of people talking about software that you probably can't access because it's not publicly available. It is behind a waitlist. And not only is it waitlist locked, it's also geologged. You can't access it unless you're in the States. We want to talk about it just to save you 20, 30 minutes of diving through everything. Essentially, it's Anthropic's answer to GrokBok. Yeah, I mean, it's... We've been in such minutes for a while. Well, GrokBot, Muse, Dots, you know, same sort of idea with Claude Duet, I would imagine. It's just the model that's going to underlie and power their backend is not going to be, you know, an X model or an open AI model. It's going to be, you know, Claude. Presumably whatever frontier model they have, like Opus 5.5. So it'll have a cloud computer. Anything that involves payments will require user approval. So they are opening it up to the payment idea where agents can make payments for you and stuff like that. It's currently available only for preview demo access. It's waitlisted. So you do have to apply to a waitlist. And only a very small portion of people in the United States are going to have access to that in the near term. And the whole idea is it's interactive. It actually checks in with you as opposed to you constantly checking in with it. And so we're seeing a broad shift from AI agents as tools to AI agents as collaborative partners that do check in on you and so on. And I think it's pretty cool, Jack. I like it. I like the proactivity thing. We were actually hacking this back in the day with NA10 where we would have scheduled Lipset kickoff every X minutes to review things. So you could actually engineer the proactivity, which has almost happened. They just put it behind these parts, right? The other interesting thing I'd call out with this, I think, is the fact that it's iOS or Android as well. So it's an app that you're going to be playing around with. Very cool. Uses your existing usage. Has all your retainer called memories. Very cool. You all called locked. This is one of the reasons why I think Grokbot is interesting because it is model agnostic, which is fab. The only challenge it's got is its memory. Yeah. In that sense. So dude, it looks cool. I mean, like, I wish, guys, we could tell you that we played with it. But honestly, we haven't because it is what is known as not available. So we'll have to see what it's doing. So yeah, it seems like an interesting step forward and we'll see what it is. I think just kind of taking this to a utility conversation, I think it's interesting to see what Anthropica are doing. Initially, this is what Claude Bart should have been. This was the initial Claude Duet like a year and a half ago. That was the whole idea behind it. So it's interesting. They've kept the cards quite close to their chest with this development. But I'll tell you one thing you can take away is that for sure they see the AI agent assistant as a big part of the future. Think about this. SpaceX SAI. You've got Anthropic. You've got OpenAI. You've got Meta. The only one we haven't seen, Nick, is probably Google. They're the only one who haven't dropped one of these things. Apple's got Siri, which is what this would be. And I'd probably expect that to be something you'd see in the next 12 months anyway. They're all putting their cards on the table with this stuff. They're all investing in it. Yeah, it's actually really interesting that you put it that way. I guess this is Claude Bart or what Claude Bart could have been. Interesting to note that Claude Bart was such a major security clusterfuck. And we're talking about security right now. And Anthropic released this right as Security Gate is happening. They're calling it Black October, by the way. Very interesting to see sort of the evolution of Claude Bart. I can't imagine this will have anywhere near as many issues as Claude Bart does, although there are obviously going to be unintended uses for Duet and unintended failure modes. Dude, you know what I just thought about? What? Right now, we're using AI as like this disembodied brain. And it's just like it's a cloud brain. And we send a request to it and we get something back. But nobody really actually wants that. What we want is we want our own personal assistant that is super customized to us and is just about us and actually works with us. And I think that's Claude Duet's angle here. It's like, no, this is your assistant. We're pulling it out of the cloud. I mean, I know it's still in the cloud. We're pulling it out of the cloud. We're putting it right beside you. I think that's the marketing pitch. AI that expands your ability, not necessarily just replaces you. I think that's an interesting take. For me, one of the problems I see now is I think AI has gotten really chaotic. I think it's a very messy mandrel is how I would describe it. There's batteries, there's car keys, there's the key to the garage. No one really knows exactly what everything is right now because everybody's splattered across so many different AIs right now. They're a little bit on dots. They're a little bit on cloud. They're a little bit over here, a little bit over there. So they're kind of spread across many different systems. Yeah. Yeah, it's true. And so centralizing all that with just one simple out-of-the-box assistant, I think, would help. Although the fact that it's a new iOS app and Android app is kind of annoying. Don't you just wish they centralized all this stuff? If this were like a book on the development of technology and stuff, I would not have expected it to be so many different apps and stuff like that. I would have thought, okay, the major operating system like Android and iPhone, they just release their own AI assistants and everybody just uses that. But it doesn't really seem like that's what's going on here. It seems like it's independent and individual model companies and wrapper companies creating these sorts of advancements ahead of time. And the hardware and phone and operating system layers, they just do their own thing and then we'd tack these onto them. But I'd be really interested to see a whole system designed from the ground up by Claude and Anthropic. Yeah, they're trying to be the... They're trying to add a layer, right? They're trying to be the control layer because essentially you're going to use these things anyway. You're going to use Slack. You're going to use Gmail. They just want to be the interface, the place, the conduit through which all this stuff happens. So they're adding a layer. Funnily enough, this is what we're talking about at Glider, the startup P.W. is we wanted to think about it on a layer level, like to be the voice layer that activates those things. And I think when I think about it now at Loudmic, these companies, realistically, the personal AI assistant holds a lot of power. Because if I can control all the apps that you use and all flows through me, it is unbelievably easy to monetize. You just have to own the control. You're basically owning the infrastructure. And I think more and more, the AI model itself is going to become a bit more of a commodity. I am probably on a weekly basis now blown away by how incredible Opus 5.5 is. Like to the point where I was even doing some graphic design today. And something that I probably would have had to have done manually in Photoshop or Canva. Nick, I literally... I'm freaking blown away by like how freaking ridiculous it is now. Where I was like, wow. I mean, it really is like insane to me. And bear in mind, every other model will catch up to that level within a short period of time. So realistically, I think to some extent, you get some sense of commodity. How do you define a commodity? Commodity is the same product you can buy everywhere. Like energy, for example, right? It doesn't make a difference where you get it. Honestly, I think like if they're not careful, they will start to commoditize. So then it's just where does work happen? And that's what they want. They want to have the best agent. One thing I'm not sure about yet, Nick, is if anybody's really cracked a commercially sound ROI model with agents yet. I don't know if anyone's making millions from them as of today. When you say commercially viable model with agents, like can you give me an example of what that might look like? Yeah. So for example, who has a personal AI system that is making the money? In terms of not the individual, but the business. For example, Darts dropped, but are they making money from Darts? Because really it just tags on to a subscription. Hermes agent was cool. They raised a crap ton of money. They do have their own plans. Like, oh, you could buy the $20 Hermes plan, the $50 Hermes plan. And I like that direction, but that was always going to be a little bit of a tough sell, to be fair. So it's like I don't know if we've really super duper cracked it yet. But I think the commercialization isn't really the hard part. I think really what we're looking for now is users. They're just looking for people and users. Hermes agent is a good example of one who really just captured the fiddlers, the engineer. And if you wanted proof that you couldn't make a business that's worth seven figures plus, just targeting fiddlers, Hermes agent, noose engineering is a perfect example of that. Because Mrs. Higgins, Nick, who works at the NHS, is not really using Hermes agent, right? It's builders that are using Hermes agent. I wonder how much money they make. It says they're reporting an annualized revenue of 36 mil, which, to be clear, just means they make about $3 million US a month. That was as of mid-September 2026, so a month ago. Do you think that's gone up or down since then? I would say down, right? Yeah. I'd say like... Yeah, because the headwind they have is that these darts and grok bar and duet are just too easy to use. And they're all people already using the app, so it's like an ecosystem. Yeah, it's fascinating. Looks like they raised $90 million in a Series B at a valuation of 1.5 bill, or roughly 1.28 billion euro. They're using euro because they're French, right? Yeah. Yeah. I mean, new research, right? That kind of makes sense. Yeah, I mean, it's interesting to see all this stuff because I feel like the frequency of market opportunities is going up. The potential magnitude of market opportunities is going up, aka any one opportunity can pay you out that much more. But the amount of time that they are available to capitalize on has gone down. And so you get more bigger, but shorter. And I think a tool like a wrapper around or some sort of like open claw-like agent that just has like far cleaner design and presumably some opinionated practices there, a framework for AI assistance per se, sort of makes that really clear. That's more or less exactly what we're seeing here, right? To be honest. I feel like you can make a ton of money, but can you sustain that money for more than like, you know, six months, 12 months? Oh, bro, it's tough, right? I mean, like, let's be fair, you know, you and I are brought on to news research. I say, Jack, Nick, we've been watching the Stack Podcast. We are one of the 17 and we need your advice on how we win and how we compete. One of the first things that I would say to them is that we're going to need some good money, to be honest with you, because ultimately I think what your market is, is just technical users who want personal assistance, right? And the issue, one of the headwinds that you have is that you're now fighting SpaceX SAI, you're now fighting OpenAI and also Anthropic, and they've all put their bet on the table, their agent. And they make it really, really, really freaking easy to use, right? You don't have to touch the terminal. To use Noose Agent, to use Hermes Agent, you actually have to go over here. And I'm not saying it's difficult, but I'm saying it's more friction than Grokbot is or that Claude is or that Dots will be, just because it's an app that they're already using. I do think there's space for that kind of build-your-own market, for sure. I do think that that is a space out there. But in a way, it's great because it validates the market with all these players. So obviously, it's very, very hot. But I think the market's definitely changed over the last six months, for sure. You know what I find interesting about all this? There's always going to be a contrarian market, aka there will always be some portion of people that don't want to use the Grokbots, let's say, just because it's easy to use. And so maybe that's who they're targeting, right? Three mil a month is not a tremendous amount of money, let's say. But presumably, the market for people that don't want to use like Cookie Cutter, Grokbot, or Muse, or whatever the hell, is more than $3 million a month. I mean, that market inherently is probably like 10 mil, 15 mil, 20 mil. So maybe they're just not even going for ease of use at all. It's just like, no, no, this is the hacker platform, right? They could say, dude, we're just going to own the top five technical 5%. That's all we're going to go for. We're not trying to be mass market, Magoo. We're really just going to be the technical version. So how do you make your products undeniable for clawed users, right? I guess like, yeah, it's going to be an interesting one to see, for sure. Because obviously, OpenClaw is kind of in that space, which is slightly weird because now they're wrapped in and around OpenAI, which themselves have a competing product. So that's an interesting one. Nick, we've got some questions. Jack, shall we do a couple of questions? I'm liking the very first one. And I know you normally ask the first question, but Jack, let me ask you this one. What model do you use for your shoulders? Dude, I use Opus 5.5 for every one of my shoulder workouts, okay? Every single one of my shoulder workouts. Little tip, by the way. I mean, I didn't follow this advice, but one guy, he was like looking after like Hollywood, what do we call it? Hollywood superheroes. He said, dude, if you had six months or three months with somebody and you had to get them as jacked as possible, what would you do? He said, I would spam lateral raises. Just spam lateral raises. Because one of the easiest ways to have a better physique is basically the ratio between your shoulders and your waist. That's everything. You're looking for 1.61 to one, the golden ratio. That's what you're trying to aim for. And the wider your shoulders get, the better the physique looks. So that's a little tip from the top, from Hollywood, baby. So just bear that one in mind. Thanks, Jack. Nick, we have some interesting observations about the Tesla thrusters being of particular use in the zombie apocalypse. Thoughts on that? Yeah, that'd be pretty fun. I mean, that was for a Tesla just built a flying car video, which is correct, by the way. Absolutely. Okay, how about this one? I work with many of the citizens in the UAE. They did not have to work at all, but they came into the office every day excited about their projects. Even if I get UBI, I will still be one of those people who will work on projects. I don't want to end up like one of those people in hover chairs from a WALL-E. Well said, Jack. I love that. Yeah, because that really is like, you know, the fork in the road. But how much of that is just like, people are always going to be like that? You know, I feel like there are some people that are just restless. They'll always want to be doing something. There are also other people that are like, reasonably content with whatever they have. Well, this is like the billionaire's paradox. Nick, isn't it? Where it's like, I saw a billionaire talking about this. He said that the irony is that to become a billionaire or very wealthy, you essentially have to have a certain series of habits. Like you work the weekends and you're really passionate about what you do. But strange, what happens is that turns you into a certain kind of person. And when you hit the goal, you're no longer the kind of person that can just relax any longer. Like you can't just stop working the weekends. So you have to relearn to stop. And they just don't because like, it's no longer about money. So ironically, I think the people that really, really, really want UBI is not going to change their productivity. The people already getting after it, it's just going to enable them to do cool things. So I really hope that UBI would encourage other people to do predictive stuff. I think the latest study showed that it didn't, unfortunately, but I really want it to be the guest. Yeah, I think some people always take advancements and opportunities like that as a positive and then a bunch of people won't. And that's just kind of how it is. Yeah, absolutely. We have some good questions about who's watching the Watcher's Watcher. Great question. It would just watch ad infinitium. Yeah, it's just Watcher's all the way down. I'm really interested in this chain of thought style on Astra. There's no spaces and some combination of words seem like one token only. Do you guys think OpenAI should be called out and pressured to change that? Jack, what do you think? Interesting. There's almost no spaces. A combination of words kind of seems like one token only. Well, we do that because I think we're trying to protect their model. They're trying to protect how they think. Because I think that is like the IP. Like you can maybe distill the answers, but you don't want to show the way the model's thinking of the protection. So do we need to show how the model's thinking? They don't have to show that if they don't want to. No. I would quite like to know it's thinking personally. But you can just ask it for that. Like how did you arrive at that decision yourself? Yeah, from like a monitorability standpoint, it would be nice if we could actually parse it out without needing to use other AI models in between the AI model and then us. Like it'd be nice if we didn't have to translate. I would like them to speak in English, even if it's less efficient. I feel like that's an algorithmic advancement that you could probably beat through scale. Like we should know what they're thinking about in English, right? That just makes sense. So even though it's less efficient and obviously less economic incentives for it, why not just like keep that aspect of chain of thought reasoning without the algorithmic improvement and then just make it bigger? Then we'll be able to more safely work our way through the weeds of superintelligence. And Nick, they're still king of the appers. Nick, we have some interesting comments on Vibe coding websites being destroyed by PHP. Nick, what's happening with Vibe coding websites right now in 2026, Matt? I don't think they're really destroyed because of PHP. But yeah, I mean, Vibe coding websites are definitely entering commoditized territory. The reality is you can just one shot a site. And so the site really is less so about like kind of like being really flashy and more so about taste. So for example, this is a website I made for myself. And I thought it's kind of cool. And, you know, this is functionality that like, I guess you could do, but it's also like a very particular type of taste, I would say. And so this is like a very simple one page website I basically made for myself with all my major blog posts and stuff like that. You know, back in the day, this sort of project might have taken me like a few weeks to really like make work really well. But now I do it in like five minutes. And I think it's just an expression of my taste more than it is really my capability because we all have the same capability nowadays. But it's cool. We are living in the era of a designer now. I think designer. You need to have the eye of a designer and the brain of an engineer if you want to win. You want to be the best. I would agree, Jack. Okay. We got time for more questions or is that it? We can't. Well, one little pre-question. Where do we get the glorious stacked hat? Guys, the stacked hats are currently on the way to Nick. So there'll be some stacked hats. But we may do a exclusive stacked hat merch run at some point. We'll have to think about that. So one key thing we have to mention, guys, is if you want to win a McNinney or cash equivalent, you're going to drop a comment down below. And on Friday, we're going to give one to you or cash equivalent in a very epic style. And if you want to be into it, go drop a comment. And if you like this kind of free thinking, no control commentary to get your head, be sure to subscribe. Have a beautiful day. And we will catch you in tomorrow's episode.