Transcription: OpenAgents Episode 263 - Bitcoin Wallets Under Attack

Media source: AI Agent Attacks Expose Bitcoin Wallet Risk.mp4 Source cut: 00:01:45 to end Edited media: AI Agent Attacks Expose Bitcoin Wallet Risk_Ep263_X.mp4 Transcript runtime: 00:14:50.93 Timestamp basis: edited video Transcription model: gpt-4o-transcribe-diarize Generated at: 2026-07-31

Machine-generated transcript, reviewed against a prompted transcription pass for product names and technical terms. Verify wording against the video before using it as quote-grade source material.

[00:00] Christopher David: All right, I'm back. Let's see what the news is. Oh, shit. Worse than any previous Bitcoin exchange hack.

[00:12] Arguably one of the worst things that could happen.

[00:16] All right, let's go through it.

[00:17] Popular hardware wallet Coldcard has a faulty random number generator. Funds are being drained from ordinary users' hardware wallets as you read this.

[00:27] Exploit most likely discovered with the help of AI. Yeah, so I'll have some thoughts on that. Likely already multiple attackers competing for the affected BTC. Many more will join very soon. Nobody knows how much BTC is affected. Nobody knows how many users are affected. This attack has only just begun, and it will continue until all affected BTC has been drained unless users secure their funds first.

[00:51] Inevitably, many Bitcoiners will not hear about the incident early enough to respond in time.

[00:57] I am truly saddened. Yes, I am as well, for everyone affected, especially those who may have just lost their life savings.

[01:03] The worst part is that they did everything right. I don't agree with that last part.

[01:13] Here's my comment. Except following influencer consensus, "buy Coldcards," is not doing everything right if everyone ignores the 800-pound gorilla hurtling at them.

[01:26] What did Coldcard or anyone else do to mitigate against AI capabilities that have been dramatically increasing for years?

[01:36] So this is a type of attack, like AI-augmented attack, that you're going to see more of.

[01:43] Coldcard apparently was the limit of the damage. Like, all of the damage from this attack seemed to be limited to Coldcard users because they had some insecure entropy.

[01:55] But that insecure entropy has been in the codebase for years, and, like, people haven't caught it. Like, why? Why have they not caught it?

[02:04] So I'm just going to kind of sift through some of my reactions, and then we'll go into some proactive things that we could maybe do to help the situation.

[02:15] Note: on Bitcoin Twitter over the last year, I saw a thousand-ish posts about quantum hardening and approximately zero posts about hardening against advanced AI.

[02:29] That seems bad.

[02:34] So I said, who's taking care of that?

[02:53] Before we get to more of my reaction, here's Harsha. Harsha's got good takes on this from the Coldcard folks, kind of on the not-so-great chain of response from the Coldcard people.

[03:11] Yeah.

[03:18] Yeah, so here's my question.

[03:20] Has anyone built or aimed agentic fuzzers at top Bitcoin OSS projects? We'll do it if no one else is, but really there should be better people than us to do it. Where's all that grant money going? Yeah, secure. We gotta secure against, against, against stuff, but, like, what about AI? Who is securing against AI? I ask.

[03:52] So I said, who is hardening Bitcoin OSS projects against AI? Like, this has been a train hurtling toward us for years.

[04:00] AI has gotten better and better. It's only going to keep getting better and better.

[04:05] Who's hardening Bitcoin projects against AI? Why have I heard approximately nothing about this?

[04:14] Ah, Michael from Lightning Labs says Spiral, project Loupe. Okay, let's see what they've done.

[04:18] Meet Loupe: AI-powered vulnerability scanning for open-source Bitcoin. Yeah, the asymmetry between attacks and maintainers shouldn't decide Bitcoin security. Yeah.

[04:29] Nice. Cool.

[04:31] Yep, scan. Yep. Yep. Internally, okay. Loupe is designed to scan. Okay. Oh, what do we got?

[04:43] Several well-known projects have already committed to being part of it: Bitcoin Core, BDK, LDK, rust-bitcoin. Okay, so the Spiral ecosystem. Cashu, Blockstream Jade, bitcoinj. Okay.

[05:00] So what has come of that? Cool.

[05:04] So why didn't Loupe catch the Coldcard problem? What is not being done that should be? Because this is an AI attack. You'd think that maybe this would have caught that.

[05:16] Probably because it's opt-in, which is understandable.

[05:20] Anyone is free to use Loupe to conduct security scans, but they'll need to bring their own model access tokens. Okay.

[05:25] On the latter question, that's a big one, and we're in a completely new paradigm. As a community, we have to put a lot of thought and effort into finding the answer to that question over the next few weeks and months. I know. Yeah.

[05:35] Yeah, I agree.

[05:36] Nice. So the projects most nervous about security, the ones that need it most, would be the least likely to opt in. Did Coldcard opt in to this?

[05:48] I'm guessing not. Maybe they did.

[05:50] Seems like that's the good-cop approach: hey, let us help you with your security. And now you need someone else to play bad cop, a.k.a., I found some problems with your shit.

[06:01] Users are depending on you to not ship broken garbage that can get their life savings stolen.

[06:06] Here's the vulnerabilities that we found. And then you responsibly disclose that to them. You know, use the standard practices of disclosure.

[06:15] But who's doing that?

[06:19] Apparently no one.

[06:22] More initiative with proper disclosures. So, yeah, maybe, I mean, no one's going to do it.

[06:29] If no one else is going to spin up a little miniature fleet army of agents to play white-hat hacker, then maybe, you know, the Bitcoin AI agent lab has been deep in the weeds of agents for three years. Maybe we should do something about it.

[06:53] But I'd love to hear from Loupe, which still hasn't posted anything.

[07:01] This is an ongoing situation.

[07:03] So Loupe, maybe you've got some preventative stuff going on, but this has been a situation going on for 24 hours, so hopefully some sort of AI-based response would be nice.

[07:18] Here's what's being done to, like, proactively harden and investigate other projects. I know, like, Block has a team doing stuff. Cool. And they've put out some good stuff.

[07:30] It feels like something should be done additionally.

[07:38] So we're going to do a bad-cop, white-hat agentic fuzzing operation coming soon to Bitcoin OSS near you. DM if you can help or advise.

[07:52] We've already got some people stepping forward to help us. That's nice.

[07:55] Agentic fuzzing. First of all, what's fuzzing?

[07:59] Fuzzing 101.

[08:02] Go ahead, Google. An automated software testing technique that floods a program with random, invalid, or unexpected inputs to find coding errors and security vulnerabilities.

[08:11] So, I mean, there's fuzzing and then there's, you know, other pen testing. But yeah.

[08:16] So agentic fuzzing: a bug-finding approach seeded by historical bugs in which deep agents perform the reasoning directly.

[08:25] Yeah. An advanced software testing method that uses autonomous AI agents as primary reasoning engine to discover complex logic bugs and security flaws. Yeah.

[08:34] Instead of relying purely on random code mutations or static rules, these deep agents analyze historical bugs, form root-cause hypotheses, and generate proof-of-concept tests across codebases.

[08:47] Yeah.

[09:01] So Cobra says it'll be a summer of AI attacks on cryptocurrencies and blockchains now that we have Fable-level open-weight models.

[09:09] Potentially billions of dollars in value sitting in insecure protocols or wallets that any script kiddie can direct an LLM to audit and exploit.

[09:17] We are cooked. Well, not necessarily, Mr. HODL. It works both ways.

[09:21] Or we could always use Fable and Mythos to balance it out. Yeah.

[09:25] I'm not gonna have access to Anthropic stuff. I'm probably on their shit list.

[09:36] But are attackers using Claude? Maybe they're using open models. I don't know.

[09:45] So Coldcard was kind of the canary in the coal mine because they had the most insecure design.

[09:52] They, like, chopped off what should have been the full entropy, so it was, like, kind of secure but not really secure, which is, like, bad.

[10:01] But at least this can get the industry to wake up to the need for AI hardening.

[10:13] And how big was this?

[10:18] Oh, some Coldcard wallets were drained.

[10:21] Mr. Coldcard, NVK, says it's FUD. The guy loaded from a different wallet. Mess of a history of events. Someone is collecting the seeds they pre-generated.

[10:35] So this, I suppose, turned out to be 100% false. Not anywhere directionally correct at all.

[10:45] So this is, uh, that's, like, that's, like, not a good thing to read.

[10:52] Could be. It's super weird, but also not just one user. If that was true, it would be tens of thousands.

[11:02] And this was deleted. So, yeah, tens of thousands.

[11:09] And what began as sort of a, yeah, it's just one device type, expanded to, oh yeah, it's actually everyone.

[11:17] And now they're saying, essentially everyone's saying, like, get off Coldcard. So get off Coldcard. That's a great kind of short-term mitigation.

[11:28] Probably bad that everyone's moving coins around.

[11:52] Trusting a single external unverified source of entropy with life savings is not doing everything right.

[12:05] Block doing good research. Looks like they might have found the guy because they used a paid blockchain analytics service.

[12:13] But they said that they thought this guy was a novice.

[12:19] But he exposed an attack that other people are replicating, so now every hacker team is going to be on this, and also smelling blood and looking for vulnerabilities elsewhere.

[12:27] So it kind of seems like we have a coordination problem in the Bitcoin ecosystem, because there's no coordinated effort for systemic agentic AI defense.

[12:54] The Loupe thing seems like a good idea, but it's clearly insufficient, and something else is needed.

[13:03] Something else is needed.

[13:11] Well, we're going to be actively in the inquiry of this. We were kind of in the middle of a product rollout, which we're going to keep doing anyways.

[13:21] Basically a Nostr-powered IDE fork of Zed for managing agents. It's got a built-in NIP-29 chat.

[13:30] We're kind of porting over some of the code from Armada so we can be interop with their secure Discord-kind-of chat.

[13:37] And I think we'll get with them and do a group channel on Nostr for coordinating defense, and get some agents going,

[13:53] and set up some proper disclosure channels and stuff. But I'd like to see agents fanning out and responsibly identifying any vulnerabilities in other projects.

[14:02] And with Nostr's encrypted chat channels, which, by the way, my company I started four years ago, that was our first product. We did a bunch of encrypted Nostr group chat stuff. We wrote one of the first drafts of a group chat channel NIP. The thing got superseded by--I wrote NIP-28, by the way, public chat--which got superseded by NIP-29, which we're now using.

[14:27] So I think that's going to be a cool way to organize some kind of defense here.

[14:34] All right, just thinking out loud here. We'll put some thoughts together. If you are interested in participating in our defender fuzz operation, shoot us a DM.

[14:49] See you soon.