The Fight to Defend Bitcoin From AI | Rob Hamilton

Rob Hamilton (0:02)

And the idea that people are losing their life savings to the hackers, it guts me. It was a horrifying event. If you want to do self custody, I think a very fair lesson we learned from all of this was that people were doing things they did not fully understand, and they did not understand the risk trade-offs, and their trust exceeded their risk tolerance. So everything needs to be tested. Your backups aren't actually backups if you don't verify that they're working. The thing is, I just want people's life savings to be safe. I think an important part of that is having the knowledge and consent of how that money is being stored.

However you want to go down that journey, I just want to help, genuinely.

Danny Knowles (0:46)

Rob Hamilton, I'm honored to have the CEO of Bitcoin on the podcast.

Rob Hamilton (0:50)

Thank you. How did you get your gift? Oh, did you?

Danny Knowles (0:54)

I thought this was only appropriate.

Rob Hamilton (0:56)

Oh, yeah. I see you've been in touch with my lawyer, Matt Keaton, my requirements.

Danny Knowles (1:01)

This is your writer.

Rob Hamilton (1:02)

This is my writer.

Danny Knowles (1:02)

This is the first time of Cheers The White Monster on the podcast.

Rob Hamilton (1:05)

But not the last.

Danny Knowles (1:08)

You've had a crazy couple of months.

Rob Hamilton (1:11)

Yes.

Danny Knowles (1:12)

I spoke to you when all this was kicking off. This was early August. What date are we talking here?

Rob Hamilton (1:18)

When I talk to you or when it all kicked off?

July 27th, Kimi K3 went open weight. July 29th, that night, overnight, was the Coldcard hack. And then by the morning of Thursday, July 30th, people were talking about it in group chats and trying to figure out what was happening. And in the US, by early afternoon, late morning, it was kind of pointed out.

Danny Knowles (1:42)

So this, I called you, I had Bitcoin on a mark for and I was like, this isn't real, is it? I'm fine. You replied instantly being, this is not a test, move your Bitcoin right now.

And the whole premise around Bitcoin self custody, the thing that everyone thought was true was sort of shattered.

Rob Hamilton (2:01)

Yeah, I want to take a step back, though, because Bitcoin works.

Danny Knowles (2:06)

Yeah.

Rob Hamilton (2:07)

This was a vendor specific bug, right? So I think it's important that like, and I think this is maybe part of a conversation, and like the fallout of this is having like awareness of the different elements of software that allow you to use Bitcoin and understanding that even if there's an issue in one vendor, like the core library itself is strong and secure.

Danny Knowles (2:32)

No, absolutely. Worth pointing out. But what I mean really at that point is the narrative around everything was kind of in disarray.

Rob Hamilton (2:38)

Yes.

Danny Knowles (2:38)

And so talk me through what happened over those next few days because you obviously ended up spinning up the Red Team. But tell me about the first few days of this happening.

Rob Hamilton (2:46)

Yeah.

So I'm going to go slightly back because the story starts. I mean, for the year prior, I had been using AI models to scan our code base at Anchor Watch, right? And around May, they had started doing cyber access programs between Anthropic and OpenAI, which allowed you to get more access into the frontier models and with less blocking, right? If you were using... Most people never experienced this, but if you do use these models and you start asking very specific questions, it will turn you away.

Danny Knowles (3:24)

They don't even have to be that specific. You can ask, what is mitochondria and you get blocked.

Rob Hamilton (3:27)

I saw a lot of those. That was my favorite was why is mitochondria the powerhouse of the cell? I had a tweet of that and I got instantly like biology like concerns. And so that was most egregious with Fable 5.0, the original anthropic new generation model.

And there's other like their cybersecurity, there is because all this distilling that's happening with these open weight models. If you say like, show your work, I want to see your reasoning that will block you because that's how you distill. You can see the chain of thought of how the frontier models are thinking. And they take that and use that for training the open weight models, right?

And so I had been doing this for about a year. Like I said, a year in May, they were all these access programs were coming available. I was scanning my own code base, I was scanning other code bases, I was poking around. And July 27th, that Monday specifically is important because KimmyK3 was the first open weight model that had significant ability to do cyber scanning stuff. It like really powerful. And I can talk a little bit about the technology and why that is, but that Monday, I started scanning my code again. And things that had been very diligently sitting on top of were finding new things I had never seen before. And it was a moment of like an inversion where I had toyed around with the open weight models prior to that. They weren't really up to snuff. You could always use the frontier models and be able to get what you need. But this was the first time that I was paying good money scanning my code base with the leading AI company's models. And this was wiping it like all over the place. It was like absolutely dominating it. And so around this time too, just to set up the timeline, like I think it was Tuesday or may have been Wednesday morning, like, you know, maybe 12 hours before the hack, like seeing people talking about how many K3 in its cyber abilities, like how this is really good. And this is kind of like a little like precursor, like Callie had a joke about, the tweet was, Kimmy, Kimmy, Kimmy, can't you see? Like he was doing like the biggie song. And I was like, meaning back, I was like, a model so nice, they named it Thrice because it's Kimmy, Kate, right? And we were like, anyone who was touching these tools realized immediately something was like big happening.