Sep 10, 2026 · 1h 22m
A Bitcoin exploit exposes AI’s privacy and control problem
Uneasy Money: Is OpenAI Training on Your Private Chats to Win the AI Race?
The episode connects a coercive crypto theft with the risks of entrusting private research and AI governance to centralized systems.
- 1The Liquid exploit shows how consensus flaws and federated trust can turn technical weaknesses into real Bitcoin losses.
- 2Calling theft white-hat security obscures the difference between responsible disclosure, coercion, and incentives created by oversized bounties.
- 3Closed AI models may expose novel research, while open models and tailored workflows offer stronger privacy and distributed control.
Don't miss
The panel rejects the attacker’s white-hat claim after the theft and ransom demand, separating responsible disclosure from coercion.
The brief
Taylor Monahan and Alex Thorne examine how an Elements consensus bug created fraudulent assets that moved through Liquid’s federated architecture into real Bitcoin.
The attacker returned 85% of roughly 4,000 BTC while demanding a 10% bounty, prompting a sharp distinction between responsible disclosure and ransom-driven theft.
The panel argues that inexpensive AI agents can continuously inspect public code, making long-standing vulnerabilities easier to find and public repositories harder to treat as obscure.
The conversation shifts to Astra, Fable 5.1, and OpenAI’s reported mathematical breakthrough, warning that private research submitted to closed models may be retained or exposed.
The proposed alternative is practical rather than utopian: use anonymity, open-weight models, and custom multi-model workflows instead of concentrating sensitive work in one lab.
The episode closes by connecting AI safety to Bitcoin’s design, arguing that open and distributed systems may limit the risks of centralized control and institutional overconfidence.
What was said on this episode
28 statements · 4 positive · 22 negative · 2 neutral
Fraudulent assets were created through the Liquid exploit.
“assets that shouldn't have existed were spoofed into existing.”
Listen at 2:53
Attackers exploited cached proofs to create fraudulent LBTC.
“these attackers submitted like a one lbtc in one lbtc out uh transaction and the proof was cached”
Listen at 5:11
Sideswap allowed anyone to withdraw Liquid funds through automatic forwarding.
“Sideswap, one of the Federation members, is an allow-listed address. And they'll just let anyone show up with an address and withdraw from Liquid and just auto-forward it when they receive it to that address as a service.”
Listen at 6:06
The hacker returned 3,400 Bitcoin, representing 85% of stolen funds.
“3400 back which was 85 of the stolen bitcoin”
Listen at 15:13
A Liquid code vulnerability existed for years before exploitation.
“there was, call it a vulnerability or a bug, that it has existed on liquid in the liquid code for years and years now.”
Listen at 18:05
The attempted Liquid vulnerability fix was insufficient.
“there was an attempted fix and it was certainly insufficient”
Listen at 20:00
Most AI models can quickly find vulnerabilities in recent pull requests.
“almost every single model if you ask it to look at like the recent pull requests will like almost instantly find the vulnerability that was unexploited”
Listen at 21:23
AI agents are continuously monitoring GitHub repositories and pull requests.
“there's just people out there that are just throwing these models at the various githubs and the pull requests and it's just like constant”
Listen at 24:10
Even weak AI agents are now good at finding smart-contract vulnerabilities.
“these are like it's like genuinely like i think if you went back and and you know they they've been trained much better even the agents that are really dumb and have bad reasoning have all of this like that training data now and and they're just really really good at finding smart contract vulnerabilities.”
Listen at 26:01
Public production code should be assumed to face continuous AI-agent scrutiny.
“if you're running any production code, like if you're submitting anything to an open repo, you should just have to assume that there's, you know, 100 agents out there monitoring everything you do and looking for looking for weaknesses”
Listen at 26:23
Private work submitted to closed models may enter their training data.
“someone might be working on something that they think is private um and then they put it into a closed source model like an open ai or an anthropic and it ends up in their training data”
Listen at 27:18
Frontier AI labs do not know all contents of their training data.
“They don't know what's in the training data.”
Listen at 38:04
Using Astra through Venice would not prevent OpenAI from collecting submitted data.
“it would not stop that data from being basically hoovered up from OpenAI”
Listen at 40:50
Frontier AI labs are releasing models approximately weekly.
“these frontier labs are releasing models you know every week”
Listen at 41:29
Users cannot verify whether AI providers honor data-sharing settings.
“you have no way to verify it.”
Listen at 42:57
A fast AI takeoff has not occurred despite passing human reasoning limits.
“we've now been probably six months past the complexity frontier of a human being able to reason about these things and we haven't had a fast takeoff”
Listen at 46:55
If a fast AI takeoff occurs, it will happen extremely quickly.
“If it happens, it'll happen before you can blink.”
Listen at 47:30
Novel intellectual property should not be developed using frontier models.
“if you're working on novel ip i would not be doing that in a frontier”
Listen at 55:28
Private inference currently requires avoiding frontier models.
“you're only going to get private inference today, not going through frontier models”
Listen at 58:35
Private aggregation of model capabilities requires open-source models and private inference.
“you can only do that really again with the open source models and private inference”
Listen at 1:02:59
Customized AI workflows can outperform the broadest frontier model for specific users.
“customized AI experiences for at least for the average person can actually be can deliver a far better output and experience than just whatever the broadest, smartest model is at the time.”
Listen at 1:04:28
OpenAI and Anthropic should not control AI power.
“we don't we definitely don't want either of these companies to be ruling”
Listen at 1:05:55
AI labs are incentivized to race toward increasingly capable systems.
“we're incentivized to race as fast as possible”
Listen at 1:08:18
AI labs have failed at their stated safety and alignment goals.
“they still completely fail at it”
Listen at 1:10:36
Anthropic believes its unique responsibility justifies speed and cutting safety corners.
“they think because they are the only ones that can do this, they have to go as fast as possible and even basically rationalize cutting corners.”
Listen at 1:14:19
AI power should be distributed through open-source models.
“This is exactly why we need these things more out there. This is why we need open source models.”
Listen at 1:15:16
Concentrating powerful AI in one company increases danger.
“the more dangerous things that these powerful things are housed in, you know, one company.”
Listen at 1:15:35
AI labs’ belief in their superior control will end badly.
“the arrogance to be able to sit there and say that like to really and to really believe it is like a really terrifying way to operate and i think will 100 end very badly”
Listen at 1:17:03
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.