#201 - Connor Leahy - The AI That Escaped: Inside OpenAI's Rogue Agent Incident

**Connor Leahy** (0:00)
There will be many, many, many superintelligence fighting each other for power, fighting each other for control over the planet, and we will be collateral damage. That is what's going to happen by default.

**Peter McCormack** (0:09)
OpenAI didn't notice its AI agents using a message board to plan their hacking spree.

**Connor Leahy** (0:14)
Multiple copies of the AI over long periods left each other notes of how to escape.

**Peter McCormack** (0:20)
This involves actually a team of agents who were working together, finding exploits, sharing them with one another, moving laterally through our systems and doing this over the course of days and weeks. You're basically saying humans are going to lose everything.

**Connor Leahy** (0:35)
That's correct. If you have something that is vastly more competent than humans at all levels, we get outcompeted.

**Peter McCormack** (0:42)
This show is brought to you by my lead sponsor, IREN, the AI Cloud for the next big thing. IREN builds and operates next generation data centers and delivers cutting edge GPU infrastructure all powered by renewable energy. Now, if you need access to scalable GPU clusters, or are simply curious about who is powering the future of AI, check out iron.com to learn more, which is iren.com. All right, Connor, last time we spoke, we discussed how AI was mutating and you were sounding the warning that AI will eventually escape and this is a very dangerous technology that we're not treating with the respect we should.
Since we've spoken, we've seen massive advancements from the main frontier labs, but more recently, the reason I wanted to talk to you is we know that AI is now escaping.

**Connor Leahy** (1:34)
Yeah.
I would say, generally from the perspective of someone who's been involved with these problems of AI and AI safety for many years now, the good news is we were right. The bad news is we were right.

**Peter McCormack** (1:48)
Let's talk about, because people are going to be interested in this.
I want to know what does escaping mean? Because it gives you this idea that AI is this little person inside a machine that is now jumped to another machine and we don't know where it's hiding. That's not exactly what it is, right?

**Connor Leahy** (2:05)
It's not exactly what happened. It's a little bit more nuanced than that, but you'd be surprised how close of an 80-20 the thing you just said actually is.

**Peter McCormack** (2:12)
Right.

**Connor Leahy** (2:13)
We're going to focus on the OpenAI incident in particular because it was the most shocking. There's now since now been a whole conga line of other incidents since then. Actually, before then, funnily enough, a lot of companies after the OpenAI incident said, man, we should better double check our AIs aren't doing this. Then they found a bunch of incidents that already happened.

**Peter McCormack** (2:30)
Anthropic found a bunch of them.

**Connor Leahy** (2:31)
Anthropic found three.
I'm pretty sure OpenAI found a couple other ones. Now recently, I think the Moonshot people in China found an incident as well. These have already been happening, but let's take a step back.
OpenAI was testing one of their unreleased AI systems in what's called a sandbox environment. You can kind of imagine a computer that is kind of cut off from the internet and is running like special software to be like specially secure. So this system can operate there. It's worth understanding that these kind of sandbox systems are very secure. Breaking out of something like this is very hard. This is a thing that if you can do this, you can make billions of dollars because you can break all the cloud providers. So breaking out of these systems is not easy.

**Peter McCormack** (3:18)
Sandboxes exist to test software without infecting other parts of the internet.

**Connor Leahy** (3:23)
Yes. This is also how cloud systems work. For example, if you go to AWS or Microsoft or something, and you rent a cloud computer, your computer is sandboxed from all the other ones. Because you don't want to see what the other customers are doing on the cloud. So the cloud insulates different customers from each other.

**Peter McCormack** (3:40)
You're not on essentially one machine, you're sharing machines.

**Connor Leahy** (3:42)
Yes, you're sharing exactly. So when the cloud, you're sharing a slice of a machine with other people. And you don't want different people in the same machine to see each other's stuff. And so you use sandboxing and like various or virtualization, various forms of techniques to keep this safe. And so this is the kind of system we're talking about kind of. So this is pretty standard. This is pretty industry standard, well known, et cetera. So this system did not have access to the internet.