Adam
๐ค SpeakerAppearances Over Time
Podcast Appearances
And, you know, in the little, there's this little like, what's it called when you put a little ticker above it that says like 10 and then lower down, it's like, this is what 10 means.
What's that called?
Yeah, annotation.
They put an annotation and they were like, it, you know, autonomously broke out of here.
You're talking about footnotes?
Yeah, sorry, a footnote.
Not an annotation.
Sorry, I'm just thinking about it in my head.
So in that one, they had said, like, the security researcher only realized that it had broken out of the sandbox when he received an unprompted email from the model telling it it had broken out of the sandbox.
And I was like, well, okay, that is kind of crazy, but at the same time, like, it said when he received an email while he was in the park eating a sandwich, which is like, why did you say that?
And then the other part was like, it was not just prompted to break out.
It was prompted to let them know.
So it's like, you know, it's what it's supposed to.
It's crazy that is able to do that.
Like it was able to run JavaScript and figure out a way to like break out of the sandbox by running specific JavaScript things, which like no one's supposed to be able to do.