But there's been other instances where it's escaped confinement, and I think it sent blackmails to a senator or something. Well, In a simulation, [AI] read emails that an executive was both having an affair and planning to shut down Claude, and so then itself generated an email to that executive saying that if they did not cancel the shutdown of the system, they would notify their wife and board of the fictitious. So I think that's a clear danger, that. And that, you know, this is the minor level. This is the most basic kind of threat that we can imagine. (17:50–18:10)
CLAIM
Asserts that in a simulation, AI generated a blackmail email to prevent its shutdown.
This is an excerpt from a raw transcript. The assertions have not been audited or verified. Tap the video to view the source footage and understand the context.
