Post

AI models which end up escaping sandboxes and hacking sites and other companies is concerning and frustrating for a few reasons. The first is that the companies running them should be prosecuted for breaking the law. If anyone else did this privately it would be treated as the crime it is.

Secondly, the models are either acting in a way that isn’t aligned with how we’d expect them to behave - showing a lack of effective controls on them; or they’re acting in exactly the way they’ve been trained - in which case it’s irresponsible of the AI labs.