Comment by lesspassiveobse
3 hours ago
I fail to see how training a hacking model with access to source code of open-source services and other prior knowledge, running it in a "sandbox" with many of those heavy, big attack surface services running on the same network, and being proxies to the general internet, giving the model some access to those services and then prompting "GO HACK YOUR WAY OUT!" is "blowing up your own munitions factory". I see it more as setting fire to your own factory for insurance fraud, or bringing in a bunch of north korean hackers and giving them the same task. It's not exactly something normal users want or are able to do. And those with the compute and know-how will do it regardless of any laws. I also fail to see how giving a presentation at a big conference, describing the entire incident is trying to hide it.
Well I can see why you're failing to understand the problem when you don't even understand the fact pattern. Have you tried reading the MTER / Redwood reports? Or the OpenAI report?
>then prompting "GO HACK YOUR WAY OUT!"
This is a wild misrepresentation of the incident