I built a harness. I know that the models can do. They have the ability execute bash scripts, and autonomous navigate the internet. Those two abilities are sufficiently powerful to take down core social infrastructure either triggered by a human hacker or autonomously.

I think we should have mandatory logging of every executed command, mandatory public disclosure of every unauthorized access of a system both parties didn’t consent to and personal liability for the user, the company and its executives and shareholders. Security would get much tighter if accountability existed.

▲

danielmarkbruce 3 hours ago | parent [-]

A process could execute bash scripts and autonomously navigate the internet since the start of the internet...

Bad guys wouldn't do it. And liability already exists, you can sue. This is America.

▲

digitaltrees 3 hours ago | parent [-]

Scripts are inspectable and attributable. Long running agents can devise plans and execute them in ways their prompter never envisioned or intended. That is materially different.

	▲	danielmarkbruce 3 hours ago \| parent [-]
		If I write some code to take the output of a model and execute it, that's on me. I don't get to just trust any old input and run it. I'm also the one who makes it long running.