Item — HN Reader

I have time machine and just let them fly with --dangerously-skip-permissions on my Mac. Worst thing it's done is back up a database, delete the database, and then run git clean locally which also wiped out the backup, so I'm not saying there are no dangers but honestly I've made worse mistakes and probably more frequently so I generally trust Claude with about the same level of access as me now.

Most common is deleting files etc but if you're using git and have backups it's barely noticeable

2 comments

Yeah I've got hourly backups out to multiple remote servers. My dev machine is in essence fungible. If it gets hosed, I'll wipe the drive and drop a good backup in. If it catches fire, I'll pick up a different machine and drop in the good backup.

I have more important things to waste my time on than writing absurd sandboxes to run AI agents without guardrails in. What even?

How are you going to notice that while working on ~/projects/acme3000 it for some reason deleted ~/photos/2003/once-in-a-lifetime-holiday/?

Backups are great when you know you need to restore.

I could ask this question without AI. How are you going to notice that while you were working on ~/projects/acme3000, you for some reason deleted ~/photos/2003/once-in-a-lifetime-holiday/?

Of course, AI is not a real person, and it does make mistakes that you or I probably would not. However, this class of mistake—deleting completely unrelated directories—does not appear to be a common failure mode. (Something like deleting all of ~ doesn’t count here—that would be immediately noticeable and could be restored from a backup.)

(Disclaimer, I’m not OP and I wouldn’t run Claude with —-dangerously-skip-permissions on my own system)

Isn't the problem that of finding out a consistency heuristic? For example, test that the resulting state is consistent with your test suite.

If it is a directory that gets deleted, then you can diff it with a previous state. If you don't control the state and don't know the surface area that you should observe, then yes, you're inviting trouble if agents run amok.