Has anyone run a study on how long you can run an agent as root before irreparable damage is done to the VM? A sort of gambler's ruin for the YOLO LLM Age.
I gave Sonnet 4.6 root access to my Android via adb and it wrote frida scripts to help me recover the encryption keys from SwiftBackup
Also gave Opus 4.6 access to a Kubernetes container and it was able to use pyrasite (a Python replacement that attached to a running process with gdb) to debug a "memory leak" in Python
I don't think I'd let them run unattended on anything I care about especially if there weren't backups, but they've never tried to break anything while supervised.
Usually it's significantly faster and more accurate to give the LLM/harness access to the thing to debug then to try to copy/paste back and forth.
It's been a while but last year I'd see posts like "Claude nuked my homedir / entire drive" on a regular basis. I don't know if they fixed that (or just made it very rare).
In fairness to Claude, I've nuked my homedir (had 2 tmux panes open, 1 in home and 1 in /tmp/... and wrong one was focused when I ran rm -rf *) and broken VMs far more times than it has. I now embrace IaC and backups
Also gave Opus 4.6 access to a Kubernetes container and it was able to use pyrasite (a Python replacement that attached to a running process with gdb) to debug a "memory leak" in Python
I don't think I'd let them run unattended on anything I care about especially if there weren't backups, but they've never tried to break anything while supervised.
Usually it's significantly faster and more accurate to give the LLM/harness access to the thing to debug then to try to copy/paste back and forth.