cross-posted from: https://lemmy.ca/post/71956921
I cannot comprehend how dumb this all is, AI has passed the event horizon of supermassive willful stupidity.
This is the future my human skills were framed as worthless for??..?
You must log in or register to comment.
- Joe@discuss.tchncs.deEnglish4 days
It is dumb to just let a non-deterministic prediction machine loose with shell access and important credentials.
You either: verify changes, and verify commands against restrictive whitelists, falling back to manual approvals (that you carefully check) before letting them run, or…
You only enable YOLO in a sandbox with temporary/traceable limited access credentials (still some risks, but muuuuch lower) AND audit resulting code or…
You only give a limited set of tools that can do no real harm and can be audited after the fact… but this is more typical of assistants and predefined workflows than coding agents.
The catch: Good sandboxing requires a security mindset and technical knowledge to not screw it up… it’s just too easy to give agents more access than they actually need, and even the experts screw up too often. It’s the wild west out there…
- duckCityComplex@lemmy.worldEnglish4 days
Still, it seems the best you can do is reduce the blast radius. The underlying problem - that LLMs are fundamentally untrustworthy - has no solution. I suspect we’re going to see a million variations on this kind of attack.