Comment by ectoloph
1 day ago
Is it naive to assume that the agent will try and access anything on your disk, either accidentally or maliciously?
Permissions classifiers in auto mode are just models trying to guess if they're doing the right thing.
Claude Code will tell you that it went around a sandbox because the sandbox blocked it. At which point, you ask yourself the point of the sandbox.
I always put the agent harness in an ubuntu-based Docker, with a /workspace folder where it can work and occasionally some other stuff mounted as read-only. The LLM server itself (llama-server) is running on a different more powerful computer on the local network, connected through Tailscale so I can also use it away from home.
I honestly don't trust these things to not accidentally mess something up, otherwise.
Now I think it's still technically possible to break out of that with some clever hacks? But the moment I see a model even vaguely considering that, I will never run it again.
(I don't use Claude but currently Qwen3.8 27B)
You need to treat agents as an independent user you're allowing on your machine.
Give them their own account. Give them only the access you want them to have. If they "hack" around that, do what you'd do to any other malicious user: kick them off.
It's not a sandbox if you can just snap your fingers and wish your way out of it.
It's not naive it makes running these ai agents inside the sandbox even more important
Not naive at all, which is why there are so many AI sandboxes: https://pleasedonotescape.com/