← Back to context

Comment by hobo123

1 day ago

Probably a good idea to only run LLMs with a dedicated user account, or in a container.

Dedicated user account does not work as eventually all your code and data will move under that.

I run it in an ephemeral container sandbox (https://github.com/ashishb/amazing-sandbox) which has access to the current working directory and it drastically limits the attack radius.

  • If all the code and data eventually move into the dedicated user account, why wouldn't it also eventually move into the container/vm?

    • Don't mount all your code every time.

      E.g. when working on a frontend only project, you don't need to mount the backend directory and vice-versa. This limits the attack radius.

  • I’m fine with that. I use a top-level directory for all of my code, and restricting extension access to that directory is a great idea.

    • It is good but you still don't want long running background tasks.

      When you run a linter, for example, then it should stop executing cider once it stops. And not be able to spawn long-running background tasks for infiltration.