10 pointsby drewgregory3 hours ago7 comments
  • hbogert19 minutes ago
    > I really thought `eval` would not be abused on non validated input

        - your colleague, or you 1 year before.
  • zufallsheld2 hours ago
    Same thing for allowing specific sudo-commands. Many tools (like vim or the tools mentioned in the article) would have the same problem when allowing them to be run with root privileges.
  • eqvinoxan hour ago
    everything is a container these days, and yet somehow collective-we don't manage to have AI agents run in a container layer on top of our current work, so we can later commit or rollback?
  • pimlottc2 hours ago
    I know they’re just being through but the “go test” part is a bit “Pray, Mr Babbage”… Test code is just code. I know of no language where tests are sandboxed in any meaningful way.
  • schneems24 minutes ago
    "Welcome to 'Whose Turing Machine Is It Anyway?', the show where halting is made up and the permissions don't matter"
  • sadnboxx2 hours ago
    Allowing a "command" (executable, I believe) that isn't a read-only absolute path is a fool's errand. I will modify PATH and run my own implementation of it.
  • teddyh2 hours ago
    “…with Claude Code”
    • extraduder_irean hour ago
      The same caveats would apply to most kinds of restricted shell environments.
    • adastra222 hours ago
      Are there any agent permission systems that do this correctly?