Keys set up. Passwordless sudo. claude --dangerously-skip-permissions. Feet on the desk. Four panels later, there's a mushroom cloud where the server room used to be.
The joke is the setup, not the agent. Every item in panel one is a reasonable convenience on its own. Stack them and you've removed every wall between a model's bad guess and your whole machine.
The real version already happened
In October 2025 a developer filed issue #10077 against Claude Code. He was working in a firmware project when Claude Code 2.0.22 ran what looks like rm -rf starting from root. He saw it and interrupted it. Every file he owned in his home directory was already gone.
Here's the part that matters. His log is thousands of lines of rm: cannot remove '/bin': Permission denied. /boot. /etc. The agent wasn't root, so the operating system said no. Ordinary Unix file permissions are the only reason the system files survived.
He says he wasn't running with skip-permissions. Another user in the same thread was, watched the same kind of delete start on his own machine, and hit Escape in time. He credits pushing to git for letting him sleep at night.
Now look at panel one again. Passwordless sudo takes down the Permission denied wall. Skip-permissions takes down the approval prompt. What's left is luck and reflexes.
Free range has fences
I let agents roam on a dev box, and I'd tell you to do the same. Production is a different pasture. The fences there have to be structural: its own identity for every agent, scoped privileges, a human gate before prod, sudo you have to ask for. Not a sentence in a system prompt. I laid out the whole setup in Your AI Agent Doesn't Need Root. It Needs a Leash.
Be stingy with passwordless sudo. Keep skip-permissions inside a sandbox you can throw away. Push to git like you mean it.
Don't be that guy.
If you want your agents set up this way, start with a free call.
— J.P. Howlett
