When an AI agent can run arbitrary shell commands on your machine, the question is not whether it will make a mistake, but how you contain the blast r