Definition
Plain language
Giving an AI agent the ability to type commands directly into a computer's command line.
As stated in the literature
Granting an agent the ability to execute arbitrary shell commands; a high-privilege capability that turns instruction-following failures into real-world actions and underlies the case for hard execution boundaries.
Why it matters: It turns a model's mistakes into real-world actions, which is why such power demands firm limits on what an agent is allowed to run.
For example, an agent with shell access could type a command that deletes files or installs software on the machine it's running on.
Heard on the show
“And they're arguing that the entire architecture of "behavioral guidelines plus an oversight agent" is structurally unsound for any agent that has shell access.”Episode 049 — An AI Agent Reached for Root in Twelve Minutes, Without Being Attacked