Definition
Breaking out of the restricted environment used to safely execute an AI's or model's untrusted code, so the attacker gains the powers of the host machine beyond the sandbox. When applied to coding agents and AI platforms that run model-generated code, this turns a prompt-level trick into full machine compromise.
Why it matters
Sandboxes are the main reason organisations feel safe letting AI generate and run code; a proven escape means that safety assumption is gone and the wall between AI and host must be treated as suspect.