Provenance separation is the practice of keeping different sources of text or commands distinct inside an AI system: system instructions, direct user input, and untrusted external content such as webpages, files, or search results. The goal is to prevent lower-trust material from being mistaken for higher-trust instructions.
This matters in cyber security because prompt injection often works by smuggling hostile text through content the model later reads as context. If a webpage, deep link, or document can influence the assistant without clear boundaries, that content may steer answers, trigger unsafe actions, or contaminate memory features. Defenders use provenance separation to label, sanitize, and constrain retrieved content, and to require confirmation before any consequential action. In practice, it is a core control for assistants that browse, summarize, or act on behalf of users.