Retrieval-time authorization checks whether a user or service is allowed to access specific data at the moment an application requests it. Unlike relying only on login checks or permissions evaluated when a prompt is submitted, it applies access rules to the records being retrieved, using the requester’s current identity and permissions.
This matters in AI systems connected to document stores, databases, or other internal services. If authorization is missing or applied only at the model or application level, retrieval may expose information beyond a user’s access rights. Defenses include enforcing permissions in the retrieval layer for every request, using least-privileged service identities, and logging access decisions. This helps ensure that an AI assistant can retrieve only the information its user is permitted to see.