Authority drift is the gradual shift of real decision-making power from people to automated systems, without a corresponding change in accountability. In cyber security, it matters because an AI tool may start as an advisor, then become a default chooser, and finally trigger actions such as approvals, routing, account changes, or downstream API calls. The workflow still appears human-governed, but the machine is effectively making the call.
This drift creates security risk because errors, prompt injection, bad data, or overbroad permissions can lead to unauthorized actions at machine speed. Defenses focus on separating recommend, decide, and execute; applying least privilege; logging overrides and exceptions; and keeping a manual fallback path. Good governance also tracks which material decisions the system can influence, who owns them, and when automated authority must expire or be reviewed.