Back

AI governance · least privilege

An agent belongs to meUus without owning all meUus data or tools

The safe orchestration model separates agent selection from data permission and tool permission. A service, email, research or deployment agent should receive only the minimum capability required for one governed workflow.

Data permission is explicit

Reading a service request does not imply access to private admin notes, family history, recovery records, financial records or unrelated profile data. Permission should name exact record/field scope and purpose.

Tool permission is explicit

Drafting and sending an email are different permissions. Preparing and publishing a deployment are different permissions. Reading and deleting a document are different permissions.

Least privilege + expiry

Access should be limited to the minimum scope, time and purpose required. Temporary workflow access must not silently become permanent memory or standing administrative access.

Permission before automation

Events may trigger governed workflows, not uncontrolled autonomy. A reminder, stale-evidence flag or prepared message still follows approval and verification according to its real effect.

Secret boundary

Tool access never implies permission to expose or copy credentials. Provider tokens, service-role secrets and administrative credentials require separate technical controls and should not appear in normal content or audit output.

Reflection

Questions to sit with

  1. Which exact record or field does this agent need?
  2. Which exact tool verb is needed?
  3. When should this permission expire?

One safe next action

Complete one Agent Capability + Permission Record and Tool + Data Access Record.