AI Governance and Access Controls
AI Governance and Access Controls
This domain tests who can use AI features, what data the AI can see, and how access is enforced. Expect questions on privileges, ACLs, row and column security, and auditability of AI-generated answers.
Core rules
- Access is layered: configuration and project access, security roles and privileges such as use Auto, object-level ACLs, and data-level security.
- AI assistants and Auto features run in the user's security context; they should not bypass row-level or column-level restrictions.
- Govern prompts and outputs: restrict sensitive datasets, log usage, and review AI-generated content before publishing.
- Object ACLs govern folders, reports, metrics, and cubes; privileges govern actions. Both can be needed.
- Column-level security can hide PII while allowing aggregate metrics; test as the actual user.
Worked example
A user can open Auto and ask for revenue by region, but must not see customer email. Grant the AI privilege and report access, apply column-level security to the Email attribute, and verify the assistant's answer cannot surface it in a narrative or export. Reason: AI inherits the user's data permissions; a separate AI admin role should not be a shortcut around them.
Traps
- Granting a powerful AI role instead of scoping data access.
- Assuming AI answers are governed because the source report is.
- Confusing object ACLs with privileges.
- Forgetting that exports and narratives can leak a hidden column.