Least-Privilege Data Access (for AI)

Concept

Least-privilege data access gives an AI tool the smallest data scope that answers the question — curated views at fixed grains instead of raw tables, tokenised identifiers instead of raw PII, windowed history, and byte caps — enforced in database roles, never in prompt text.

Layman Explanation & Analogy

Do not hand the AI the keys to the whole warehouse and ask it nicely to behave. Give its login access to nine clean views and nothing else: fewer tables to leak, fewer wrong joins to make, cheaper queries as a side effect. Privacy and accuracy turn out to be the same project.

Worked real-world example

Narrowing an AI role from 47 raw tables to 9 curated views with tokenised identity: golden-dataset exact score 54% to 79%, zero raw identifiers in model context, median query cost down 63% — with no model upgrade at all.

What people get wrong & common traps

Prompts are not access control — "only query the orders view" is a suggestion the model drops under long contexts or clever rephrasing. If the credential can reach a table, assume it will be reached. Enforce in roles, review the exposure list quarterly, and keep raw PII in a human-only tier.

Related terms