Question 4Cloud, deployment and security
How do you make sure an AI search tool never shows people documents they aren't allowed to see?
This is the first thing a customer's security team asks about. Getting it wrong ends the deal.
“Strong background in on-premise, private cloud, and multi-cloud environments (AWS, Azure, GCP), including experience with containerization and orchestration (Docker, Kubernetes).”Forward Deployed Engineer (AI Applications) · Seekr
The answerFour steps
- 1
Carry each document's permissions from the source system into the index
Say it like this
“When we index, each chunk keeps its source document's access list: the users and groups allowed to open it.”
Why it scores
Carries permissions into the index, where retrieval can use them.
- 2
Filter search results by the user's identity before the model sees them
Say it like this
“At query time we resolve the user's groups from their SSO identity and filter the search to chunks they can open, before anything reaches the model.”
Why it scores
Filters by identity at retrieval. A prompt rule is not access control.
- 3
Re-sync permissions when they change at the source
Say it like this
“Sharing changes all the time, so we re-sync permissions every few minutes, and when someone leaves, SCIM deprovisioning removes their access right away.”
Why it scores
Stale permissions are the quiet leak. Covers changes and leavers.
- 4
Log who saw what, and test with users who should be denied
Say it like this
“We log which documents fed each answer, and before launch we test with a contractor account that should see almost nothing.”
Why it scores
Tests the deny path with a real low-access user and keeps an audit trail.
StrongAll together
About a minute, out loud. The marked parts are the four steps; hover one to find it above.
Say it's AI search over SharePoint and Google Drive for a 5,000-person company. When we index, each chunk keeps its source document's access list: the users and groups allowed to open it.1 At query time we resolve the user's groups from their SSO identity and filter the search to chunks they can open, before anything reaches the model.2 The model can't leak what it never saw. Sharing changes all the time, so we re-sync permissions every few minutes, and when someone leaves, SCIM deprovisioning removes their access right away.3 We log which documents fed each answer, and before launch we test with a contractor account that should see almost nothing.4
WeakThe answer that fails
Sounds fine. Common. Scores low.
“I'd tell the model in the system prompt to only use documents the user has access to, and add a filter on the output that catches sensitive things like salaries.”
Interviewer's note
Prompt rules and output filters aren't access control. One clever question and the salary file is in the answer.
PracticeTwo minutes, then they push back
Your turn
How do you make sure an AI search tool never shows people documents they aren't allowed to see?
Answer like you're in the room: about two minutes, no notes. Say it out loud, even alone. It feels different from thinking it. Then check your answer against what interviewers score.