Permissions are part of the answer

A relevant answer can still expose a restricted record. Check what the reader may access before source material reaches the model.

Access is checked before source material reaches the model.

Start with the source of authority

A company may hold general procedures and restricted records in one search service. Membership is not permission to read every record.

Microsoft documents a security-trimming pattern where identity filters search results, and it works only as well as permissions are maintained. Read Microsoft’s security-trimming documentation.

Enforce access before restricted text reaches the model

Do not send unauthorised passages to the model and trust instructions to keep them out of its response. Apply the decision first, to titles and snippets as well as the answer.

A retrieval service may examine candidates before filtering; that alone is not disclosure. Verify where protected material could cross a boundary: a model provider, a log, another user’s cache.

Permission changes have a path through the system

A person can leave a group after a document is indexed, or a cache can retain an answer from an earlier decision. Design for both as routine events.

Document how identity and cached results update, and set revocation behaviour the business requires.

Test with people who should get different answers

Use approved test accounts at different access levels and ask each the same question.

  • Confirm a restricted record never appears in the answer, snippet or citation for a reader without permission.
  • Remove access, then repeat the query, cached path included.
  • Verify a retrieved instruction cannot grant access or widen tool permissions, and check that logs and support exports open no second route to restricted content.
  • Confirm source links still enforce the source system’s own rules.

Keep corpus approval distinct from reader access

An approved corpus means the organisation selected the material, not that every user may see every document, or that a filter covers retention and incident response.

Agree who owns identity, who owns permissions and who investigates a failure before building. AURA LOGIC LABS treats those as part of Knowledge Systems and Production Governance.

Plan access to your sources.

Bring your access rules and the sources they protect.