Reduce disclosure risk by limiting the source set, checking the intended reader's access before retrieval, and testing allowed and denied paths. Treat uploads and cached copies as separate access decisions. Use supported field restrictions or exclude mixed-sensitivity files, scope agent actions, and retain available evidence with its coverage limits. These checks do not prove that nothing can leak.
1. Establish source and copy boundaries
Choose sources and connectors that your account actually supports. A connected source, an indexed or cached representation and a manually uploaded file can have different access and refresh behavior. Check what is stored or indexed, including derived text or embeddings, and how edits, deletions and permission changes reach each copy. Establish where each copy lives and which permissions apply before adding sensitive information.
2. Govern at the point of retrieval
Before returning information, confirm the requester and evaluate the access rules that apply to the retrieval path. Test search, direct reads, source links and cached results with intended and restricted accounts. Include a permission change and revocation so you can observe propagation rather than assume an immediate update.
3. Establish field restrictions
Where the system supports the required field or section boundary, test it with harmless mixed-sensitivity material. Otherwise exclude that document or provide a separately approved version. Do not assume that an allowed file automatically has usable field-level controls.
4. Check access-record coverage
Retain the records available for the operations you use, then inspect which events and fields they cover. Treat logging as best-effort unless stronger coverage is demonstrated. Metadata can still be sensitive even when document text is excluded; restrict record access and review it before sharing.
5. Put guardrails on agents
Autonomous agents act faster than people and at larger scale. Give them explicit limits, human-in-the-loop on sensitive actions, and a kill switch. The same governance that covers people should cover agents.
6. Be able to prove it
Bring evidence of the specific paths and scenarios you tested, including denials, revocation and cache behavior. Explain missing fields and unrecorded operations. Tamper evidence can support record integrity. It cannot establish that nothing leaked or that every event was captured, or recall information already returned. On-chain anchoring remains on the roadmap.