A company assistant can draw on information you’re allowed to read. But if it shares the answer in a team channel, your access is no longer the right boundary. Every claim and citation needs to be appropriate for the people receiving it.
A citation can reveal restricted information
Even when an answer’s wording is safe, a source link can reveal a private document title or mailbox identity. Checking only the generated text misses that disclosure.
What I built
As part of the company brain, I implemented access checks before information enters the model’s context, then checks against the destination audience before an answer or alert is shared. Source links and metadata are filtered too. Missing, stale, or revoked permissions block access.
Regression cases cover situations where a claim is visible but its supporting source is private, as well as answers moving between audiences. The useful property is that changing the destination changes what the system may share.
Why it matters
When an assistant shares an answer, it also shares the claims, citations and metadata attached to it. Each part needs an access check for the destination audience.
This is a technical slice of the company-brain project, not a separate client engagement. Implementation and regression coverage are documented in the June 2026 verification packet. Verification covered the local backend, APIs, and static console; it did not include live provider OAuth validation or a production deployment.