Why enterprise AI assistants overshare, and the control that stops it

In short: enterprise assistants overshare because they retrieve from everything their index can reach, not from what the person asking may open. The control is permission-aware retrieval: search as the user, fix loose sharing before indexing, keep sensitive collections out, log what was retrieved, and treat retrieved text as untrusted. A two-person test before rollout shows whether it works.^1,2

One of the most common AI data incidents inside organizations is not a breach from outside. It is an internal assistant answering a question with information the person asking was never supposed to see: a salary file, a draft restructuring plan, a client’s confidential terms. The assistant did nothing wrong by its own logic. It searched what it could reach and summarized what it found.