In the database, on every table, through row-level security. Not in the user interface, and not in the description an AI reads before calling a tool.This is worth stating plainly because it changes what a bug can cost. If a permission check lived in the application and someone forgot one, the data would be exposed. Here, forgetting a check means the query returns nothing — the failure mode is an empty page, not a leak.
An AI client acts as the person who signed in, with exactly their role. The server resolves the user, organization and membership from the verified token before every tool call; it never takes an organization or user id from the model.A model cannot ask for another organization’s context, however it phrases the request, because it is never asked which organization it wants.
Assistant
Responses are generated using AI and may contain mistakes.