A research team built working prototypes, proposed a new way to classify privacy choices, and audited 30 platforms to find out how well they let humans set privacy rules for agents acting on their behalf. The short finding: most platforms give humans no way to express what data an agent may share, with whom, and under what conditions. The gap is not a technical limit. It is a design choice no one has made yet.
The research separates two things that current products blur. One is granting an agent access to a system. The other is telling the agent what it may do with the information it finds there. Humans can often do the first. Hardly any product supports the second. An agent can read a calendar, a document store, or an inbox, and the human who granted that access has no way to say which parts stay private, which can be quoted, and which can be forwarded to another agent or service.
The audit of 30 platforms turns this from a theoretical gap into a measured one. Product teams building agents that read personal or organisational data should treat privacy preference setting as a first-class design problem. The research gives them a taxonomy to work from, not just a critique to absorb.