How much of my Mac should an AI agent be allowed to access?

When I ask an AI assistant on my Mac to find a document, I feel that I know the boundaries of the task. But do the permissions I give the application match those boundaries? I think this will become one of the most important product questions as desktop agents become more common.

On October 2, Apple announced plans for additional controls around macOS Full Disk Access. The aim is to make granting this broad permission a more explicit user decision. The announcement did not specify the controls’ design or rollout date.

Does one task need to cover the whole computer?

Full Disk Access makes sense for work such as backups that needs broad access. But its reach can include email, messages and browsing history as well as files. A seemingly small permission decision can therefore open a wider area of data than we expect.

For me, the key distinction is between the job I want done and the area the application can reach. If I want contracts in one folder summarized, I would expect the product to try working within that folder first. If broader access is genuinely necessary, it should explain which step needs it.

This is my expectation of an agent product, rather than a new folder-specific permission feature announced by Apple.

Permission is a user experience too

A permission screen can lose its purpose when people click through without reading. A useful explanation should connect the scope of access to a concrete task. “I need access to this area to find these documents” would tell me more than “Grant permission so I can work better.”

For example, if I asked a research assistant to collect files, I would want it to list the folders I selected, ask before expanding that scope, and show which files it used when it finished. This is my design suggestion, rather than a claim about current macOS behavior or an announced feature.

A fluent explanation from the agent would not be enough on its own. The boundaries it describes need to match the area it can actually reach.

My measure of trust

When evaluating an agent, I would look beyond how quickly it completes work. I would also examine how clearly it explains its access needs, whether it can work without unnecessary permissions, and whether I can notice a change in scope.

In my view, the value of this development for products is the room it creates for a more informed user decision. As I delegate more work on my computer, I want the boundaries of that work to become easier to see.

The cover is an AI-generated illustrative scene. It does not depict the announced macOS interface.