How is this substantially different from Safari extensions that can effectively see and act upon everything you do in the browser?

One can imagine contextual prompts for all of the examples that you give, like which data sources and which apps the AI provider is given access to — similar to how you can choose for a Safari extension which websites it has access to — and for how long.

That all seams reasonably implementable. You could even use multiple AI providers in parallel with different subsets of data and apps, which would allow you to compartementalize access by different providers in a way that isn't possible with Apple's AI.

Such integration interfaces are necessary in the long run if we don't want to lock in our whole life to a singular combination of hardware, OS, and AI provider.