Was eading up on the use cases for this, one of the big ones is mixed or augmented reality. But they don't give developers access to the cameras yet. There isn't really much you can do with it outsode of games like the Oculus but they aren't marketing it as such.
They're not giving direct access to the cameras, but developers don't need that raw data. Developers still have access to the data the cameras see. The choice to do this probably stems from a security perspective. This way there's no possible malicious app or site that can actually record you in an intimate environment like your home. It's the same way ARkit has worked for years on the iphones and ipads.
That's a very good point re: privacy. ARKit does work really well for scene detection like the room shapes, plane detection stuff etc but I was mostly thinking about the VisionKit stuff too (not the VisionOS ones) wheee you want to be smart about things. E.g. if you want to be able to detect text in mixed reality (like real time translation) or if you want to recognize objects not just by their shape.
It is going to be a privacy nightmare I agree but it is also severely limiting the possibilities of the software right now