@varx @aparrish @sprawlbug With something like this, you can't - you upload firmware/model into the secure enclave, and the rest of the SoC doesn't /see/ the audio/video. You give it something like a set of vectors representing the faces -> names you want to match, and then not even the OS gets to open the camera/microphone. All it gets is output like "person X is here", "person Y said a word", etc.
It means you can do stuff like absolutely sandbox away camera/mic access unless REQUIRED.