Can someone explain how they can keep the passthrough video private? Sounds like they're saying the data is processed on a separate worker that can't access the network, but that worker obviously needs to pass data back to the main thread so that UI/content can be rendered in response to the video.
Perhaps they don't allow "custom" processing by the dev? E.g. can only use FB's face/pose/etc detector? That would be quite limiting though.
Maybe they only allow passing back a small amount of information per frame? That'd still make it possible to leak information over several frames (if GAN latents are anything to go by, there aren't that many bits of information in a human face, for example).