But it would be very hard to spot new weights of the neural network and a new list of target hashes. These are pretty much guaranteed to change regularly as they retrain the embedding network and/or change the list of known target images. So it will be very hard if not impossible to see what they're searching for. That latest update could just add the ability to recognize CSAM pictures that had meme texts added to them. Or it could change the embeddings and target list to spot unlicensed posting of copyrighted still frames from movies. Or it could now retrieve any picture of people in police uniform. No way to know if you don't have a hunch and a targeted picture you want to test with.
Do you know this for certain about any vendor? If a company the size of Apple were pushing the same update to everybody, then it would likely be known about by the world pretty quickly. But... if a targeted signed update is sent to a handful of selected devices, that's harder for the world to find out about. It's risky, but it's definitely technically possible.
Could they write and deploy something overnight that hoovered up everyone's data? Maybe on iOS, less so on OS X, but now they're going to ship with that capability. I don't understand how some people can't see the difference between "they could always push some nasty update," versus literally shipping hardware and software with a backdoor.