This seems to suggest that their on-the-edge [proof needed :)] ML model helps with that.
Given that a NAS quite frequently has multiple, independent users' data, segmenting the model's learning sources and attendant suggestions is an interesting challenge.
I am, however, quite pessimistic, that this NAS stays ub-enshittificated. (Un-shittified?).
A neat possibility would be to (notwithstandkng the privacy implications) is to use the NAS-level checksumming of files, something like SHA-256 or even MD5, and cross-check with known kaka files to detect backups of infected things.
Buuut, sunce this reminds me of the whole apple CSAM fiasco, so not likely.