> they should be filtered too (not "blocked", it will always be probabilistic)

The article’s point is we haven’t precisely studied what those probabilities are.

No it isn't actually, those kind of low level technical questions have been somewhat studied. It's the end to end interventions which have been poorly studied.