Yeah this has always been the glaring blind spot for most of the "AI Safety" community; and most of the proposals for "improving" AI safety actually make these risks far worse and far more likely.