ariana_azarbal comments on Generalisation Hacking: a first look at adversarial generalisation failures in deliberative alignment