Florian_Dietz comments on Auditing language models for hidden objectives