Auditing language models for hidden objectives