Investigating unintended model actions in our evaluations and internal use