A key challenge in alignment research is measuring malicious behavior accurately and distinguishing it from simple misgeneralization