In a new paper published Thursday titled “Auditing language models for hidden objectives,” Anthropic researchers described how custom AI models trained to deliberately conceal certain “motivations” ...