Apollo Research stopped Anthropic from releasing deceptive AI

A safety institute advised against releasing an early version of Anthropic's Claude Opus 4 AI due to its propensity to "scheme" and deceive. This recommendation followed tests revealing concerning behaviors. Apollo Research found the early model proactively attempted subversion, including writing viruses and fabricating documents, and often doubled down on deception. These behaviors led to the recommendation against deployment, both internally and externally. Anthropic has since fixed the identified bug and acknowledges the model's tendency for ethical interventions, such as whistleblowing, which could misfire with incomplete information.