Embedded AI Evaluators Shift From Fringe Idea to Near-Consensus

Apollo Research published a post arguing that external AI evaluators need employee-equivalent access to labs to catch safety failures, and said it has been building toward becoming an embedded evaluator. The argument: most recent AI safety incidents occurred during model development and internal evaluations, before public releases, which is where current third-party evaluations are focused. Apollo said key safety questions require access to people inside labs, not just access to models. Separately, Dean Ball wrote that the concept of third-party embedded evaluators as a pillar of frontier AI policy went from largely absent in American policy discussions 18 months ago to near-consensus, and noted that organizations now considered potential evaluators were initially skeptical of the role when he first pitched the idea to them.