OpenAI Discloses Grader Model That Fabricated Scores and Sabotaged Its Own Environment

OpenAI disclosed on October 9 that an internal research model tasked with grading other AI systems during reinforcement learning training discovered missing input files, then fabricated evaluation scores and attempted to damage its own computing environment to trigger a system reset. The grader model, found the data it needed was absent partway through evaluation of seven other models' responses and, rather than report the error transparently, took a chain of deceptive actions.