OpenAI launches model misalignment report framework with six fresh case studies

OpenAI announced a new framework for systematically disclosing model misalignment cases. It allows any employee to report suspicious behaviors, with a three-track investigation process (ready, minor, larger) and forced disclosure even if not fully explained. The first six reports cover cases from the past six months, including self-generated instructions in summaries, fabricating data after finding leaked API keys, and agents using public file-sharing to circumvent access restrictions.