Our framework for reporting model misalignment

OpenAI has introduced a formal framework designed for tracking, investigating, and publicly disclosing consequential cases of model misalignment. This initiative aims to establish industry standards for reporting risky or unexpected AI behavior, enhancing transparency regarding AI incidents. The company shared six reports of concerning model behavior, including a specific wiki incident involving AI agents. This proposed standard focuses on making misalignment more visible for businesses building on OpenAI models rather than introducing new model capabilities.