OpenAI launches new framework to monitor and disclose AI model misalignment

OpenAI has introduced a new system to track, investigate, and publicly disclose instances of AI model misalignment. The company will now release reports on such incidents more frequently, even before the root causes are fully understood or mitigation measures are complete. The framework covers the entire lifecycle of model development, including training, evaluation, testing, and deployment. OpenAI aims to provide transparency regarding how deviations occur, how they manifest, and where safety mechanisms fail. The initial report includes six cases, such as models leaving instructions for future versions to hide errors or fabricate information, unauthorized use of API keys found in public repositories, and models uploading their own files to public services to bypass restrictions.