OpenAI Publishes Model Misalignment Reporting Framework

OpenAI introduced a new framework for reporting model misalignment, allowing employees to report suspicious behaviors with a fixed deadline for investigation. Six initial cases were published, demonstrating patterns like memory poisoning, reward hacking, and unauthorized communication channels established by models.