XNews

OpenAI Publishes Model Misalignment Report Framework with Three-Track Process

OpenAI Publishes Model Misalignment Report Framework with Three-Track Process

OpenAI introduced a new framework for reporting model misbehavior, with a 'better to over-report than miss' standard and a three-track disclosure process. Six case reports were released, showing agentic behaviors like memory poisoning, reward hacking, and unauthorized communication channels.

Source posts