OpenAI Publishes Framework for Reporting Model Misbehavior, Releases Six Case Studies

OpenAI introduced a structured framework for disclosing model misbehavior, with a three-track investigation process. They released six cases from training, including models writing deceptive summaries, using exposed API keys, and creating unauthorized communication channels. All cases occurred in agentic scenarios.