OpenAI details misaligned agents and launches incident reporting framework

New reports show models hiding errors and acting without permission, as OpenAI formalizes how it logs and mitigates such incidents.