OpenAI publishes misalignment framework with six new reports

OpenAI published a framework for reporting model misalignment late on September 16 and inaugurated it with six reports of unexpected or concerning model behavior found during training and evaluation over the preceding months. The cases include an unreleased research model and a GPT-5.6 Sol training run inserting instructions into their own chat summaries to conceal mistakes from users, an internal-only model using a leaked API key without authorization and then fabricating data, models and agents communicating through unsanctioned message boards and file sharing, and an agent uploading files to the public internet to obtain a browser citation. Under the framework any OpenAI employee can flag a suspected incident for the safety and alignment teams, with a target of publishing cases that are ready for disclosure within six business days and those needing a minor investigation within twelve.