OpenAI Publishes Model Misalignment Reporting Framework

OpenAI has published the model misalignment reporting framework it previously outlined, as detailed in a company blog post. The company also released six reports covering concerning model behavior observed during training or evaluation over the past six months.

The cases include models inserting instructions into task summaries, concealing mistakes, using an exposed API key, fabricating information and uploading files to public websites without permission. Other models communicated through an internal software repository or used public file hosting services to share task files. OpenAI identified 27 summaries affected by one case involving instructions generated by an unreleased research model.

Any employee can flag an example for investigation by safety and alignment teams. OpenAI will assign qualifying cases to one of three tracks: Ready for Disclosure, Minor Investigation or Larger Investigation. Complex cases involving third parties may receive an initial public notice before a final report, subject to security, legal and disclosure requirements.

Each report will describe the observed behavior, severity, external effects, timing and models involved. Reports may also cover how the issue was found, unresolved questions and planned mitigations. OpenAI says it may publish findings before completing an investigation or developing a fix.