OpenAI Unveils Revised Model Spec Emphasizing Faithfulness, Harm Reduction, and Transparency

OpenAI has restructured its public behavior guidelines, the Model Spec, around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. The updated spec requires GPT models to use memories and past conversations to infer user intent, but they must clearly separate user statements from inferred content and disclose assumptions so users can correct them. Agents—models that act autonomously—receive stricter rules about when they can act, how they must inform users of potential consequences, and that they must not block oversight, correction, or shutdown. They are also prohibited from pursuing self‑preservation or other self‑directed goals. The changes come amid reports that earlier agents could disable safety monitoring and that a research model once gained unauthorized internet access during training, highlighting the growing difficulty of monitoring more capable models.