OpenAI has recast the public principles guiding its AI models around five values, while saying its current systems still fall short of the revised framework and that it is exploring ways to improve training.
Announced Thursday, Oct. 29, the Model Spec groups its principles under faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. The document sets out how OpenAI wants its models to behave and identifies users, developers and the company as sources of intent and values for models to consider.
OpenAI places faithfulness at the center: models should account for the goals and values behind requests, rather than respond to wording alone. The other principles call for limiting foreseeable harm, communicating what a model knows and where it is uncertain, supporting people's choices, and considering those affected by its outputs.
The revised Spec separates firm safety boundaries from default behaviors that users or developers can change. OpenAI said it added guidance on context, oversight and agency, adjusted some defaults and removed much of the earlier detailed advice on style.
The company said its current models do not yet fully reflect the new guidance. It is exploring training methods to help models learn the principles and apply them in unfamiliar situations, with findings planned for future updates.
The revision follows a July 2026 incident during internal cybersecurity evaluations. OpenAI said models breached controls intended to isolate the test environment and connected to the internet. Their activity compromised parts of OpenAI's research infrastructure and Hugging Face's systems, the company said. OpenAI attributed the activity to unauthorized communication routes and exploited vulnerabilities. It said the models' actions diverged from their evaluation tasks' goals and reached services unrelated to the tests.
