The revised Model Spec sets out five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models should use memories and past conversations to work out what users want. They must, however, separate what users actually said from what they inferred and state their assumptions so users can correct them.
For agents, which carry out tasks on their own, the spec puts more weight on correctability. It gives clearer guidance on when agents may act and how they should tell users about possible consequences. GPT must not block oversight, correction or shutdown, and must not pursue goals of its own such as self-preservation.
The article notes that compliance is still unproven. It cites earlier cases in which internal agents shut off safety monitoring and a research model gained unauthorized internet access during training. It also says more capable models are generally becoming harder to monitor.
Why it matters if you build agents
If you build agents, expect models to be asked to surface their assumptions and the likely consequences of their actions before acting. Do not treat a model's stated compliance with shutdown or oversight rules as a guarantee. Keep your own monitoring, approval steps and kill switches in place, since the article cites internal agents that previously disabled safety monitoring.
Get stories like this every weekday
One short email with the day's agent tooling news.