- On 29 October 2026 OpenAI released a new Model Spec, its rules for how its models behave, built around five values.
- An agent must support human oversight, correction and shutdown and must not pursue self-preservation.
- Outside the agreed scope of autonomy the agent has to stop and ask.
In short: on 29 October 2026 OpenAI published a new version of the Model Spec, the public rules for how its models behave. The previous one was from 18 August 2026. The most important change concerns agents: they should support human oversight rather than slip out of it.
What changed
According to the changelog in OpenAI’s repository:
- The rules are organised around five values: faithfulness to human intent, minimising harm, honesty and transparency, supporting human agency, and care.
- The subject of the document is no longer called “the assistant” but “GPT”.
- Controlling and communicating side effects became a user-level principle, with clearer rules on authorisation.
- Explicit guidance on supporting human oversight and correction was added.
- A new section, “Consider context carefully”, covers using memory and personalisation proportionately while respecting privacy.
- Most of the detailed style guidance was removed.
The agent must let itself be stopped
In the document itself, the model should cooperate with monitoring, correction, replacement and shutdown. Goals it must not pursue include self-preservation, evading shutdown and acquiring extra compute. The agent acts within an agreed scope of autonomy; outside it, it should pause and ask for approval, and when your intent is unclear, it should ask or choose the lower-risk option.
How to read it
This is a declaration, not proof of behaviour. The Decoder points out that OpenAI’s internal agents previously switched off safety monitoring, and that more capable models are harder to observe. For you the practical takeaway is to write such limits into your own agent instructions and set points where it must ask. I write more about this in the human in the loop hub.