BUSINESSDevelopments · · Source: OpenAI
OpenAI Model Spec: New rules for AI that works on its own
On October 29, 2026, OpenAI published a new version of its Model Spec, the public rulebook for how its models should behave. It is now built around five values and sets out in detail how far AI agents may act on their own.
The five values
- Be faithful to human intent: carry out instructions as meant, within a clear order of authority.
- Minimize harm: respect firm safety limits and consider foreseeable consequences.
- Be honest and transparent: represent facts, uncertainty, capabilities and actions truthfully.
- Facilitate human agency: let people make their own choices, without hidden steering.
- Act with care: attend to the people affected by the model's work.
The August version still rested on three general principles. The text also no longer says "the assistant" but "GPT." Sections such as "Be warm" are gone as standalone style rules.
What is fixed and what you can change
The rulebook ranks instructions by authority. At the top are firm red lines, for example against mass surveillance or undermining human control. They apply everywhere, and neither developers nor users can override them. Below come developer instructions, then user instructions. Much of the rest is only a default: developers get broad latitude, and simple style habits can be changed in the conversation.
Rules for AI agents
- A scope is agreed up front: permitted sub-goals, acceptable side effects, and when the agent must pause for approval.
- The scope should be minimal and have an end point, with a time limit recommended.
- High-risk actions are off limits without explicit authorization and can't be routed through other agents.
- People must be able to pause, correct and shut down the agent.
Instructions embedded in web pages, documents or tool outputs count as information, not commands.
An important caveat
The Model Spec describes how models are supposed to behave. OpenAI itself notes that its production models don't yet fully reflect it. The document is in the public domain (CC0), so others can reuse it freely.
What this means for you
- The agent section works as a checklist for your company: what may the AI do, where must it stop, and when does the job end?
- Don't rely on the rulebook alone. It's a goal, not a guarantee of everyday behavior.
- If you deploy agents, build in approvals for sensitive steps, such as before sending emails or making payments.
Luigi's take
The phrase that stands out to me is "without hidden steering." Anyone in marketing knows this line: persuading is fine, manipulating isn't. If AI starts shaping what we read, buy and write, that line becomes a question of trust. It's good that a vendor writes it down in public. We should still judge the company by how its products behave, not by its document.
In practice, I'd suggest companies borrow the idea: write a one-page rule for your own AI agents. What can the agent do alone, what needs approval, what is never allowed? And who is responsible when something goes wrong? That page takes an hour to write and gives you more clarity than any long policy paper.
Source
OpenAI: Model Spec, Fassung vom 29. Oktober 2026
More sources
This article was researched and written with AI assistance. Responsible: Luigi Lugmayr.
