OpenAI updates its Model Spec around five behavioural values

What Changed

OpenAI published a new Model Spec dated 29 October 2026. The document is organised around five values: be faithful to human intent, minimise harm, be honest and transparent, facilitate human agency, and act with care. It defines an authority chain from root and system instructions through developer and user levels, and says assistant and tool messages have no authority by default. It also calls out untrusted third-party data, the scope of autonomous actions, side effects, and the tension between immediate instructions and longer-term goals. The spec is published under CC0 and points to a changelog and earlier versions.

Why It Matters

Agent builders rely on model behaviour that is often described only in marketing terms. A public spec gives teams a vocabulary for prompts, evaluations and incident reviews: who has authority, what counts as an untrusted instruction, and when a model should ask rather than act. The emphasis on tool messages having no default authority is directly relevant to prompt-injection defences in browser and coding agents.

Practical Impact

Use the spec as a test rubric, not as proof that a deployed model follows it. Add cases for conflicting developer/user instructions, retrieved content that contains commands, and tool actions with irreversible side effects. Make trust boundaries explicit in your own system prompt and tool permissions rather than relying on the model to infer them. Record which model and spec version you evaluated, because OpenAI says production models do not yet fully reflect the document and the spec changes over time.

Sources

• An updated Model Spec for more capable and beneficial AI
• OpenAI Model Spec (2026/10/29)

Related games

Q1K3UNDERRUN

Keep reading

GitHub Copilot Fusion routes each coding task across multiple models

Copilot Automations can now be shared, billed and owned by a repository

Cloudflare open-sources Void to connect Vite apps to Workers

GitHub Copilot arrives in JetBrains Air with local and cloud agents

Chrome DevTools adds safer controls for coding agents

Node.js 27 alpha lands V8 15.6 with new memory defaults

Node.js schedules high-severity security releases for 11 November

Claude Opus 5.5 targets agentic coding at a lower cost per task

Godot 4.8 beta 2 fixes regressions and raises the .NET floor