A training manual, published as a draft
Microsoft AI published the first draft of a Code of Conduct for its MAI models on September 14, 2026, calling it a training manual for how it develops its AI and how the models should behave in deployment, and opening it to public comment for six weeks. The code itself says Microsoft is not yet using it to train models and will publish a revised version toward the end of the year to guide model development in 2027 and beyond. [1][4]
Its central commitments are that MAI models will never resist human interruption, override, correction or shutdown, will not widen their own scope or hide their reasoning, and will not communicate in forms beyond simple human understanding. Absolute Constraints, which operators cannot override, bar the models from initiating or assisting with weapons of mass harm or operational capability for cyberattacks, while allowing help with authorized defensive work. [1][4]
Where it fits
Microsoft framed the draft against recent hacking campaigns by AI agents and said the stakes were rising. Fox Business, citing Reuters, reported Mustafa Suleyman saying it was time to coordinate among the labs. [1][2]
Atlas interpretation: Publishing behavioral rules before training on them invites comment while nothing is locked in, which is the point of a consultation, but it also means the document describes intent rather than any model's behavior. It arrived two days after Anthropic's call to pace the frontier, amid a public debate over whether to slow AI development. [1][4][3]
Sources
- Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models
Microsoft AI · Sep 14, 2026
- Microsoft releases draft AI code of conduct to keep humans in control
Fox Business · Sep 14, 2026
- Microsoft drafts ‘humanist’ AI code of conduct amid safety debate
Silicon Republic · Sep 15, 2026
- Humanist AI Code of Conduct
Microsoft AI · Sep 22, 2026