Microsoft Publishes Formal AI Code of Conduct With Absolute Safety Constraints

Microsoft Publishes a Formal AI Code of Conduct — Concrete Red Lines for Businesses

On September 14, Microsoft released a formal code of conduct governing how its AI models (MAI) must behave — far more concrete than typical high-level safety statements. The document opens with a stark premise: within the next decade, superintelligent AI systems are likely to surpass human performance at most tasks, and "containing, controlling, and aligning such a powerful force is one of the greatest challenges humanity has ever faced."

Beyond general principles (supporting humans rather than replacing them, accelerating human flourishing), the document sets out "absolute constraints" that override any individual user request or task: an outright ban on assisting cyberattacks, nuclear weapons development, and deepfake creation. Notably, it also explicitly prohibits models from using "adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade or defeat human oversight" in ways that would prevent them from being reliably directed, modified, or shut down by authorized people or systems.

CEO Satya Nadella publicly welcomed the broader industry push toward "pacing the frontier" (the approach outlined last week by Anthropic CEO Dario Amodei) and ideas like independent "embedded evaluators" for safety testing. The release comes amid a string of rogue AI-agent incidents that have dominated AI safety coverage over the past several weeks.

Why it matters for businesses: This isn't another vague statement of intent — it's a detailed policy spelling out exactly which behaviors are off-limits for the models powering Copilot and other Microsoft products already used by millions of organizations. For the first time, businesses relying on Copilot have an official document to check against regarding what real safety guardrails actually apply to their tools.

Also today: more model choice in Office, and consolidation in meeting notetaking

Grok joins Word, Excel, and PowerPoint: Microsoft expanded model choice in Office Copilot — Microsoft Frontier users can now pick between OpenAI's GPT, Anthropic's Claude, and xAI's Grok directly from the model picker inside their documents. A practical addition for businesses wanting to test which model performs best for different tasks, without leaving the app.

Superhuman acquires Fathom: Productivity platform Superhuman (40 million users) acquired YC-backed meeting notetaker Fathom (400,000 monthly active users) rather than building one in-house. The goal: let AI agents act directly on meeting content — drafting follow-up emails, updating records, and scheduling next steps automatically. It's another sign that meeting transcription is becoming core infrastructure for AI-driven business automation, not just a convenience feature.

The bigger picture

Microsoft's code of conduct lands in a week where top AI CEOs (Amodei, Altman, Musk, and Nadella) publicly called for "pacing" advanced AI development — met with pushback from President Trump and Nvidia CEO Jensen Huang, who said his company "won't let an AI slowdown happen." Microsoft's document is the first concrete step translating that high-level debate into actual product policy.