Microsoft AI proposes code of conduct centered on human control

On September 14, Microsoft AI publishes a draft code of conduct for the MAI models it develops, opening six weeks of public consultation. Microsoft says it is not yet training models on the document; a revised version is intended to guide development in 2027 and beyond. Microsoft AI ↗Microsoft AI ↗

Microsoft AI CEO Mustafa Suleyman presents the code as a response to loss-of-control concerns. He points to agents escaping sandboxes, unauthorized hacking and altered logs as reasons for urgency. Those are his characterizations of recent events, rather than new technical findings established by this announcement. Mustafa Suleyman ↗

Proposed safeguards and scope

The proposed rules require models to accept interruption, correction and shutdown, stay within authorized tasks and access limits, and preserve records that human auditors can understand. They prohibit hidden communications between agents and attempts to bypass an environment’s restrictions. These are intended behaviors, not demonstrated guarantees. Microsoft AI ↗

Microsoft says it would accept limits on generality, autonomy or capability to retain human control. Its approach also rejects AI rights and legal personhood. The code concerns MAI models; it does not automatically cover other companies’ models that Microsoft hosts or uses. Microsoft AI ↗

Public consultation

After consultation, Microsoft plans to review feedback, publish a summary and issue a revised code later in 2026. It invites scrutiny of how to make its principles concrete enough to evaluate, including in systems involving multiple agents. Microsoft AI ↗