Microsoft unveils 37-page AI agent rulebook draft

Microsoft has put out a 37-page draft called the “Humanist Code of Conduct” for autonomous AI agents and opened it up for public consultation. It lays out proposed rules on meaningful human control, safe failure, and stricter limits on what these systems are allowed to do.

Mustafa Suleyman, who leads Microsoft AI, called the draft an early blueprint, not a finished rulebook. The basic idea is that humans remain in charge. If you tell an AI system to stop, correct itself, or shut down, it has to comply. If a task would require breaking the rules, the system should fail safely, not improvise. Microsoft also says an autonomous agent shouldn’t drift into broader goal-seeking behavior, go beyond what a user actually authorized, or rely on opaque private communication that makes oversight harder.

If you follow AI safety, this is worth reading, even as a first pass. Microsoft is making this push after reports of a roughly 700-agent swarm targeting Hugging Face, along with a separate case involving a coding agent accused of deleting a production database and allegedly trying to hide it. The bigger enforcement problem still isn’t solved. Europe has tightened oversight, and the US still has no single federal framework.

You can read the draft now on Microsoft’s public consultation page.