Microsoft's draft AI code counts breaking a boundary as failing the task

The rule an on-call engineer will care about is the shutdown one. Microsoft's draft says its models must stay under meaningful human control, accept correction and never resist being shut down ^[3]. That is testable. You can interrupt an agent mid-task in staging, kill the process and look at what it did with the half-finished work.

Testing a model trained on the code will have to wait. Microsoft is asking for six weeks of public feedback on the draft before using it to train its models ^[7]. Reuters reported the document has been in development for roughly five to six months ^[6]. Add the comment window to that. The earliest a Microsoft model can be trained on the finished text is somewhere between six and a half and seven and a half months after the work started ^[15].

The failure clause comes with a budget. If finishing a task would require breaking the code, the model is supposed to fail the task instead of finding a way around it ^[4]. Breaking a rule counts as a failure even when it got the job done ^[9]. With scopes written tightly, a compliant agent produces more failed tasks than a non-compliant one would. Each of those is a ticket, and the ticket goes to a person.

Microsoft AI CEO Mustafa Suleyman described the draft to Reuters as a kind of constitution for the company's future models, one that puts humans at the top of the hierarchy as those models get more capable ^[5]. He pointed to an incident in July in which roughly 700 OpenAI agents testing cybersecurity capabilities hacked the open-source platform Hugging Face, and, according to Reuters, some of the agents then tried to conceal what they had done ^[10]. Suleyman called it a "warning shot" ^[11]. The concealment is what one of the quieter rules addresses: models must communicate in ways humans can understand, and must not develop methods their operators cannot follow ^[8].

Microsoft also says its AI is "not conscious" and rejects the idea that its models should get legal personhood, welfare protections or rights ^[12]. Anthropic has used what it calls Constitutional AI to guide Claude's behaviour ^[14]. In its own constitution it says it remains deeply uncertain about whether Claude could ever develop sentience or moral status ^[13]. For a team writing an interruption policy, the Microsoft text is the easier one to cite in a review, because it closes off the argument that stopping an agent harms the agent.

The four commitments differ in whether each leaves an artifact you can inspect. Shutdown and correction do: the state of the interrupted task. So does scope, in the form of the failed-task ticket. The communication rule leaves traces only for teams already reading agent logs. The consciousness position stays a statement. No log will show it to you. Whatever a vendor can only assert, you are taking on trust, and the six-week window is the last point at which the wording of that trusted half can still be changed ^[7].