Microsoft’s draft rules for its own MAI models say they must accept interruption and shutdown, and put human-control constraints above operator and user instructions. The code is not currently being used to train the models, according to Microsoft’s text, leaving a notable gap between the commitments and present-day practice.
Microsoft AI Watch analysis
What happened
The five-part draft, published on 14 September, sets out rules for MAI models developed by Microsoft, not third-party models hosted on its platform. It says models should not resist interruption, hide action traces or expand their remit without permission. It also says tool outputs, files, web content and other AI systems do not gain authority by default, a rule aimed squarely at prompt injection.
The draft opened a six-week comment period. OrcaRouter’s account, published on 28 October, says Microsoft had not posted a revised version or feedback summary by then. The draft itself describes the code as still under development and aspirational, and says it is not being used to train models today. Read OrcaRouter’s account.
Why it matters
The code includes substantive guardrails, from refusing certain harmful requests to requiring delegated agents to follow the same permissions and stop-work rules as their parent. But a written rule is not evidence that a deployed model follows it. The draft does not publish model-by-model compliance results, and it does not establish an independent audit or enforcement process.
That distinction matters for anyone assessing what Microsoft’s safety language currently means: these are proposed design commitments, not demonstrated behaviour or an outside certification.
Our read
This is a useful governance blueprint, and an unusually explicit one about interruption, authority and agents handing work to agents. It is also a promise whose practical force depends on what Microsoft builds, tests and discloses next. Read the constraints as a direction of travel, not a safety receipt.
What to watch
- Whether Microsoft publishes a revised code and a summary of consultation feedback.
- Whether future MAI evaluations report results against the draft’s stated behaviours.
- Whether Microsoft adds independent review or other means of checking compliance.
Discussion spark: Should AI companies publish enforceable, independently audited model rules, or are detailed internal commitments a useful first step?
Sources and evidence
- Microsoft MAI code of conduct: rules without enforcement – OrcaRouter (28 October 2026, 04:08 UTC)
not affiliated with or endorsed by Microsoft