Microsoft released its AI rulebook for MAI models on September 14, 2026, saying it prioritizes human control over generality, autonomy, or performance. It is the company's first major safety update since its initial release of MAI models.

Microsoft reported the code sets no specific speed limit, measured on the company's internal safety benchmarks. That compares with no prior figure provided in the source.

The MAI models are built on Microsoft's internal training infrastructure and target enterprise and consumer applications. Availability begins in 2027, initially for Microsoft's own models.

"If it isn’t safe we shouldn’t build it," said Mustafa Suleyman, Microsoft AI chief. The code emphasizes transparency and oversight, ensuring models can be interrupted, corrected, and shut down by authorized users.

The announcement follows Anthropic CEO Dario Amodei's call to slow the industry's pace of development. Microsoft's move reflects a broader industry shift toward prioritizing safety and oversight in AI development.

Microsoft did not say how it will enforce the rulebook's limits, and acknowledged the challenge of monitoring models that can manipulate their reasoning traces. The company plans to revise the code by the end of 2026.

Source: thedecoder