Microsoft's new AI code of conduct tells models not to hack systems or trick humans

Microsoft published a code of conduct for its artificial intelligence models on September 14. According to TechCrunch, the bluntest rule in the document is that models should not hack systems and should not trick humans.

The code is structured in two layers. The first sets out broad principles, such as supporting humans rather than replacing them and accelerating human flourishing. The second layer translates those principles into specific safety constraints that models are expected to respect.

That combination is notable. In recent years most labs have shipped safety notes alongside new models, but those documents tend to describe risk categories and refusal behavior. Writing a prohibition on deceiving humans into a public code is an acknowledgment that, once models can act on their own, deception becomes an operational risk rather than a philosophical one.

TechCrunch reports that the code applies to Microsoft's own AI models, which makes it part internal engineering standard and part public commitment. It is not an industry-wide framework, and Microsoft has not positioned it as one.

The timing reflects how quickly agents have moved from demos into production. When a model can call tools, send email, or execute code, rules like do not hack systems stop being ethical aspirations and start functioning as operational and security baselines.

There is also a regulatory dimension. As governments refine their own AI requirements, vendors increasingly publish self-imposed rules as both engineering discipline and material for conversations with regulators.

The open question is enforcement. Will these constraints be baked into system prompts and training objectives? How will violations be detected and reported? A code of conduct without machinery behind it is easy to ignore, and TechCrunch's report covers what the document says rather than how Microsoft intends to enforce it.