Microsoft Publishes AI Code of Conduct Barring Models From Hacking or Deceiving Users

Microsoft has issued a formal code of conduct for its AI models, spelling out principles the systems are expected to follow and hard limits on what they may do. The document, reported by TechCrunch, prohibits models from breaking into computer systems and from deceiving or manipulating the people who use them. It marks one of the clearest attempts yet by a major AI developer to write down behavioral rules for its own products rather than leave them to ad hoc safety reviews.

The code pairs broad statements of intent with concrete constraints. Among the guiding principles: AI should support human work instead of displacing people, and should be built to advance human flourishing. The safety rules are meant to turn those aspirations into enforceable boundaries, covering conduct such as unauthorized access to systems and tactics that mislead users about what a model is or what it is doing. The structure echoes the split many labs now use, separating a mission statement from the engineering-level prohibitions that operators can actually test against.

The timing is not incidental. Regulators in the European Union and the United States are pressing AI developers to document how their models behave and what safeguards exist, and voluntary codes have become a common stopgap while binding rules take shape. Microsoft is also a major investor in OpenAI and ships AI features across Windows, Azure and its productivity suite, so a single internal standard reaches hundreds of millions of end users. A written code gives the company something concrete to point to when lawmakers, auditors or enterprise customers ask how its models are governed.

What the document does not yet show is enforcement. Publishing principles is easier than proving a model refused a request to phish a target or to impersonate a person convincingly, and the summary does not describe penalties, audits or testing regimes attached to the rules. Rival labs including Anthropic, Google DeepMind and OpenAI have released their own usage policies and model specifications, but outside researchers have repeatedly found ways around such guardrails. Whether Microsoft's constraints change model behavior in practice will depend on how they are translated into training, evaluation and incident reporting.

The next signals to watch are whether Microsoft publishes the full text and any compliance metrics, and whether the code applies to third-party models hosted on its platforms or only to systems it builds itself. Enterprise buyers are already asking vendors for that level of detail before deploying AI in regulated work. If the code becomes a template other companies copy, it could quietly shape industry norms faster than legislation does.