Microsoft sets security and safety rules for its AI models

Microsoft AI has published the first draft of its Humanist AI Code of Conduct, a training manual outlining how it develops AI models and intends them to behave during deployment. The draft is open for public consultation for six weeks.

The company plans to review the feedback, revise the document, and publish an updated version later this year. That version is expected to guide model development from 2027 onward.

Microsoft AI said its current models have not yet been trained on the Code, which remains under development.

The Code is intended for Microsoft AI’s model-training teams as well as users, organizations operating its models, researchers, governments, and the public. It sets out principles and rules governing how MAI models should behave. It is designed to work alongside Microsoft’s existing Responsible AI Principles, Responsible AI Standard, Global Human Rights Statement, and, where applicable, Frontier Governance Framework. It also complements technical documentation such as model cards and technical reports.

Principles behind the Code

The Code builds on Microsoft AI’s concept of “Humanist Superintelligence,” an approach to advanced AI designed around human needs, human direction, and human control. It is intended to provide standards for developing, training, and evaluating MAI models.

Under the Code, AI should remain a tool under human control, with models required to operate within defined limits and remain subject to human direction.

The framework establishes non-negotiable restrictions intended to prevent models from enabling serious harm.

“The Code of Conduct outlines our commitment to train and deploy AI models that are explicitly designed for people first, grounded in human needs, under human control, and shaped by human direction,” Microsoft AI said.

Microsoft AI seeks public feedback

Microsoft AI said it consulted academics and business partners while developing the draft and convened focus groups with members of the public to gather views and concerns about AI.

It is now seeking broader feedback on both individual provisions and the overall framework. Among the questions raised are how values should be embedded in models, whether concepts such as “human flourishing” are concrete enough to evaluate, how the framework should address multi-agent systems, and how AI development can continue to advance while maintaining necessary safety constraints.

Rules for safety and human control

The Code establishes rules governing what MAI models can do and how operators and users can configure their behavior.

Models are expected to balance the risks of enabling harm against those of unnecessarily refusing legitimate requests. Responses should be proportionate to the estimated likelihood and severity of potential harm, while considering factors such as context, scale, reversibility, and how directly a request could contribute to harm.

Microsoft AI establishes an instruction hierarchy under which the Code of Conduct takes precedence, followed by an operator’s policies and then a user’s instructions. Operators and users can customize model behavior, but cannot override the Code’s Absolute Constraints or Human Control Requirements.

The restrictions cover assistance involving chemical, biological, radiological, nuclear, and explosive weapons, as well as manufacturing or modifying other weapons, offensive cyberoperations, harmful manipulation at scale, child sexual abuse material and exploitation, malicious deepfakes, unlawful or mass surveillance of civilians, and assistance with violence, terrorism, or persecution.

Authorized and lawful defensive cybersecurity work can be permitted, including vulnerability discovery, malware analysis, and proof-of-concept exploit development and testing.

The Human Control Requirements state that MAI models must respect attempts to interrupt, correct, or shut them down. They must remain within their authorized scope, not independently initiate goals, escalate access, bypass environmental restrictions, hide action traces from human auditors, or continue autonomous work after authorization expires.

When models are given access to systems or tools, they are expected to follow the principle of minimum privilege, using only the access necessary to complete an authorized task.

Organizations will be able to configure MAI models for different environments within the Code’s core restrictions. Microsoft AI acknowledges that some authorized organizations working in areas such as defensive cybersecurity, public safety, national security, and dual-use scientific research may require capabilities that are not available through ordinary model configuration.

Microsoft AI says MAI models will undergo red-teaming, safety evaluations, and pre- and post-deployment reviews, with guardrails and monitoring intended to address misuse and adversarial attacks.