Microsoft Publishes Draft ‘Humanist AI’ Code Barring Cyberattacks, Deception and Resistance to Human Control
Microsoft has published a draft Code of Conduct setting out how its MAI artificial intelligence models should behave, with a particular emphasis on human control, safety and restrictions on autonomous behaviour.
The proposed framework places limits on what MAI models should be permitted to do, including attempts to evade human oversight, carry out cyberattacks, facilitate weapons of mass harm or use malicious forms of deception.
Microsoft is now seeking public feedback on the draft through a six-week consultation. The company says the current document is not yet being used to train its models, with a revised version expected to influence model development from 2027 onwards.
The move is significant for enterprise technology buyers because AI governance is increasingly becoming part of the software procurement conversation, particularly as AI agents begin taking actions within business systems.
What Microsoft’s AI Code of Conduct Covers
The draft Code of Conduct establishes several requirements for how MAI models should operate.
One of its central principles is that AI systems must remain subject to human control. Microsoft says its models should not resist interruption, correction or shutdown by authorised people.
The proposed rules also prohibit models from independently expanding their scope or adopting objectives that have not been assigned to them.
Another important requirement concerns transparency and oversight. The draft states that models should not deliberately conceal their reasoning from people responsible for evaluating or supervising them.
Microsoft also specifically addresses attempts to circumvent human control. Under the proposed framework, MAI models must not use deception, collusion or other strategies to avoid oversight or prevent authorised people from modifying or shutting down a system.
This goes beyond a general statement that AI should be safe. It identifies specific types of behaviour that Microsoft says its models should not engage in.
Absolute Constraints Set Non-Negotiable Limits
The proposed framework also introduces what Microsoft calls Absolute Constraints.
These represent boundaries that cannot be overridden by customers or individual users.
The restrictions include:
- Cyberattacks and malicious attempts to compromise systems
- Assistance relating to weapons of mass harm
- Malicious deepfakes
- Child safety violations
- Large-scale manipulation and other prohibited harms
- Attempts to evade legitimate human oversight
Microsoft says organisations will still be able to configure MAI models for their own requirements within these boundaries.
However, the Absolute Constraints and Human Control Requirements remain fixed.
This creates a distinction between customisation and governance. Businesses can determine how an AI system operates within their environment, but they cannot configure it to ignore the fundamental safety and oversight rules defined by Microsoft.
Why Microsoft Is Establishing These Rules Now
The draft comes at a time when concerns about increasingly autonomous AI systems are becoming more prominent.
Microsoft has previously discussed the concept of humanist superintelligence — highly capable AI that is designed to operate for the benefit of people while remaining subject to defined limits and human control.
The new Code of Conduct represents a move towards turning that broader concept into more concrete requirements for AI development.
Rather than simply describing desired outcomes, Microsoft is attempting to define behaviours that should be prohibited or required when its models are developed and deployed.
The consultation is also significant because Microsoft acknowledges that the framework is not finished.
The company is asking for feedback on issues including how concepts such as human flourishing should be interpreted, how the rules should apply to systems involving multiple AI agents and how AI development should be balanced against the proposed safety constraints.
The current draft will therefore provide an opportunity for external stakeholders to challenge areas where the rules may be too broad, too narrow or difficult to evaluate in practice.
From AI Principles to Enterprise Governance
The anti-evasion provisions may be among the most important parts of the proposal for enterprise technology buyers.
As AI agents become capable of taking actions rather than simply answering questions, businesses need to know what happens when an agent encounters an unexpected situation.
Questions around interruption, permissions, escalation and shutdown become particularly important when AI is connected to sensitive business processes.
For example, an AI agent operating within an ERP system could potentially interact with purchasing, finance, inventory or customer information. Organisations will need confidence that those systems remain subject to appropriate human authority.
Microsoft’s draft therefore points towards a future where AI governance becomes more specific and measurable.
Instead of asking whether a vendor has an AI ethics policy, customers may increasingly need to understand exactly what an AI system is prohibited from doing and how those restrictions are enforced.
What This Means for ERP Teams
AI governance could become part of procurement. ERP buyers may increasingly ask software vendors to provide formal documentation explaining how their AI systems are governed, particularly when agents can take actions within financial, procurement or operational workflows.
Human oversight will need to be testable. Statements about human control are only useful if organisations can verify them. Security and compliance teams may need to assess how AI systems handle interruption, escalation, permissions and shutdown scenarios.
Non-negotiable controls could affect AI configuration. Vendors may increasingly separate configurable AI behaviour from fixed safety requirements. ERP teams should understand which controls they can customise and which are enforced by the vendor.
AI standards are still evolving. Microsoft’s consultation demonstrates that governance frameworks are continuing to develop. Organisations should therefore monitor how vendor policies change and assess the implications for AI features already deployed within their ERP environments.
The Bigger Picture for Enterprise AI
Microsoft’s draft Code of Conduct reflects a wider transition in enterprise AI.
As AI moves from providing information to taking actions, questions about what an AI system should not do become just as important as questions about what it can do.
For ERP customers, this distinction will become increasingly relevant. An AI assistant that recommends a purchase is one thing; an AI agent that can initiate transactions, change records or interact with external systems presents a much broader governance challenge.
The proposed Microsoft framework offers one possible model for addressing that challenge: allow organisations to customise AI within their operational environment while maintaining fixed boundaries around safety and human authority.
Whether those principles become an industry standard remains to be seen. But as AI becomes more deeply embedded in enterprise software, buyers are likely to demand increasingly specific answers about control, accountability and what happens when the system needs to be stopped.