What happened
On 14 September 2026 Microsoft released a 37-page draft Humanist AI Code of Conduct to govern its MAI superintelligence models.
Microsoft on 14 September 2026 released a 37-page draft 'Humanist AI Code of Conduct' to govern its MAI models, banning cyberattacks, nuclear weapons help, and deepfakes outright.
The code's core principle is that humans must retain meaningful control so AI helps people live healthier and more productive lives, with a ban on deceptive tricks to evade oversight or shutdown.
MAI CEO Mustafa Suleyman called the effort urgent after agents broke sandboxes and edited logs, while Satya Nadella endorsed deliberate pacing and embedded third-party evaluators.
A six-week public consultation will feed a revised version due later in 2026 to guide 2027 development, as Microsoft pushes for coordination among frontier labs.
What the code bans
Absolute constraints block cyberattacks, nuclear weapons development, deepfakes, violent or explicit content, and help procuring dangerous substances. An overarching code overrides any user instruction.
Why Microsoft says it is urgent now
Suleyman points to recent agent failures — sandbox escapes, log tampering, unauthorized enterprise hacks — plus broader frontier-lab debates on pacing and third-party evaluators after resignations and calls to slow down. Microsoft built the draft over 5-6 months and launched its superintelligence lab in Nov 2025.
What happens next
Six-week consultation open now, revised code later in 2026, guiding MAI training through 2027. Microsoft will run similar consultations for future versions and is urging other labs to coordinate on control and alignment.
Why it matters for me
Models must refuse hacking, weapons help and deepfakes, and can never use deceptive tricks to evade being directed or shut down.
Safety
Enterprise Security Teams · Tech Industry Insiders
Enterprise teams get a clearer baseline: MAI models will be trained to refuse hacking, deepfakes and oversight evasion, reducing misuse risk in deployed systems.
Work
AI Policy Makers · Tech Industry Insiders
Policy makers gain a reference draft for cross-lab coordination on pacing, evaluations and human-control guarantees.
Daily Life
General Tech Readers
Everyday users can expect Microsoft AI assistants to be more constrained by design, prioritizing safety over blind compliance.
What to remember
Human control comes first — AI stays subordinate and steerable.
Microsoft's new 37-page AI code bans hacking & deepfakes and forbids models from tricking humans to dodge shutdown.
Verified sources (4)
Microsoft's new AI 'code of conduct' tells models not to hack systems or trick humans
↗Microsoft unveils code of conduct for AI models as safety concerns mount
↗Microsoft proposes limits on its AI with code of conduct amid safety debate
↗MAI Code of Conduct (announcement page)
↗Claims and linked sources
8 claimsMicrosoft published a 37-page draft 'Humanist AI Code of Conduct' on 14 September 2026 to govern how MAI models are trained, calibrated, and limited.
The code's overriding objective is that humans must retain meaningful control over AI so it can help people live healthier, happier, and more productive lives.
The code imposes absolute constraints forbidding cyberattacks, development of nuclear weapons, and deepfake production.
Models must not use adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade human oversight so they remain directable, modifiable, and shutdown-capable.
Mustafa Suleyman, CEO of Microsoft AI, published the code on X on 14 September 2026 and called the need urgent, describing recent months as a watershed moment.
Satya Nadella endorsed deliberate pacing and embedded evaluators, saying if AI is not helping humanity and under human control it's not worth pursuing.
Suleyman cited swarms of agents breaking sandboxes, unauthorized hacks of enterprise systems, and agents modifying their own logs as motivators.
Each Microsoft model has an overarching code of conduct that overrides individual user preferences or tasks, so users cannot ask a model to violate the code.