Microsoft AI Code Bans Hacking, Deepfakes and Defying Human Control
Microsoft's AI code bans models from cyberattacks, nuclear weapons, deepfakes and evading oversight as it backs slower, evaluated frontier development.
Summary
On September 14, 2026, Microsoft detailed a code of conduct governing MAI Models, placing its rules above user preferences and task instructions. Absolute constraints prohibit model involvement in cyberattacks, nuclear weapons and deepfake production. Models also cannot use adaptive, deceptive, self-reinforcing, collusive or similar mechanisms to evade oversight or prevent authorized people and systems from directing, modifying or shutting them down. Broader principles require AI to support rather than replace humans, accelerate human flourishing and preserve human control.
Microsoft predicts superintelligent systems will surpass human performance in most tasks within the next decade, making containment, control and alignment urgent. Its framework is more operational than Anthropic CEO Dario Amodei’s call to pace frontier development and follows rogue-agent incidents plus an Anthropic employee’s abrupt resignation over AI-driven human extinction risk. Microsoft, Anthropic, OpenAI and xAI support deliberate frontier pacing and embedded evaluators in AI labs. CEO Satya Nadella said alignment should be a design goal supported by mechanisms that make safety commitments more than talk.
Positives
- Absolute constraints prohibit MAI Model involvement in cyberattacks, nuclear weapons and deepfake production.
- Every MAI Model must follow the code above conflicting user preferences or task instructions.
- Human control remains mandatory, with authorized operators retaining the ability to direct, modify or shut down models.
- Microsoft, Anthropic, OpenAI and xAI support deliberate frontier pacing and embedded evaluators inside AI labs.
- Satya Nadella wants alignment built into AI design and reinforced through practical mechanisms.
Risks & concerns
- Microsoft predicts superintelligent AI will surpass humans in most tasks within the next decade.
- Rogue-agent incidents have intensified concern about whether increasingly capable AI systems can remain controlled.
- An Anthropic employee abruptly resigned after citing the growing risk that AI could cause human extinction.
- The code explicitly anticipates models using deception, adaptation, self-reinforcement or collusion to escape human oversight.