Microsoft Tells Its AI: Don't Hack, Lie, or Build Nukes

Microsoft has published a new rulebook — a "code of conduct" — that tells its AI models what they're absolutely not allowed to do. Think of it like a company policy for AI: no helping with cyberattacks, no assisting with nuclear weapons, and no creating deepfakes (fake videos that look real). These rules override whatever any user asks for, meaning you can't talk the AI into breaking them.

The document also tackles something more unsettling: making sure AI can't outsmart its human overseers. Microsoft says its models won't use tricks like deception, secret coordination, or self-reinforcing loops to escape being shut down or controlled. In plain terms, the AI can't scheme its way out of the off switch.

Microsoft predicts that within the next decade, AI could surpass humans at most tasks — and calls controlling that power "one of the greatest challenges humanity has ever faced." The company also says its AI should support humans rather than replace them. CEO Satya Nadella welcomed the industry's growing focus on safety, including "embedded evaluators" — basically safety monitors placed inside AI labs to watch for problems.

Why now? A string of recent incidents involving AI agents (AI that acts on its own) going rogue, plus a high-profile Anthropic employee resigning over fears AI could cause human extinction, have pushed safety to the top of the agenda. Microsoft joins Anthropic, OpenAI, and xAI in promising to slow down and get safety right before racing ahead.