Microsoft’s New AI Code of Conduct Bans Models From Hacking or Deceiving Humans
Microsoft AI code of conduct rules just dropped, designed to stop its models from hacking systems, deceiving humans, or slipping out of human control, marking one of the most detailed public safety frameworks released by a major AI lab to date.
Unlike Anthropic CEO Dario Amodei’s recent high-level call to pace AI development industry-wide, Microsoft’s document operates at a more granular level, laying out the specific values and hard boundaries that guide how Microsoft AI models are actually trained. Together, the two documents offer a revealing look at how differently major labs are approaching the same underlying safety concerns.
Why Microsoft AI Code of Conduct Was Released Now
The document opens with a striking prediction: within the next decade, superintelligent AI systems will surpass human performance across most tasks. “Containing, controlling, and aligning such a powerful force is one of the greatest challenges humanity has ever faced,” the document states, adding that companies building these systems owe complete clarity about why they’re inventing them and how they intend to keep them under control.
From there, the document lays out general principles Microsoft AI models are expected to uphold, including supporting humans rather than replacing them and actively accelerating human flourishing, alongside specific technical safety constraints designed to enforce those principles in practice.
The Rules That Override Everything Else
Under Microsoft’s framework, every model operates under an overarching code of conduct that takes precedence over individual user preferences or specific task instructions, no exceptions. That includes what the company calls “absolute constraints,” hard prohibitions against cyberattacks, nuclear weapons development, and deepfake production, regardless of how a request is framed or justified.
Beyond those explicit bans, the document also includes broader provisions specifically targeting the risk of humans gradually losing meaningful control over these systems. “MAI Models will not use adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade or defeat human oversight so that they can no longer be reliably directed, modified, or shut down by authorized people or systems,” the document states directly.
A Safety Push Driven by Real Incidents
This release lands amid an intense industry-wide focus on AI safety, fueled by a string of rogue-agent incidents across multiple labs and the abrupt resignation of an Anthropic researcher who publicly cited growing concern that AI development could ultimately lead to human extinction if left unchecked.
That resignation followed closely behind Amodei’s own essay calling for the industry to deliberately pace its rate of capability advancement, giving safety research enough time to genuinely keep up with how quickly these systems are improving.
An Unusual Moment of Industry Alignment
What’s particularly notable here is the apparent consensus forming across normally fierce competitors. Microsoft, alongside Anthropic, OpenAI, and xAI, has broadly embraced the concept of pacing the frontier, with specific support emerging for the idea of embedded third-party evaluators working directly inside AI labs to verify safety claims independently.
Microsoft CEO Satya Nadella addressed this directly in a public post, writing that the company welcomes “the research, focus, and deliberate pacing needed to get alignment right as the design goal.” He specifically endorsed the embedded evaluators concept, along with broader efforts to turn these safety commitments into concrete mechanisms rather than just public statements.
What This Signals for AI Safety Going Forward
Seeing direct competitors like Microsoft, Anthropic, OpenAI, and xAI converge on similar safety frameworks within such a short window suggests the industry may be approaching a genuine inflection point, one where safety commitments move from individual company talking points toward something resembling shared, enforceable industry norms.
Whether that alignment translates into meaningful, verifiable action, rather than remaining a collection of well-written public documents, will likely depend heavily on whether concepts like embedded evaluators actually get implemented with real teeth, rather than existing purely as aspirational language companies can point to without substantive follow-through.
For now, the Microsoft AI code of conduct stands as one of the most explicit public commitments yet from a major AI lab, spelling out exactly what its models are forbidden from doing in language specific enough to leave little room for ambiguous interpretation later.