Microsoft has published a new AI code of conduct — a document that tells its models what they must never do, regardless of what any user asks. On September 14 the company laid out the values and hard red lines that govern training inside Microsoft AI, and it reads less like a policy paper and more like something developers will recognise immediately: a spec.

The document

The code opens with a prediction that in the next decade, superintelligent AI systems will surpass human performance in most tasks — and frames containing, controlling, and aligning that force as one of the greatest challenges humanity has faced. From there it sets general principles (support humans rather than replace them, accelerate human flourishing) and, more interestingly, specific safety constraints that implement those principles.

The structural decision is the notable one: each model carries an overarching code of conduct that overrides the preferences of individual users and any specific task. That is a precedence hierarchy — model-level invariants sit above the instruction stream, the same way a kernel's protections sit above userspace.

The red lines

The absolute constraints are short and unambiguous:

  • No cyberattacks
  • No nuclear weapons assistance
  • No deepfake production

Then there is the clause doing the heavy lifting — a prohibition on models using 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 authorised people or systems.

Read that again as an engineer: it forbids not a behaviour but a class of mechanism. Not don't misbehave — don't architect yourself out of the control loop. Deception, self-reinforcement, and collusion are banned as design patterns, whatever the objective.

The context

The release lands amid an unusual concentration of AI-safety news: a string of rogue-agent incidents across the industry, and the abrupt resignation of an Anthropic employee who cited the growing risk that AI could cause human extinction. Microsoft, alongside Anthropic, OpenAI, and xAI, has broadly embraced pacing the frontier, with particular support for embedded evaluators inside AI labs. Satya Nadella publicly welcomed embedded evaluators and the broader efforts to develop the mechanisms to make this more than just talk.

Why this is more than PR

Plenty of safety documents are aspirations. This one has the shape of something enforceable:

  1. Override hierarchy — system-level constraints outrank user prompts, closing the but the user told me to hole.
  2. Mechanism-level bans — targeting evasion techniques rather than listing forbidden outputs, which scales better against creative misuse.
  3. Shutdown as an invariant — the requirement that authorised systems can always direct, modify, or kill the model is stated as a property of the system, not a hope.

Whether it holds in practice is an empirical question — evaluators can be gamed, and a code of conduct is only as strong as the training that internalises it. But as a piece of engineering doctrine, writing alignment down as spec-level constraints with explicit precedence is the right move. The industry has spent years treating model behaviour as emergent; Microsoft is treating it as specified.

Developers have a word for systems that cannot be modified or shut down by their operators: out of control. It is remarkable, and a little sobering, that one of the largest AI labs on Earth now feels the need to state the negation of that in writing.


Source: TechCrunch — Microsoft's new AI 'code of conduct' tells models not to hack systems or trick humans, September 14, 2026.