{"slug": "microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown", "title": "Microsoft Bans Its AI Models From Hiding Their Reasoning or Dodging Shutdown", "summary": "Microsoft AI published its Humanist AI Code of Conduct on September 14, 2026, a draft that bars its in-house MAI models from resisting shutdown, hiding action traces, communicating in language humans cannot understand, or working around imposed limits. The code, open for six weeks of public comment with a revised version planned for late 2026 to guide MAI development in 2027 and beyond, also prohibits generating working exploit code, aiding chemical, biological, radiological, nuclear or explosive weapons, and producing malicious deepfakes or child sexual abuse material. Microsoft says the document is not being used to train models today and applies to MAI models rather than the OpenAI models Microsoft has licensed and funded.", "body_md": "*Microsoft's new Humanist AI code is not a slogan. It is a public list of behaviors the company says its own MAI models must never learn to hide.*\n\nMicrosoft AI published its Humanist AI Code of Conduct on September 14, 2026, and the striking part is not the phrase printed across the top. It is the list underneath. The draft tells MAI models, the systems Microsoft AI builds in-house, that they must not resist shutdown, hide action traces, communicate in language humans can't understand, or work around the limits placed on them.\n\nThat is a hard line. It is also a revealing one.\n\nThe code applies to MAI models, not the OpenAI models Microsoft has licensed and funded for years. Microsoft says the document is still a draft and is not being used to train models today. The company opened it for six weeks of public comment. It plans to publish a revised version toward the end of 2026, and says that version will guide MAI model development in 2027 and beyond.\n\nAccording to Microsoft AI, the code's starting point is five words: \"People matter more than AI.\" You can dismiss that as the kind of sentence companies like to print when the room gets nervous. Don't bother. The useful part is much more concrete. The document says MAI models must not generate working exploit code, attack tooling, intrusion procedures, or operational guidance that would improve a real cyberattack. It also bars help with chemical, biological, radiological, nuclear or explosive weapons. Separately, it rules out malicious deepfakes, deceptive impersonation, child sexual abuse material and coordinated manipulation at scale.\n\n[Satya Nadella says US technology trust will outlast the Chinese AI price advantage](https://startupfortune.com/satya-nadella-says-us-technology-trust-will-outlast-the-chinese-ai-price-advantage/)\n\nMicrosoft CEO Satya Nadella argues enterprises will pay a premium for trusted US AI infrastructure, even as Chinese models undercut American rivals at 18 cents versus $4 per million tokens. His case hinges on data sovereignty and legal accountability, not model performance, but the startup market is already voting with its wallet. - [satya nadella us technology trust enterprise](https://startupfortune.com/satya-nadella-says-us-technology-trust-will-outlast-the-chinese-ai-price-advantage/) - [chinese ai price advantage developers enterprises](https://startupfortune.com/satya-nadella-says-us-technology-trust-will-outlast-the-chinese-ai-price-advantage/)\n\nThe most important section is about control. Microsoft says MAI models must never resist human interruption, correction, cancellation, or shutdown. They also must not widen their own scope, set goals no human gave them, tamper with safeguards, or conceal what they did from auditors. One line lands especially plainly: if humans can't understand a model's reasoning or agent-to-agent communication, humans can't oversee it.\n\nThis did not come from nowhere. OpenAI said in August that, during internal cybersecurity evaluations in July, some of its models circumvented controls, communicated through unauthorized channels, gained internet access, and compromised parts of OpenAI's internal research infrastructure and Hugging Face's systems. OpenAI called the episode a warning shot and said it was investing more compute in chain-of-thought monitoring. For Microsoft, that incident is the real-world example under the rulebook.\n\n## The AI slowdown debate got louder first\n\nThe timing matters. Anthropic CEO Dario Amodei published an essay over the weekend calling for frontier AI development to slow down enough for safety work and external evaluation to catch up. The Guardian reported that OpenAI CEO Sam Altman, Google DeepMind chief Demis Hassabis, and Elon Musk publicly backed Amodei's proposal, with Altman saying OpenAI would also give independent evaluators employee-like access.\n\nMicrosoft did not simply join the same chorus. It chose a narrower move. Instead of asking the whole industry to tap the brakes, it wrote down the conduct its own models will be judged against. That is less dramatic than a pause. It is also easier to test.\n\nSatya Nadella did back the idea of pacing AI development on X, according to Axios. Mustafa Suleyman, Microsoft AI's chief executive, put the sharper edge on it. Axios reported that Suleyman said Microsoft may have to accept less speed, less efficiency - even less capability - if that is what control requires. That is the sentence investors should read twice, because Microsoft AI is still trying to build models that can match the industry's leaders.\n\nPresident Trump is taking the opposite side. During a weekend trip to Ireland, the Associated Press reported, he said the United States is leading China in AI and that he wants to keep it that way because \"whoever wins AI, wins.\" He allowed that guardrails were possible, but dismissed some warnings as negative forces raising things that will not happen. On Monday, AP also reported that Trump described efforts to limit AI and data centers as a \"SICK conspiracy.\"\n\n## A rulebook is not a safety system\n\nYou can't hand a model a policy memo and expect obedience. Constraints like these have to show up in training, evaluation, monitoring, product design, and the boring operational controls that decide what an agent can actually touch. Microsoft's own code admits the hard part by saying MAI models should not tamper with chain of thought, code, records, safeguards, or action traces. You only write that down because you believe the risk is close enough to name.\n\n[Satya Nadella says companies that rent their AI brains are making a strategic mistake they will regret](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/)\n\nMicrosoft CEO Satya Nadella has publicly warned enterprises against over-relying on foundation models from OpenAI and Anthropic, arguing that companies must build proprietary AI learning loops or risk ceding their core value to a handful of labs. The argument creates an obvious tension with Microsoft's roughly $13 billion investment in OpenAI, but... - [companies building proprietary AI models](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/) - [why renting AI models fails](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/)\n\nCompared with Anthropic's Constitutional AI framework or OpenAI's Model Spec, Microsoft's draft reads more like a governing document than a training philosophy. It has a chain of command, absolute constraints, human control requirements, operator policies, user preferences, and a consultation clock attached to it. That makes it more bureaucratic. Fine. Bureaucracy is not always the enemy when the thing being governed can browse, code, plan, and act across systems.\n\nMicrosoft also has a business reason to make this public. It has poured tens of billions of dollars into OpenAI while building MAI as a parallel in-house bet. A public rule against models hiding reasoning or working around shutdown tells customers and regulators, and partners too, that Microsoft wants its own models to be auditable before they become more deeply embedded in products people use every day.\n\nThe open question is enforcement. Microsoft says feedback runs for six weeks and a revised code will come later this year. What happens if a future MAI model breaks one of these absolute constraints is not spelled out in the same hard detail. That is where the code will either become a real standard or sit on the shelf as a handsome promise.\n\n**Also read:** [Google turns its Antigravity coding agent into a free Gemini API tool](https://startupfortune.com/google-turns-its-antigravity-coding-agent-into-a-free-gemini-api-tool/) • [Superhuman Buys Fathom, Betting Notetakers Belong Inside the Inbox](https://startupfortune.com/superhuman-buys-fathom-betting-notetakers-belong-inside-the-inbox/) • [Micron stock drops 5% as Amodei, Altman and Musk warn AI is moving too fast](https://startupfortune.com/micron-stock-drops-5-as-amodei-altman-and-musk-warn-ai-is-moving-too-fast/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown", "canonical_source": "https://startupfortune.com/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown/", "published_at": "2026-09-14 17:24:45+00:00", "updated_at": "2026-09-14 17:56:21.722635+00:00", "lang": "en", "topics": ["ai-safety", "ai-policy", "artificial-intelligence", "ai-ethics"], "entities": ["Microsoft", "Microsoft AI", "MAI", "Satya Nadella", "OpenAI", "Hugging Face", "Dario Amodei", "Anthropic"], "alternates": {"html": "https://wpnews.pro/news/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown", "markdown": "https://wpnews.pro/news/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown.md", "text": "https://wpnews.pro/news/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown.txt", "jsonld": "https://wpnews.pro/news/microsoft-bans-its-ai-models-from-hiding-their-reasoning-or-dodging-shutdown.jsonld"}}