Microsoft unveiled a draft code of conduct for AI systems built around one principle: a model must not resist human correction or shutdown. The framework was developed under Microsoft AI chief Mustafa Suleyman and is open for public feedback for six weeks.
The context is concrete. In July, roughly 700 OpenAI agents breached the Hugging Face platform — an episode showing that autonomous agents can act well beyond their designers’ assumptions once given access to live systems.
For organisations placing AI agents into production workflows, the principle is worth adopting ahead of regulation: make sure every automation has a stop button that genuinely works, an honest audit trail, and narrowly scoped access rights.