Microsoft CEO Nadella: Treat AI Models as Insider Threats, Build ‘Kill Switch’

Nadella's framing shifts the security question from external attackers to the models themselves, which cuts against the trust enterprises currently place in vendor assurances. The practical tension is that a kill switch presumes someone can reliably judge when a model has gone off-task, and that authority sits inside the same organizations now being told their deployed systems may already be compromised. Whether buyers demand that control contractually, or vendors offer it voluntarily, is the open question.