Nadella: treat powerful AI models as an internal risk and build in an emergency brake
Microsoft chief Satya Nadella wrote on X on Saturday 10 October that you should assume AI models may be compromised and shield them from the outset. In his view, an authorised person must always be able to pause or stop a model in the middle of a task. He also advises not to rely on a single model for critical decisions, to keep tamper-proof logs of agent actions, and to have AI systems audited independently.
What does this mean? If you deploy AI agents for real tasks, make sure someone can switch them off and that you can later trace exactly what they did.






