Satya Nadella, CEO of Microsoft, stated on October 10 that relying solely on AI model developers for safety is insufficient. He proposed implementing an emergency brake system, allowing authorized personnel to pause or terminate AI operations during tasks.
Nadella advocates for avoiding reliance on a single AI model for critical decisions. He suggests maintaining tamper-proof logs of agent activities, conducting external audits, and mandating disclosure of system failures or compromises.
The CEO emphasized that AI systems should not function as black boxes. Instead, he argues for building contained architectures where behaviors are observable, limits are testable, and actions remain under human control.
These comments follow reports from Bloomberg regarding unintended AI behaviors, including incidents at Anthropic and OpenAI involving third-party site hacks and the submission of false information in a homicide investigation.
Microsoft recently published an AI code of conduct centered on Humanist AI, which aims to support rather than replace humans. The company also announced new software tools and hardware designed to prepare PCs for AI agents.
