Satya Nadella: AI agents introduce an insider threat

‘In fact, the attack can just come from the agent itself’

EXCERPT

NADELLA: "With these powerful AI agents, you kind of have an insider risk, right? It's not... I mean, if you look at the historical cyber risk was all about, uh, the bad actor using cyber capability, uh, to attack. Um, and the question now is, in fact, the attack can just come from the agent itself, and so therefore having, uh, notifications between, uh, those two states is a fantastic start. And I think the US leading with a framework for how to think about safety, and even when we think about safety, there's multiple levels of it, right? Mm-hmm. There is, how do we protect ourselves against bad actors using AI? How do we protect against, uh, an environment that needs to be contained that is not contained, right? That is where the notifications. And then there is essentially the AI alignment challenge, right? Yeah. Which is scientifically a challenge because it's an experimentational sci- it's an experimental science. We don't yet have the science for really thinking about con- you know, really alignment."

Source: Deirdre Bosa · Sep 27, 2026 · 56s clip