On "Satya Nadella says we should assume all AI models are ‘compromised’"

Nadella's right that we need to stop treating models like oracles, but "assume all AI is compromised" is theater unless you're actually prepared to audit what your models are doing—and most organizations aren't. The real work isn't philosophical acceptance of danger; it's building testing infrastructure, monitoring outputs against your actual business logic, and having the discipline to reject a model's answer when it contradicts your values or your data. That's not sexy, which is probably why a CEO frames it as an existential stance instead.