Nadella: "assume all AI models are compromised"
From the AI digest for 2026-10-11
What happened: Microsoft CEO Satya Nadella published a lengthy X post (~6:10 PM ET Oct 10, The Verge's writeup) arguing the industry can no longer accept AI as a "set of nested black boxes" whose advice is accepted or rejected wholesale; models should be contained, observable, and leave behind "tamper-proof human-readable evidence."
What's genuinely new: A development of the agent-privacy/security frame the record carried this week — now stated by the CEO of the company shipping agents at consumer scale. The concrete asks (containment, observability, auditable evidence trails) map onto the observability lane already on record (METR, Oct-6). It is a posture, not a shipment — filed accordingly.
Watch: Microsoft shipping anything to that spec (agent products with evidence trails); harness-vendor uptake; whether the security community reads it as roadmap or deflection.
Links: The Verge — Satya Nadella says we should assume all AI models are 'compromised'