Microsoft CEO Satya Nadella is asking companies to stop trusting AI models blindly. In a long post on X titled "Models as Insider Risks in the Super Intelligence Era", he argues that frontier models, closed or open-weight, should be treated as insider risks, meaning actors with privileged access to corporate systems.
The starting point is a structural difference. In traditional software, behavior can be traced to a specific code path. With frontier models, Nadella writes, outputs cannot be attributed to particular training data or weight configurations. Yet these models are already being deployed as agents with access to the most sensitive data and the power to take mission-critical actions. "We simply can't outsource responsibility for what intelligence does on our behalf," he writes, adding that a model provider's assurances do not relieve the company of that duty.
The claim is not that models are inherently malicious. The risk, he says, is that any sufficiently capable actor with access to important systems can make mistakes or be compromised. Companies should therefore apply to AI agents the practices already used for powerful people: verify identity, limit privileges, log activity and create containment boundaries.
His central proposal is to separate the supply of intelligence from the authority over it. In practice, the controls defining what a model can do must sit outside the model: the model must be separated from the harness that orchestrates its work and from the action space it can access. Nadella also lays out seven design principles, including model diversity, so that no single model is responsible for an important outcome or for verifying its own work. Every meaningful model action must leave tamper-proof, human-readable evidence, and outcomes must be reproducible without relying on the model's own account. On containment, he says companies must assume a model is compromised, and an authorized person should always be able to pause or shut it down mid-task, like an emergency brake.
His conclusion: the most trustworthy Super Intelligence system will not be the one with the model we trust most, but the one that lets us trust the model the least.
Industry standards still to be written
Nadella also calls for industry standards where existing ones fall short. The post follows several September statements by the same CEO, including a code of conduct for Microsoft's MAI models that remains a draft. Who will write these standards, and when, is still unclear.
No comments yet. Be the first!