Nadella: treat every AI model like a compromised insider

The news

Microsoft chief executive Satya Nadella published an essay titled Models as Insider Risks in the Super Intelligence Era, arguing that companies deploying frontier models should treat them the way security teams treat a powerful employee who might be compromised. 'We must assume a model is compromised and contain it from the start,' he wrote, TechCrunch and CNBC reported. 'Think of it like an emergency brake.'

Why it matters

The framing puts controls outside the model, in logs, harnesses and off-switches, and it will likely show up in enterprise procurement checklists.

The details

Nadella lists seven principles, including separating the model from the harness that orchestrates it, tamper-proof logs of every meaningful action, independent auditability, incident disclosure, and containment, with an authorised human always able to pause or shut down a model mid-task.

Days later, Microsoft president Brad Smith told the BBC that voluntary pledges are not enough and called for binding global AI safety rules, outside approval before release and kill switches, TechCrunch reported.

What we could not check

• The essay names no Microsoft product, so it is unclear how the principles apply to Copilot today.
• We relied on reporting of the essay's text.