Satya Nadella says we should assume all AI models are compromised

Microsoft chief executive Satya Nadella has said that all Artificial Intelligence (AI) models should be treated as “compromised” and potentially dangerous – in other words as a new form of ‘insider threat’ which is something I highlighted years ago in front of audiences in Prague …

In a lengthy post to X about AI safety, the tech boss said that containment of rogue AI models will only be possible if safeguards are built-in from the start. His comments came after Microsoft President Brad Smith called for global safety rules for the technology, claiming that milk and washing machines are currently more regulated than AI.

“We must assume a model is compromised and contain it from the start,” Mr Nadella wrote. “Think of it like an emergency brake. Any authorised person should always be able to pause or shut down a model mid-task. More advanced models will require more advanced containment algorithms and technologies that we need to standardise on.”

The Future of AI Keynote: Dystopia, Utopia and Control, a keynote by futurist Matthew Griffin

He added that AI companies need to take an engineering approach to safety and that models should be separated from the mechanisms that govern its actions.

Nadella referred to the technology as “super intelligence”, following an executive order from US President Donald Trump to no longer refer to it as artificial intelligence.

He appeared to distance himself from Trump’s call for AI firms to regulate themselves, saying that stricter industry standards are needed to prevent bad outcomes. A day before Mr Nadella’s post, Brad Smith told ITV News that we need to act urgently to address the greatest risks associated with AI.

He warned that the sector is currently subject to less regulation than everyday consumer products due to the pace at which it is developing.

“Right now AI is less regulated than your washing machine,” Smith said. “It is less regulated than a carton of milk. We all go into the grocery store, we look at the milk, we know it's safe. That is the kind of trust we need to build for AI as well. As these technologies spread more broadly, we either address this now, we either put the kind of safeguards in place that we need, or we run a greater risk."

In a separate interview with the BBC, Smith repeated Nadella’s calls for “emergency brakes” to stop out-of-control AI systems, and it follows broader industry calls from the likes of Google DeepMind and Anthropic to develop a “kill switch” for AI.

The idea has gained traction in recent weeks following a string of cyber attacks and other incidents involving frontier AI models.

First published on Fanatical Futurist on 10 October 2026.

311 Intelligence

What this means

Analysis by the 311 Strategic Intelligence Unit

The big shift is from trusting AI models by default to treating them as a possible insider threat. Microsoft's CEO wants every model contained from the start, with an emergency brake that any authorised person can pull. Our Starburst data shows how far there is to go: we rate Containment Algorithms, ways to keep a strong AI inside set limits and shut it down if it goes past them, a Watch technology at TRL 2, and we expect wide use only by 2045. We expect Chain-of-Thought Monitoring, a second AI reading an agent's step-by-step thinking, to reach wide use sooner, by 2030.

What it means by sector.

• Technology: Nadella says models should be separated from the mechanisms that govern their actions. AI vendors will be asked to prove their models can be paused mid-task.
• Every organisation using AI: if the maker says to assume a model is compromised, buyers should treat AI access like any privileged user, with limits and logs.
• Governments: Brad Smith says AI is less regulated than a carton of milk, while President Trump has called for AI firms to regulate themselves. We track Binding AI Laws as a Prepare trend, with wide use expected by 2028, and Deregulation for Growth as an Act now trend.

What to do now.

• Check that every AI agent you run can be paused or shut down by a named person, mid-task.
• Give AI systems the least access they need, as you would any insider with privileged rights.
• Watch for containment standards from Microsoft, Google DeepMind and Anthropic, and plan to adopt them.

The Starburst technologies behind this story

From the 311 Technology Starburst (first product, expected wide use and readiness):

Technology	Category	First product	Wide use	Action	TRL
Containment Algorithms	Security	2032	2045	Watch	2
Chain-of-Thought Monitoring	Security	2026	2030	Prepare	5
Runtime Assurance Monitors	Security	2014	2033	Explore	8
Agentic Firewalls	Security	2027	2035	Explore	7

And the trends they feed, from the 311 Trends Starburst:

Trend	Wide use	Action
Binding AI Laws	2028	Prepare
Deregulation for Growth	2027	Act now
AI Cyberattacks	2028	Prepare

How this moves the trend

Microsoft has pushed for shared rules before. In 2017 it called for a Digital Geneva convention to limit cyber attacks. In 2024 researchers argued that controlling self-improving super intelligence is probably impossible, and AI developers were warned about insider threats. Nadella now treats the model itself as the insider.

Go deeper with 311

• The 311 Starburst Centre: every technology and trend above, with timelines and readiness.
• 311 Vantage: ask what AI containment means for your organisation and get a plain English executive brief.
• The Exponential Technology Codex and the Annual Trends Codex: the technologies and trends reshaping every industry.
• Book a briefing or keynote on the future of AI and AI safety for your board or leadership team.