The more a model can do, the more depends on the architecture around it. On 10 October 2026, Satya Nadella published “Models as Insider Risks in the Super Intelligence Era” on X. Microsoft's CEO describes how organisations can give Super Intelligence (SI) systems access to data and actions and keep control. CNEXT shares this view; it matches how we build AI-native solutions.
What Satya Nadella describes
With traditional software, you could trace behaviour back to a code path. Today's SI systems lack that mechanistic understanding, even though frontier models are far more capable. Agentic systems still receive access to sensitive data and mission-critical actions. Nadella concludes that organisations need to rethink their trust architecture. Responsibility for what an intelligence does on their behalf stays with them. A model provider's assurances do not take it away.
His proposal: build SI systems whose behaviour can be observed, whose limits can be tested and whose actions can be contained. The supply of intelligence and the authority over it belong in different hands. Nadella deliberately sets the question of alignment aside and starts with engineering: deterministic system design around non-deterministic models, human controls, reliable operating procedures, and industry standards where none exist yet.
The core of the post is a change of perspective. Closed and open-weight frontier models should be treated like insider risks. The reason lies in access: any capable actor with access can make mistakes or be compromised, whatever its intentions. Organisations have handled powerful insiders for decades, with identity, least privilege, logging and clear boundaries.
Nadella considers transparency into a model's chain of thought non-negotiable, yet insufficient on its own. Models can test each other adversarially, but then one opaque model is watching another. Controls over access and actions therefore belong outside the model, separated from the harness and the action space, an information security principle dating back to the 1970s.
The principles
Summarised in our own words:
- 1Model diversity: No model is the sole dependency, and none verifies its own work.
- 2Observe everything: Every meaningful action leaves tamper-proof, human-readable evidence that can be reproduced without the model's own attestation.
- 3Verifiability: The whole system is tested continuously, including failures, attacks, edge cases and changes.
- 4Independent controls: The organisation decides what a model can access and what it may do.
- 5Independent auditability: Validation is independent of the intelligence being validated.
- 6Containment: Compromise is assumed from the start. An authorised person can always pause or stop a model mid-task, like an emergency brake.
- 7Incident disclosure: What failed is disclosed promptly and shared across the industry.
His closing thought: the most trustworthy SI system is the one that requires the least trust in its model.
Our view: AI-native means rebuilding the work
We see every day how much SI can already do. Agents research, write code, reconcile contracts, prepare decisions and work directly in business systems. That is why we build AI-native solutions: we design a workflow so that agents carry it out from the start and people decide where it matters. Existing processes were built for people and their handovers, and they serve that purpose well. Bolting AI onto them afterwards achieves little.
We expect that, over time, one agent will win: one capable agent per person or organisation that does most of the work. We work with frontier organisations in the sense of Microsoft's Frontier Firm, where people and agents already share workflows, as our Agentic AI page describes.
Because such an agent receives broad access, AI-native work needs strong architecture. Nadella's principles describe what it has to deliver, and they shape every Enterprise AI platform we set up.
How we put the principles into practice
In the Microsoft stack, we translate the principles into concrete building blocks:
- An identity for every agent: With Microsoft Entra Agent ID, every agent gets its own governed identity. Conditional Access, permissions and sign-in logs apply per agent, just as they do for employees. The extended security features for agents require Microsoft Agent 365. More under Zero Trust & Identity Security.
- Inventory and control: The agent registry in the Microsoft 365 admin center, part of Agent 365, lists every agent with its owner and risk signals. Agents without an owner can be blocked or deleted.
- Logging and retention: Microsoft Purview records agent interactions in the audit log, and Data Security Posture Management shows which agents work with sensitive data. Agents can even be included in Insider Risk Management policies, fitting Nadella's framing.
- Least privilege and the action space: Connectors and MCP tools receive only the permissions a task needs. In Microsoft Foundry, we restrict an MCP server's tools to an allow list. Which content an agent can see at all is settled during AI data preparation.
- Human approval: Consequential actions such as a payment, an email to a customer or a deletion wait for a person. In Foundry, the MCP tool requires an approval for them; in Copilot Studio, an agent flow pauses with a request to the responsible reviewer.
- Model diversity: Foundry offers models from OpenAI, Anthropic, Mistral AI, Meta, DeepSeek, Cohere and xAI. We have work checked by a different model from the one that produced it.
- Continuous testing: Maester regularly checks your tenant configuration against Microsoft's recommendations and your own rules, as part of our Security Operations. For agents, we add test sets with edge cases and manipulated inputs.
- Emergency stop: In the admin center, agents can be blocked and Foundry agents stopped. We also build the mid-task stop into the harness: revoke tokens, end sessions, halt scheduled jobs.
What you can check now
If you run agents in production, Nadella's principles make a useful checklist. Do you know which agents are running and who owns them? Does every agent have its own identity with minimal permissions? Are logs produced outside the agent? Can an authorised person stop an agent, and has that path been rehearsed? An assessment shows where you stand. With AI governance, the answers become rules and evidence for your board and auditors.
Capable models and strict controls belong together. The more work an agent takes on, the more valuable an architecture becomes that places as little blind trust in it as possible.
