Microsoft CEO Satya Nadella is urging businesses to approach artificial intelligence with caution, arguing that organisations should not automatically trust AI systems simply because they are developed by reputable companies.
In comments reported by Business Insider in October 2026, Nadella advocated applying a zero-trust mindset to enterprise AI. The approach assumes that systems must be continuously checked and restricted rather than granted unrestricted access based on trust alone.
His position reflects a growing concern across the technology industry: as AI becomes capable of taking actions across business applications, a mistake or security compromise could have consequences far beyond an incorrect chatbot response.
What zero trust means for AI
Zero trust is a cybersecurity principle built around the idea that no user, device or application should automatically receive unlimited access to a system.
Instead, access is granted according to identity, permissions, context and the sensitivity of the requested action. Organisations verify requests and limit the amount of damage that any compromised component can cause.
Applying this approach to AI means treating an AI agent as a system that requires controls, even when it is operating inside a company’s own environment.
An agent that summarises documents may need permission to read selected files. It does not necessarily need permission to delete company records, approve payments or send confidential information externally.
The principle is simple: give AI only the access required to complete its assigned task.
Why AI agents create new security challenges
Traditional software generally follows predefined rules. AI systems, by contrast, generate outputs based on learned patterns and the context they receive. Their responses can be unpredictable, particularly when they encounter unfamiliar instructions or misleading information.
This becomes more consequential when an AI model is connected to external tools.
An agent might be authorised to search email, update a project management system or retrieve customer information. If it misunderstands an instruction or follows malicious content embedded in a document, it could attempt an action that the user never intended.
A system that has excessive permissions can turn a relatively small error into a serious operational incident.
Businesses therefore need safeguards that operate independently of whether the AI model behaves as expected.
The controls companies should consider
Nadella’s reported position points towards several practical security measures.
First, AI agents should have narrowly defined permissions. Access to sensitive systems should be limited, and high-impact actions should require additional approval.
Second, organisations should separate AI decision-making from critical control systems. An agent may recommend an action, but a separate, deterministic process should verify whether the action is permitted.
Third, companies should maintain audit logs showing what the agent accessed, what it attempted to do and which actions were approved.
Businesses also need ways to stop or isolate an agent quickly if suspicious behaviour is detected.
These controls should be tested regularly. Simply writing an AI policy does not prove that the system can resist malicious instructions or prevent unauthorised actions.
Why this matters for the enterprise AI market
AI agents promise to automate work across customer support, software engineering, finance and administration. Their commercial value depends partly on how much responsibility companies are willing to delegate.
But businesses are unlikely to grant broad access to systems that cannot be monitored or controlled reliably.
Security may therefore become a competitive advantage for AI platforms. Companies that demonstrate clear permissions, transparent activity logs and reliable human oversight may find it easier to win enterprise customers.
The challenge is not to eliminate AI autonomy entirely. It is to establish boundaries that allow automation without surrendering control.
The bottom line
As AI agents gain access to more business systems, trust alone is not a security strategy. Companies need restricted permissions, independent verification and human oversight for consequential decisions.
