Context

The organisation had an established zero trust security posture for users, devices, and traditional applications. As AI systems and agent based capabilities began to operate across enterprise environments, they increasingly interacted with sensitive data and critical systems, often acting on behalf of users or processes. These systems did not fit neatly into existing trust models. Once authenticated, AI components were often implicitly trusted, even though their behaviour could change over time as data, models, and workflows evolved.

‍

The Challenge

Zero trust had been designed around relatively stable actors: human users and well defined services. AI systems blurred these assumptions. Agents could initiate actions, chain decisions, and operate continuously without direct human involvement. Granting them static permissions or broad trust undermined the intent of zero trust, while attempting to apply human centric controls directly created friction and impractical oversight. The organisation faced a growing gap between its stated security principles and how AI systems were actually trusted in practice.

‍

The Decision

The organisation chose to extend zero trust concepts explicitly to AI and agent based systems, rather than treating them as special cases. Instead of assuming that AI systems could be trusted once authenticated, they applied principles of continuous verification and least privilege to how agents accessed data and systems over time. Trust boundaries were made explicit: what an agent could do, under what conditions, and when access should be re evaluated or constrained. They deliberately rejected both extremes-implicitly trusting internal AI systems, and attempting to lock them down with static, overly restrictive controls.

‍

What Changed

AI systems were no longer treated as permanently trusted actors once deployed. Access decisions became more contextual and time bound, reducing reliance on long lived permissions. Teams designing agents were forced to be clearer about intent, authority, and scope, rather than assuming broad access would be acceptable. Security teams gained better visibility into how AI systems behaved in practice, without needing to introduce constant manual review. Some delivery paths became more deliberate, but trust in running AI systems at scale improved.

‍

Why This Matters

As AI systems become more autonomous, the risk is not just unauthorised access, but misplaced trust. Zero trust principles lose meaning if they stop at the edge of AI workloads. Extending these principles to agents and AI systems allows enterprises to scale automation without creating blind spots where behaviour is assumed rather than verified. The challenge is not adopting zero trust in name, but ensuring it applies to the systems that increasingly act without humans in the loop.