For the past few years, most conversations about AI security have focused on protecting information. We worry about prompt injection, sensitive data exposure, model security, and hallucinations. Those concerns are important, but they are increasingly only part of the story.
Enterprise AI is evolving from generating information to taking action.
Modern AI systems can retrieve enterprise data, invoke APIs, update business applications, execute workflows, and coordinate activities across multiple systems. Standards such as the Model Context Protocol (MCP) are accelerating this shift by making it easier for AI to interact with enterprise tools and services.
This changes the security conversation.
When AI begins acting on behalf of users, traditional application security alone is no longer enough. We need governance that answers three fundamental questions:
- Who can act?
- What actions can they perform?
- How do we continuously verify trust?
In my latest white paper, Securing the AI Action Plane: An Architectural Framework for Identity, Governance, and Trust in Enterprise AI, I introduce the concept of the AI Action Plane and present a governance framework built around three architectural pillars:
- Identity-Centric Governance
- Constrained Execution
- Continuous Assurance
Although the paper uses MCP as a motivating example, the framework is intentionally technology agnostic. These principles apply to any enterprise AI system capable of acting on behalf of users, regardless of the underlying model, vendor, or orchestration framework.
I believe the next generation of AI security will be defined less by protecting models and more by governing AI behavior.
I’d love to hear your thoughts.