A technical deep-dive into Microsoft's AI security architecture — covering the full stack from identity primitives and MCP attack surfaces to runtime threat detection and regulatory framework alignment. Caveats on current limitations are shown throughout.
Security Dashboard for AI is now generally available. Entra Internet Access Shadow AI Detection and Prompt Injection Protection both GA on March 31. New Security Analyst Agent and Security Alert Triage Agent announced. Sentinel MCP Entity Analyzer GA in April. Several new Purview and Entra capabilities added across preview and GA. See the Gaps & Roadmap page for the updated availability matrix.
Organisations are deploying AI copilots and autonomous agents at scale to automate decision-making and access enterprise data. Standards like Model Context Protocol (MCP) now let AI systems invoke real enterprise tools — email, file systems, APIs, SaaS platforms — not just generate text. This transforms the AI agent from a productivity interface into a privileged digital actor operating inside the enterprise perimeter.
80% of Fortune 500 companies are already using agents according to Microsoft's research. Industry projections estimate over one billion AI agents in enterprise environments by 2028. Traditional security models built for users, endpoints, and applications break down when the actor is non-human, persistent, autonomous, and potentially opaque.
An AI agent blends a user (accesses data, makes decisions), an application (runs code, calls APIs), and a service account (operates non-interactively, often persistently). No single existing security primitive handles all three. Microsoft is building toward this with Entra Agent ID — but it remains in limited preview for frontier customers only. Today, agents operate under OBO (On-Behalf-Of) delegation — inheriting the invoking user's identity and permissions, not a purpose-scoped identity of their own.
Microsoft secures AI by bringing identity-first security, runtime threat protection, and unified visibility to AI workloads, agents, and MCP ecosystems — extending Zero Trust across the full AI lifecycle. The strategy is architecturally sound. The execution is partially complete — agent identity (OBO) and per-agent licensing remain the critical structural gaps.