AI agents are becoming identities. Someone needs to own that risk.
Every AI agent your team builds on Azure AI Foundry or Microsoft Security Copilot creates a real identity in Microsoft Entra ID, with its own permissions and, usually, nobody responsible for it. AI Security is where ID Posture governs that risk, alongside every other identity in your tenant.
Why this is a real, distinct risk
Most AI agent platforms hand you a working agent fast. What they don’t hand you is a governance model. Microsoft’s own answer to this is Microsoft Entra Agent ID, a real identity object type, agentIdentity, alongside a blueprint object, agentIdentityBlueprint, that defines what every agent minted from it can do. These aren’t hypothetical. Azure AI Foundry and Microsoft Security Copilot both create them today, and most organisations have no visibility into how many exist in their own tenant, let alone who’s accountable for them.
ID Posture reads these objects the same way it reads every other identity in Entra ID, no new permissions, no separate connection. It’s the one place agent identity governance sits inside a scored pillar rather than a separate add-on category: this is Non-Human Identity risk, not a resource-configuration checklist, so it counts toward your actual posture score.
What’s covered today
Agent identity governance is the first area of AI Security coverage, checking whether each agent identity has an accountable human sponsor, whether an agent’s blueprint is owned by a person rather than another machine identity, and whether a blueprint holds Microsoft Graph permissions broad enough that every agent minted from it inherits the risk. See what ID Posture checks on every AI agent identity.
This is an admin-controlled, opt-in check, off by default until your team turns it on. More AI Security coverage is planned as the category matures, following the same pattern as every other area of the product: real findings, grounded in what Microsoft actually exposes, never a fabricated pass.
Connect your Entra ID tenant and see whether your organisation’s AI agent identities have real, accountable owners, no sample data, no demo environment.