Non-human and AI agent identity, by ear

Why 'machine identity' is a loose label for devices, services, and workloads, and why their scale requires automated lifecycle and secrets management, with regulation still catching up.

How machine identities differ from human ones, why static credentials like API keys and service accounts create technical debt, and the case for dynamic, ephemeral credentials.

The main ways attackers find and exploit secrets and credentials leaked in source code, and how developers can defend against those attacks.

How service principals, app registrations, and agent identities become hidden admins, which API permissions to hunt for, and why to replace secrets with managed identities.

Why workload identities should be treated as privileged, how long-lived secrets widen the attack surface, and how ephemeral secrets, SPIFFE/SPIRE, and policy-as-code help.

What SPIFFE is, how it places a workload or AI agent identifier inside a certificate, and how policy engines can allow-list those identifiers.

Why non-human identity risk built up while human identity matured, covering service accounts, unrotated API keys, hardcoded credentials, and zero trust for AI agents.

How GitHub Actions gets short-lived AWS credentials without stored keys: OIDC provider setup, role trust policy conditions on aud and sub claims, JWT signing, and AssumeRoleWithWebIdentity.

Where authentication and authorization standards fall short for non-human identities and AI agents, and which emerging frameworks and external guardrails aim to close the gap.

Identity challenges raised by AI agents, the role of MCP, and recursive delegation and scope attenuation, drawing on the OpenID Foundation agentic AI whitepaper.

Why AI agents need each action authorized in real time, why few MCP servers implement the OAuth the MCP spec requires, and how to roll agents out from zero privileges.

How existing OAuth specifications such as JWT authorization grant, token exchange, and client ID metadata, plus SPIFFE, apply to securing AI agents as workloads.

Why MCP security is an API, identity, and authorization problem, covering unauthenticated MCP servers, tool poisoning, prompt injection, and practical next steps.

How Entra Agent ID blueprints, agent identities, and agent users relate, and how stolen blueprint credentials could reach agent identities in other tenants.