The Shadow AI Workforce: Governing AI Agents and Non-Human Identities Across AWS

In a mature AWS estate, the humans are a rounding error. The population that actually holds privilege is roles, workloads, service accounts, CI/CD runners, Lambda execution roles, and now AI agents that assume roles and call tools on their own. Most of it was provisioned outside any joiner-mover-leaver process, carries broad or wildcard permissions, and never gets deprovisioned. Agents make it sharper: a workload chaining sts:AssumeRole or invoking a tool with a scoped-too-wide credential is a privilege decision no human ever reviewed. The table compares how teams are discovering the non-human identities they already have, scoping and expiring that access, and deciding what an autonomous agent is actually allowed to do, without grinding the platform to a halt.