Shadow AI Agents Accumulate Unaudited Service Account Privileges Across Enterprise Systems

Enterprise security teams initially blocked public AI domains to contain data leakage. That perimeter no longer contains the problem once agents connect directly to production APIs via stored credentials. The research documents agents created through SaaS automation features, MCP servers, and endpoint scripts that retain read-write-delete scopes long after their builders rotate roles. Traditional IAM policies assume human-initiated, deterministic sessions; agent workflows chain multiple systems under single inherited identities, bypassing review gates. Visibility gaps widen when agents run on service accounts provisioned through developer tools rather than central identity providers. Six months after deployment, ownership attribution becomes unreliable, allowing dormant agents to trigger downstream automations without fresh authorization. Access control failures compound faster than data-loss incidents because one agent can modify configurations across connected platforms before any alert fires.