Unchecked AI agents could be disastrous for us all - but OpenID Foundation has a solution (www.zdnet.com)

The OpenID Foundation (OIDF) published new research warning that unchecked agentic AI—autonomous services acting on behalf of users—could rapidly multiply inside enterprises, gain access to sensitive systems, and even grant other agents access without human oversight. The paper sketches realistic attack/accident scenarios (e.g., employees provisioning inboxes to productivity agents) and explains how current ad-hoc controls won’t scale as agents outnumber humans. It flags the Model Context Protocol (MCP) as a double-edged enabler: MCP makes dynamic discovery and chaining of data, compute and models easier, improving agent utility but also amplifying unpredictability and risk because agents behave non-deterministically. To address this, OIDF recommends treating agents as first-class identities within enterprise IAM and extending existing open standards and governance tools. Key technical proposals include extending SCIM (System for Cross-domain Identity Management) for automated agent lifecycle provisioning and deprovisioning, integrating AI-specific guardrails into Identity Governance and Administration (IGA), and enforcing real‑time policies at the point of action (for example, automatic PII masking before data is sent to LLMs). The research calls for interoperable, open IAM standards and cooperative industry work so IT can regain predictability—centralized, policy-driven workflows, fine-grained authorization, and runtime controls—to safely scale agentic AI.