Overview
- Researchers have shown practical attacks on agentic workflows, including a DNS TXT hijack that took over an AI coding assistant, a 'Poisoned Tenant' invitation campaign that granted owner access, and a CVSS 8.5 flaw in Amazon Q Developer that has since been patched.
- Enterprises commonly grant AI agents broad, long‑lived access to email, files, CRM and code while lacking visibility into agent actions, creating what researchers call 'identity dark matter' where non‑human identities sit outside traditional IAM controls.
- Security and industry guidance now pushes a layered operational model: discover and classify agents, assign owners and governance metadata, use non‑human identity (NHI) platforms for credentials, apply runtime validation for action‑level checks, and deploy observability for forensics.
- Open‑source toolkits such as Microsoft’s Agent Governance Toolkit have legitimized runtime enforcement, and a growing vendor stack — NHI platforms, runtime validators, guardrails and tracing tools — is emerging, but many organizations still lag on policy authorship and adoption.
- The shortfall matters for people and regulators because uncontrolled agents can move money, expose sensitive records, or alter customer data; expect tighter audit expectations, cross‑functional governance led by CIOs/CISOs and business owners, and more rules from standards bodies.