This article examines non-human identity (NHI) security as a critical 2026 cybersecurity challenge, arguing that AI agents, service accounts, and automated systems now outnumber human users in enterprises but remain governed by IAM systems designed for humans. The article identifies three core IAM assumptions that AI agents break—self-monitoring, credential rotation, and human ownership accountability—and presents five risk categories that emerge when NHI governance fails. The problem matters because uncontrolled non-human identities create an expanding attack surface that most organizations significantly underestimate and inadequately monitor.
- Scale and prevalence of non-human identitiesThe article reports that for every human user identity in a large enterprise, there are typically 10–45 non-human identities (service accounts, API keys, OAuth tokens, machine certificates, AI agent credentials). Most are undocumented, overprivileged, and completely unmonitored with no active owner. Discovery exercises typically find 3–5x more NHIs than security teams initially estimated, revealing a major gap in organizational visibility.
- Three foundational IAM assumptions AI agents violateTraditional IAM assumes identity holders self-monitor and report unusual activity (agents operate 24/7 and cannot self-report compromise), that credentials are static and periodically rotated (agents may generate and discard credentials dynamically), and that there is a responsible human owner accountable for each identity (AI agents are often deployed and abandoned with no clear owner). Every human IAM control that depends on a human noticing, reporting, or triggering an event fails for AI agents.
- Five categories of NHI risk in 2026The article defines: orphaned credentials with no active owner but still-valid permissions; over-provisioned AI agents granted broad permissions for flexibility rather than specific tasks; hardcoded credentials embedded in code, configs, or environment files; no agent lifecycle management leading to indefinite operation after project ends; and agent impersonation where attackers steal agent credentials to access systems directly without defeating MFA. Each category creates distinct attack vectors with different remediation approaches.
- Real attack scenarios enabled by NHI gapsThe article provides three composite scenarios: an orphaned integration account with read/write database access and no rotation or monitoring enabled exfiltration for 11 days undetected; a prompt-injected AI customer service agent with broad mailbox, Jira admin, and unrestricted email sending permissions was manipulated to export support ticket contents; and AWS credentials stored as CI/CD environment variables were exfiltrated via supply chain compromise to fund cryptocurrency mining. None required software vulnerability exploitation—all were credential and permission governance failures.
- Framework for AI agent identity managementThe article presents five principles: every agent must have a named human owner responsible for permissions review, monitoring, and decommissioning (with review required when owner leaves); permissions must be scoped to specific tasks, never granted broadly for future flexibility, with quarterly permission reviews removing unused access; use short-lived credentials (OIDC/managed identities) instead of static API keys where possible, with rotation enabled for static credentials stored in secret managers; monitor agent actions like privileged users with SIEM logging and alerts for out-of-pattern
- Practical NHI inventory methodologyThe article specifies where to find non-human identities: cloud identity providers (Azure AD service principals, AWS IAM roles and non-human users, Google Cloud service accounts), code repositories (GitHub secrets scanning and bot accounts, grep patterns for hardcoded credentials), CI/CD pipeline environment variables, and surveying development teams about deployed AI agents. For each identity found, document owner, purpose, permissions, last used date, and any expiry. This inventory exercise typically reveals the organization's true NHI population and defines programme priorities.
- Permission scoping as blast radius controlThe article emphasizes that most teams over-provision agent permissions significantly—granting broad mailbox access when agents only need to read specific senders, broad Jira access when only one project needs tickets, and unrestricted email permissions when only sender replies are needed. Each over-broad permission multiplies the blast radius if the agent is compromised, making explicit minimal permission scoping the most valuable security control in agent identity policy design.
- Orphaned credentials as the most persistent riskThe article identifies orphaned agents deployed for projects or proof-of-concepts that kept running after the developer moved on as the most consistently encountered problem in assessments. These agents still have valid credentials and permissions but no one monitoring them. The author recommends quarterly NHI audits as standing security activity with every agent requiring a named owner in an ownership registry; when the owner leaves the organization, the agent is reviewed on the offboarding checklist.
Bottom line: Organizations must immediately inventory their non-human identities across cloud IAM, code repositories, and development teams—a process that typically reveals 3–5x more identities than estimated—then implement explicit ownership, minimal permission scoping, automated lifecycle management, and continuous monitoring for each AI agent and service account. Without these governance controls substituting for human-centric IAM assumptions, the expanding population of always-on, unmonitored automated i

