Your identity governance program almost certainly has a blind spot. It is not a configuration error or a policy gap. It is architectural.
The IGA platforms most enterprises run were designed when “identity” meant a person. Joiner-mover-leaver workflows, access certifications, role mining, entitlement reviews — every process assumes a human on one end of the access relationship. That assumption held for decades. It no longer holds.
The Cloud Security Alliance’s May 2026 briefing puts the current ratio at 45 non-human identities (NHIs) for every human identity in the average enterprise — rising to 144:1 in cloud-native environments. ManageEngine’s Identity Security Outlook 2026, a survey of identity and security leaders, found that nearly half of organizations see ratios above 100:1, with some sectors reaching 500:1. Between the first half of 2024 and the first half of 2025 alone, the average enterprise saw a 44% increase in NHI volume.
Your IGA program was built for the 1. The 45 are largely ungoverned — and AI agents are about to make that ratio worse.
Why NHI Governance Has Fallen Behind
The governance gap did not appear overnight. It accumulated gradually as organizations added cloud workloads, SaaS integrations, CI/CD pipelines, and automation frameworks — each of which minted new machine identities — while IGA programs remained focused on human access reviews.
The result is a sprawl problem with a consistent fingerprint across enterprises: service accounts that outlived the projects that created them, API keys hardcoded into repositories years ago that are still valid today, OAuth tokens granted to third-party integrations that no one has audited since, and a credential inventory that nobody owns.
GitGuardian’s 2025 research cited by NHI Management Group found 28.65 million new hardcoded secrets committed to public GitHub repositories in 2025 alone. More troublingly: 64% of secrets first exposed in 2022 were still valid in 2026. Credentials are not being rotated. They are not being revoked when projects end. They are sitting in repositories, in configuration files, and in CI/CD pipelines — fully functional, fully exploitable, and invisible to most IGA platforms because those platforms were never configured to look for them.
This is not a secrets management problem in isolation. It is an identity lifecycle problem. The credentials are still valid because the identity lifecycle process that would have flagged them for review — or triggered decommissioning — was never extended to cover machine identities. The joiner-mover-leaver process ends at humans. Machine identities have no equivalent offboarding trigger.
The Attack Surface This Creates
Attackers have noticed the gap. Mandiant’s M-Trends 2026 report, drawing on more than 450,000 hours of incident response, confirms that machine identities have become a primary attack vector in cloud breaches. Threat actors are harvesting long-lived OAuth tokens, compromising third-party SaaS vendors to steal hard-coded keys, and then using those credentials to move laterally through environments at a scale and speed that human-account-focused detection misses.
SpyCloud’s 2026 Identity Exposure Report, via The Hacker News, flagged non-human identity theft as one of the fastest-growing categories in the criminal underground — with a third of recovered non-human credentials tied to AI tools.
Sophos’s State of Identity Security 2026, a survey of 5,000 security leaders across 17 countries, found that 71% of enterprises experienced at least one identity-related breach in the past year, at a mean recovery cost of $1.64 million. Weak NHI management was cited as a contributing factor in 41% of those incidents. And only 34% of organizations regularly audit or rotate their non-human credentials.
Read that last number carefully. In an environment where machine identities outnumber humans by 45 to 1 and are behind 41% of breaches, only 34% of organizations have a regular audit and rotation cadence for them. The attack surface is vast, growing, and largely unreviewed.
The CSA briefing adds a dimension that matters particularly for organizations with agentic AI deployments: AI agents dynamically acquire permissions at runtime, spawn sub-agents, and take autonomous action. Unlike static service accounts — which are at least predictable — AI agents create a shifting access profile that pre-AI governance processes were never designed to track. A static quarterly access review does not catch an agent that spun up, acquired elevated permissions, completed a task, and went dormant — all in the same afternoon.
What Makes NHI Different from Human Identity Governance
Understanding the governance gap requires understanding why NHI is structurally different from human identity governance — not just a larger version of the same problem.
No natural lifecycle events. Human identities have built-in lifecycle triggers: hire, role change, departure. These events prompt IGA workflows. Machine identities have none of these. A service account created for a project three years ago has no equivalent of “employee resignation” to trigger decommissioning. Without explicit lifecycle management, machine identities accumulate indefinitely.
No authentication behavior that signals compromise. When a human account is compromised, behavioral signals often emerge: unusual login times, impossible travel, failed MFA. Machine accounts authenticate continuously and predictably — that’s their function. A compromised service account looks exactly like a functioning service account, which is why Verizon’s DBIR found credential abuse accounted for 22% of breaches in 2025 with machine credentials a growing share.
No single owner. Human identities belong to a person. Their manager is accountable for their access. Machine identities are frequently orphaned — created by a developer who left, for a project that ended, integrated into a system no one fully understands anymore. An AgentMode AI analysis of the CSA’s NHI governance research found that 78% of organizations have no documented policy for creating or removing AI identities, 51% report no clear ownership, and only 20% have a formal process to offboard and revoke API keys — figures that cover the AI identity subset specifically, not the full NHI estate.
Context must come from runtime behavior. The CSA’s framework for NHI governance is precise on this: for human identities, access decisions can be derived from role and organizational context. For NHIs, context must be derived from runtime behavior — the chain of relationships between application, secret, identity, and resource. That requires continuous monitoring and access-path mapping, not periodic certification.
The Four Governance Controls That Actually Move the Needle
The IGA vendors are adding NHI modules. The PAM vendors are expanding into machine identity. The secrets management platforms are building governance layers. The market is responding to the problem — which means the tooling is catching up, and the priority for security architects is understanding which controls matter most and in what order.
Based on the current incident data and the governance frameworks emerging from CSA, CyberArk, Saviynt, and EIC 2026, four controls consistently appear as the foundation:
- Build a Complete, Continuous Inventory
Your existing IGA platform almost certainly has incomplete NHI coverage. It may govern the service accounts in Active Directory. It likely does not govern API keys in SaaS platforms, OAuth tokens in developer tools, certificates across multi-cloud environments, AI agent tokens, CI/CD credentials, and hardcoded secrets in code repositories.
A complete NHI inventory requires discovery that spans all of those surfaces simultaneously. The inventory record for each NHI should include: type (service account, API key, OAuth token, certificate, agent credential), owning system, named business owner, named technical custodian, purpose, last-seen activity, credential expiry date, and access scope.
Ownership is the most important field. Without a named owner, there is no accountability for lifecycle decisions. The NHI Management Group analysis is explicit: every machine identity needs a named business owner and technical custodian — and that ownership should be reconciled during joiner, mover, and leaver events so machine identities do not drift into orphaned status when people leave.
- Apply Lifecycle Controls With Explicit Decommissioning Triggers
The lifecycle gap for NHIs is the governance equivalent of a leaver process that nobody runs. Machine identities accumulate because there is no event that triggers review.
The fix is to build explicit decommissioning triggers into the governance process:
- Project end: All NHIs created for a project scope are flagged for review when the project closes
- Owner departure: Any NHI owned by an employee who leaves is automatically queued for ownership transfer or decommissioning
- Inactivity threshold: Any NHI that has not been used within a defined window (30–90 days depending on type) is flagged as a candidate for decommissioning
- Integration retirement: When a SaaS application or API integration is retired, all associated tokens and keys are swept
This is the same process logic as joiner-mover-leaver, applied to machine identities. The triggers are different, but the governance discipline is identical.
- Enforce Least Privilege and Time-Bounded Scope
Saviynt’s posture analysis via NHI Management Group found that 97% of NHIs carry excessive privileges — meaning the default state of machine access in most enterprises is already too broad. That is not a configuration error on a subset of identities. It is the baseline condition of the estate.
The remediation approach that works at scale is blast-radius reduction: prioritize NHIs by how far their access can spread before a human can intervene. High-blast-radius identities — those with wide reach, shared credentials, or access to sensitive data — get tightened first.
For AI agents specifically, the governance guidance emerging from EIC 2026 is consistent: ephemeral, just-in-time access rather than standing privileges. Issue credentials that expire with the task. The agent receives project access when its task starts; that access is removed when the task ends. Standing access for idle agents is unnecessary exposure — and at machine speed, unnecessary exposure is a liability.
- Continuous Monitoring and Access-Path Mapping
Periodic access reviews work for human identities because human access profiles change slowly and lifecycle events provide natural review triggers. NHIs change continuously — new tokens are minted, scopes shift, integrations evolve. A quarterly certification cycle cannot keep pace.
The architecture that works is continuous access-path mapping: tracing how each NHI and AI agent reaches high-value resources, including indirect paths through tokens, workloads, and delegated tools. The goal is not to audit all 45 NHIs per human — it is to know, at any moment, which NHIs can reach sensitive systems and whether those paths are appropriate given current business context.
This is what the CSA’s NHI governance framework, as published by Oasis Security, calls “continuous validation rather than periodic approval”: after an identity is created, governance shifts to policy-driven enforcement informed by real usage and dependency data — identifying over-privileged identities, rotating credentials when risk changes, flagging unused access, and safely decommissioning identities when their consumers are retired.
For the AI agent subset of the NHI estate, continuous monitoring has a dimension that traditional machine identity controls don’t cover: what the agent actually does with its access at runtime. Identity lifecycle governance answers who the agent is and what it can reach. Prompt-level governance answers what it sends, what it returns, and whether its behavior stays within its declared operating envelope. Those are separate control problems requiring separate enforcement points. Platforms entering this space — including SafePrompts.ai from TVR Labs, which is building identity-aware prompt governance across ChatGPT, Copilot, APIs, and enterprise AI systems — produce the session-level audit trail that the broader NHI governance program needs to demonstrate control at the execution layer. The identity governance program scopes the agent and provisions its credentials. The prompt-layer enforcement captures what happens next.
AI Agents: The New NHI Frontier
Everything above applies to the NHI estate that already exists in your environment. AI agents extend the problem in a direction that makes the existing gap more urgent, not less.
AI agents are non-human identities with properties that make them harder to govern than traditional machine identities:
- They dynamically acquire permissions at runtime rather than holding a static access profile
- They spawn sub-agents — meaning one provisioned identity can create a chain of downstream identities with their own access
- They process external data as part of their function, which means their behavior can be altered by the content they ingest (prompt injection)
- They operate at machine speed between human decision points
CyberArk’s analysis via NHI Management Group frames the governance requirement precisely: assign AI agents an owner, a purpose, an access boundary, and an expiry condition — and review those attributes continuously. Treat the agent as a machine identity with task-scoped authority rather than a generic application component. That approach makes revocation, auditability, and containment possible when the agent behaves unexpectedly.
The concept that matters most for agent governance is identity blast radius: the damage from a compromised or over-scoped agent is determined by how far its credentials, permissions, and integrations can reach before revocation. That is a measurable governance problem. Scoping it down before deployment — constraining what the agent can reach structurally — is the primary risk control.
The IGA vendors moving fastest on this — Saviynt, Omada, SailPoint — are building governance models that treat AI agents as first-class NHI types, with the same lifecycle discipline as privileged human accounts plus tighter runtime oversight. The EIC 2026 consensus was clear: first-class verifiable identities rather than shared secrets; runtime authorization checking every action rather than just registration; traceable delegation through signed receipts and standards like OAuth 2.1 and A2A.
Where to Start: The NHI Governance Sequencing
For identity architects rebuilding governance programs to cover NHIs, the sequencing that produces results:
Step 1 — Discover before you govern. Run a full NHI discovery sweep before writing policy. You cannot govern what you cannot see, and the NHI surface is larger than most programs assume. Prioritize: cloud IAM roles and service accounts, SaaS OAuth connections, CI/CD credentials, API key repositories, and AI agent tokens.
Step 2 — Assign ownership to every discovered NHI. An inventory without owners is a list, not a governance program. Every NHI gets a named business owner and technical custodian before any other governance work proceeds.
Step 3 — Rank by blast radius and remediate the top tier. Not all NHIs carry the same risk. Focus first on NHIs with access to sensitive data, wide reach, shared credentials, or no rotation history. That is your highest-impact remediation surface.
Step 4 — Build decommissioning triggers into existing lifecycle processes. Add NHI sweep steps to the project close-out process, the employee offboarding process, and the quarterly IGA review cycle. This stops the accumulation without requiring a separate NHI governance team.
Step 5 — Extend continuous monitoring to NHI access paths. Your SIEM and identity analytics tools likely have partial NHI coverage. The gap is usually in SaaS-native tokens and AI agent activity. Map the blind spots and close them before the next agent deployment.
The NHI estate you have today will be significantly smaller than the one you have eighteen months from now. Every agentic AI deployment adds to the machine identity count. The governance infrastructure you build now needs to scale with the ratio — which means the sequencing and the lifecycle processes matter more than any single tool purchase.
This post is part of TechVision Research’s June 2026 series on agentic AI security and identity governance. Next: CISO Accountability, Board Reporting & ITDR: The Identity-First Defense →
Missed the earlier posts? Start with Zero Trust for AI Agents: A 2026 Blueprint → or Closing the AI Proof Gap: A Shadow AI Governance Blueprint for 2026 →
For more guidance, catch the replay of Kevin Kampman’s webinar, “Instituting AI Governance: Guiding Risk in the Artificial Identity Era”
Recent Comments