The Uncounted Workforce: Machine Identities and the Governance Gap No One Is Closing

Ask most CISOs how many employees they have, and they can answer within a rounding error. Ask them how many identities their organization actually manages — every service account, every API key, every OAuth token, every AI agent operating with its own credentials — and the honest answer, more often than not, is that no one has ever counted. That is not a gap in reporting. It is a gap in governance, and it has gotten dramatically larger in the past eighteen months for one specific reason: AI agents create new machine identities the way human hiring never could, at a pace no existing process was built to track.

This is not a marginal, back-office problem. Palo Alto Networks' 2026 Identity Security Landscape report, drawn from nearly three thousand security leaders worldwide, puts the average ratio of machine identities to human identities at 109 to 1 — up from 82 to 1 the year before, a jump of nearly a third in twelve months. In cloud-native environments, independent research places that ratio as high as 144 to 1. One Fortune 500 financial institution, audited as part of a broader industry study, was found to have more than 4.2 million non-human identities against roughly 50,000 human accounts. The workforce doing the most consequential logging in, calling out, and moving of data across your enterprise is not your workforce at all.

Four forces compound on top of each other. Cloud adoption assigns an identity to every resource that needs to reach another service. Microservices architecture multiplies that further, since each service typically authenticates independently rather than inheriting a single credential. SaaS integration adds a token for every connected tool, on top of whatever OAuth sprawl already exists. And now agentic AI adds a category that did not meaningfully exist three years ago: autonomous systems that provision their own credentials, call APIs on their own schedule, and — increasingly — spin up sub-agents that need credentials of their own. Of the 109 machine identities the average enterprise now manages per human employee, roughly 79 are AI agents, and that segment alone is projected to grow another 85 percent over the next twelve months, well ahead of the growth rate for either human or traditional machine identities.

None of this is happening because organizations are being reckless. It is happening because the tooling that provisions new identities has gotten radically faster than the governance layer that is supposed to track them. A developer can spin up a new service account in seconds. An AI platform can generate a new agent credential automatically, as a side effect of a workflow someone approved for an entirely different reason. There is no HR event, no onboarding form, no manager sign-off — the traditional triggers that cause a human identity to enter a governance system simply do not exist for most machine identities. They appear because a system needed them to, and in the absence of a countervailing process, they stay forever.

Single sign-on, multi-factor authentication, privileged access management, and the quarterly access review — the four pillars most identity programs are built on — all share a quiet assumption: that an identity belongs to a person who can be asked to log in again, prove who they are, and answer for what they did. Machine identities break that assumption at every point. A service account cannot complete an MFA challenge. An API key does not have a manager to certify its access is still needed. A quarterly access review conducted by a human reviewer, working through a spreadsheet, does not scale to an inventory that may run into the millions and that changes hourly.

The result is a compliance grey zone that regulators are only beginning to close. Frameworks including SOC 2, ISO 27001, PCI DSS, and NIST 800-53 all carry access governance requirements that apply, in principle, to non-human identities exactly as they apply to human ones. In practice, most audit processes were built around human user reviews, and machine identities have simply never been asked the same questions. That grey zone is narrowing. Auditors are starting to ask pointed questions about machine credential lifecycle management, and "we store the keys in a vault" is no longer a sufficient answer when the vault has never been inventoried, let alone reviewed.

This is not a theoretical exposure. Roughly 90 percent of organizations reported at least one identity-related security incident in the past year, and research tracking those incidents through the first half of 2026 attributes about 70 percent of them to activity tied to autonomous AI systems — agents acting with excessive privilege, credentials that outlived their purpose, or automated workflows that no human was actively supervising when something went wrong. Separately, roughly two out of every three ransomware intrusions now trace back to a compromised identity credential rather than a traditional malware delivery vector, and machine credentials — over-privileged, rarely rotated, frequently hardcoded — are a disproportionate share of that entry point.

The credentials themselves are also leaking at a pace that has outstripped most detection programs. Security researchers tracking public code repositories found that more than 28 million hardcoded secrets — API keys, tokens, and service account credentials pasted directly into source code — were added to public GitHub repositories in 2025 alone, a 34 percent increase over the year before and the largest single-year jump on record. Every one of those secrets represents a machine identity that a developer, often without malicious intent, embedded somewhere it should never have lived, and that in many cases is still valid today.

Identity Type	Where Governance Typically Fails
Service accounts	Created for a project, rarely deprovisioned when the project ends; often over-privileged from day one
API keys & tokens	Frequently hardcoded in source or config files; long or infinite expiry is the default, not the exception
OAuth grants	Approved by individual users with no admin review; persist indefinitely after the approving employee leaves
RPA bots	Built to mimic human logins with standing credentials, often outside IT's own asset inventory
AI agent credentials	Provisioned automatically by the platform; no owner, no approval workflow, and often no expiry by design

Non-human identity governance is not a product category you purchase your way out of, though tooling has matured considerably and can help. It is a discipline that starts with visibility and ends with ownership. The following five actions represent the floor for organizations that have not yet treated machine identity as a defined governance category — not the ceiling.

• Build One Inventory, Not Five. Cloud IAM consoles, secrets managers, CI/CD pipelines, and AI platforms each maintain their own partial list of machine identities. Almost no organization has consolidated them into a single, authoritative inventory. Until that inventory exists, every other control is operating on incomplete information.
• Assign an Owner to Every Credential. A machine identity with no named human or team accountable for it is a candidate for immediate review, not a low-priority cleanup item. Ownerless credentials are, definitionally, credentials no one would notice going rogue.
• Default to Expiry, Not Permanence. Static, long-lived API keys should be the exception that requires justification, not the default configuration. Short-lived, automatically rotated credentials close the largest share of this exposure without requiring a single new tool.
• Extend Least-Privilege to Machines. Service accounts and AI agents are routinely granted broader scopes than they need, because the broader scope is easier to configure once and forget. Treat machine credentials with the same privileged-access discipline applied to a human admin account.
• Stand Up an AI Agent Registry Before Scale Forces the Issue. Every agent deployed should be registered with a defined purpose, a defined data scope, a defined action scope, and a named business owner — before it is given credentials, not after an incident forces a retroactive audit.

The uncounted workforce is not a future risk. It is already the majority population inside every enterprise network, it is growing faster than the human workforce and faster than the tools built to govern it, and in most organizations, no single person could currently produce an accurate count of it if a board member asked. That is the finding worth sitting with: not that machine identities are risky in the abstract, but that most security leaders cannot yet answer a question about their own environment that regulators, auditors, and attackers are all now asking with increasing specificity.

For executive leadership, the next security briefing is the right place to ask it directly: how many non-human identities exist across our environment, who owns each category, and what percentage currently have no expiry and no assigned owner? If the honest answer is "we're building that inventory now," that is a defensible starting point. If the honest answer is silence, the exposure was never hypothetical — it has simply never been measured.

---