For every human login you secure, dozens of machine identities operate with credentials nobody owns, rotates, or revokes. Agentic AI just turned that imbalance into a crisis.
Identity security has spent a decade getting good at one thing: protecting humans. Multi-factor authentication, single sign-on, phishing training, conditional access — an entire discipline built around the assumption that the risky identity is a person who might get fooled or breached. That assumption was reasonable when people were most of the identities on the network. They no longer are.
The numbers are lopsided in a way that should reframe how security budgets get spent. Non-human identities — service accounts, API keys, tokens, bots, and now AI agents — outnumber human users by an average of 45 to 1, and in cloud-native environments the ratio reaches 144 to 1, up from 92 to 1 a year and a half earlier. KPMG's 2026 report puts the figure at 80 to 1 and names non-human identity a critical CISO priority. Whatever the exact multiple, the conclusion is the same: the population you've optimized your defenses around is now a rounding error in your own identity estate.
And the part of the estate you've largely ignored is the part that doesn't get phished, doesn't rotate its password, and doesn't leave when an employee resigns. It just keeps holding a credential, often a powerful one, indefinitely.
Why machine identities are the soft target
A human identity has a lifecycle the organization understands. Someone joins, gets provisioned, changes roles, leaves, gets deprovisioned. There's an HR event behind every change and, usually, a process attached to it. Machine identities have none of that discipline. They get created by a developer to make two systems talk, granted broad permissions because narrowing them was tedious, and then forgotten. The governance applied to them rarely matches the rigor applied to people — only about one in five organizations has a formal process for offboarding and revoking API keys.
The result is sprawl, and the sprawl is measurable. GitGuardian's 2026 research found 28.65 million hardcoded secrets pushed to public GitHub repositories in a single year — a 34% jump and the largest on record. Each one is an orphaned, often over-privileged key sitting in code, waiting for someone to find it. These aren't exotic attack techniques. They're credentials left on the doorstep, at a scale no human-focused identity program was designed to see.
A machine identity doesn't get phished, doesn't reset its password, and doesn't quit. It just holds the key you gave it years ago — and you have dozens of them for every employee you so carefully protect.
Agentic AI turned a problem into a crisis
The reason this moved from "known hygiene issue" to "top CISO concern" in 2026 is the arrival of AI agents, which break every comfortable assumption about how machine identities behave. A traditional service account does a fixed, predictable thing. An AI agent is an autonomous actor: it acquires permissions dynamically at runtime, spawns sub-agents, calls external APIs, writes and executes code, and chains actions across dozens of systems — each step potentially using or creating a credential no human explicitly approved.
So the fastest-growing category of identity on your network is also the least predictable, the most heavily privileged, and the worst governed. The 144-to-1 ratio isn't just large; it's accelerating, and agents are the accelerant. Every agent you deploy doesn't add one identity — it adds an identity that can mint more. Govern that population with the tools you built for human logins and the math simply doesn't hold.
Visual 1 — Two identity populations, two different problems
Dimension | Human identities | Non-human identities |
|---|---|---|
Relative count | 1 | 45 to 144 |
Lifecycle | HR-driven, well understood | Created ad hoc, rarely retired |
Primary defense | MFA, SSO, training | Often none of the above applies |
How it fails | Phishing, credential theft | Orphaned keys, secrets in code, over-privilege |
Offboarding | Standard process | ~20% of orgs have one |
How to read it: nearly every defense in the middle column was built for the left and doesn't transfer to the right. The larger population is the one your program barely covers.
The uncomfortable reframe
There's a contrarian implication here that security teams resist, because it sounds like undoing a decade of progress. The human-identity work wasn't wrong — it was necessary, and phishing is still a live threat. But the marginal security dollar has almost certainly been going to the wrong population. Adding another layer of friction to human logins while machine credentials proliferate ungoverned is optimizing the small, well-defended surface and ignoring the large, open one. It feels productive because human identity is where the mature tools and clear metrics live. The exposure has moved to where the tools aren't.
What this means for leaders
Get a count before you get a strategy. Most organizations cannot say how many non-human identities they have, who owns each, or what it can access. That inventory — including secrets buried in code and the agents spawning new credentials — is the unglamorous first deliverable, and it's usually alarming.
Give machine identities a lifecycle. The core fix isn't exotic: ownership, least privilege, rotation, and a real offboarding process for credentials, applied with the same rigor you give people. A key that outlives its purpose is the single most common way this estate gets breached.
Govern agents at creation, not after. Because agents mint and inherit permissions at runtime, the control has to sit at the point an identity is issued and scoped — not in a quarterly review that's already months behind the sprawl. Treat every new agent as a new privileged identity, because that is exactly what it is.
The perimeter conversation has been settled for a while: identity is the new perimeter. The unfinished part is that most of your identities aren't people, and the defenses you're proud of were built for the minority. The attacker has already noticed which door is unlocked. The only question is whether you count your machines before they do.
A BusinessInfomatics original. Drawn from the KPMG 2026 cybersecurity report, Entro Security and GitGuardian 2026 data on NHI ratios and secrets sprawl, and Cloud Security Alliance research on non-human identity and agentic-AI governance.
