Somewhere in your company there is a drawer nobody opens.
Picture it as a real drawer for a moment. Inside are employee badges. Each one opens doors, and some open every door. None of the faces on them belong to people. One badge belongs to a deploy script written by an engineer who left four years ago. One belongs to a monitoring integration for a tool you decommissioned. One is labeled simply “temp-fix — do not delete.”
Every human in your building got a photo, a background check, and a start date. Every human will get an exit interview and a deactivated badge on their last day. The badges in the drawer got none of that, and they will never have a last day, because no one is the owner of the drawer.
Security people call these non-human identities: service accounts, API keys, tokens, certificates — any principal that can authenticate and is not a person. The industry went counting again in 2026, and the largest published tally found more than a hundred of them for every human in the enterprise. Read that again slowly. For every colleague you can see, more than a hundred pieces of software carry a badge, and most companies cannot produce the list.
For twenty years this was a quiet problem. The badges belonged to software that did exactly one thing, forever. The deploy script deployed. The backup job backed up. A stale credential was a risk, but a passive one — it needed an attacker to find it before it became a story.
Then software got hands.
The new workforce is agents: models with tools and permission to act in a loop. An agent does not do exactly one thing forever. It reads a situation, decides, and acts — with whatever badge you gave it. And because agents are built in a hurry, they tend to be handed the badge from the drawer: the old service account that already had access to everything, because someone in 2021 got tired of permission errors.
Here is the pattern I have watched, assembled from incidents I have worked and teams I have sat with. Nobody decides to give an intelligent system the master key. The decision happens in fragments. A pilot needs read access. A demo goes well. A deadline arrives. Someone reuses a credential that was already there, already broad, already invisible. Six months later an agent that answers support tickets is holding a badge that can touch the billing database, and not one person in the company knows both halves of that sentence.
The security industry has a name for what limits the damage when things go wrong: blast radius — the worst thing an actor could do on its worst day. For a human employee, blast radius is bounded by process: approvals, reviews, a manager, a legal name. For the drawer, blast radius is bounded by nothing but the permissions themselves. An agent with a too-broad badge does not need to be hacked to hurt you. It needs to be wrong once, sincerely, at speed.
So here is a Friday-afternoon exercise that costs nothing and has changed the shape of more than one budget meeting.
Count the drawer.
Not fix, not clean — count. Pull every credential your systems can enumerate. For each one, ask three questions. How old is it? What can it reach? Who would notice if it acted? You will not finish, and that is part of the finding. Somewhere in the first hour someone will hold up the digital equivalent of a badge and say the sentence that starts every reckoning: “does anyone know what this is for?”
Write the count on a whiteboard: how many badges, how many with no known owner, the age of the oldest. Age matters more than people expect. A credential is a decision someone made about trust, and the older it is, the further the company has drifted from the person who made it. In more than one room I have watched, the count ended the argument: the oldest active credential was years older than the agent it served. The room goes quiet.
The fix is not mysterious, and it is not a product pitch. It is the discipline your company already applies to people, applied to the workforce that now outnumbers them: an owner for every badge, a scope that matches one job, an expiry date, an offboarding path. Least privilege is not a poster. It is payroll hygiene for software.
The teams that do this do not start with tooling. They start with the count, because the count converts an abstract risk into a number an executive can be embarrassed by. Budgets follow embarrassment.
None of this slows the agent wave, and it should not. Agents are coming to your production systems because they are useful there. The question is the same one asked of every new hire: what exactly are you allowed to touch, and who signed for it?
Your humans could answer. Open the drawer and ask the rest of the workforce.