Your AI Agents Have Credentials. Do They Have Identities?
Ask your security team a simple question: how many AI agents currently hold credentials to production systems, and who can revoke any one of them individually? In most enterprises the honest answer is that nobody knows, and nobody can. Agents were deployed the way scripts have always been deployed — with a shared service account, a token in an environment variable, and a plan to sort out governance later. Later has arrived. The agents now read customer records, move money, file tickets, and call each other, and the identity layer underneath them is a spreadsheet at best.
The Ratio Nobody Planned For
Human identity management assumes a knowable population. You onboard people, you offboard them, and the number changes slowly. Agent populations do not behave that way. In mature deployments agents outnumber human employees by a wide margin — industry estimates commonly put the ratio near eighty to one — because every workflow, every department, and every integration spawns its own. They are created in minutes, cloned casually, and almost never decommissioned. An identity model designed for a headcount that changes monthly is now governing a population that changes hourly.
What Your Agents Actually Inherit Today
Walk the credential path of a typical enterprise agent and you find something uncomfortable. It authenticates as a service account created years ago for an integration nobody remembers. That account has accumulated permissions through a decade of "just add this scope so it works." It is shared across several automations, so its logs cannot distinguish which agent did what. And because revoking it would break four other systems, nobody revokes it. The agent does not have an identity. It has a borrowed key with the label rubbed off.
Why Human IAM Does Not Simply Transfer
The instinct is to extend the existing identity provider to cover agents. It helps, but it does not finish the job, because agents differ from humans in ways the model does not anticipate. They act continuously rather than during working hours, so anomaly baselines built on human rhythms produce noise. They act on behalf of a user, which means every action has two identities that must both be recorded. They call other agents, creating delegation chains where authority can quietly widen. And they cannot be phoned and asked what they were doing.
What Agent Identity Actually Requires
Four properties separate real agent identity from a token in a vault. First, uniqueness: every agent has its own identity with its own lifecycle, so its actions are attributable to it alone. Second, scoping: permissions are granted per agent and per task, not inherited from a shared account. Third, rotation: credentials are issued short-lived and rotated automatically, because a static secret held by a long-running process is a standing invitation. Fourth, revocation: any single agent can be shut off in seconds without collateral damage to the others.
The Kill Switch Test
There is a fast diagnostic for whether your agent governance is real. Ask what happens if one agent starts behaving badly at two in the morning — looping, exfiltrating, or acting on a poisoned instruction. Can the on-call engineer disable that specific agent immediately, without a change ticket, and without taking down the other agents that share its credentials? If the answer requires a meeting, you do not have agent identity. You have a shared account with optimism layered on top. The same test applies org-wide: a single control that revokes every agent at once, instantly.
Why This Is Being Priced In Now
The market has already signaled where this is heading. Palo Alto Networks paid roughly twenty-five billion dollars for CyberArk to own machine identity for conventional infrastructure — before AI agents became the fastest-growing category of non-human actor in the enterprise. Security budgets follow acquisitions of that size. The organizations that treat agent identity as a first-class concern now will pass the audits that the organizations improvising with service accounts are about to fail. Start by inventorying every agent that holds a credential. Most teams are surprised twice: by the count, and by what those credentials can reach.
Ready to See Decision Intelligence in Action?
Stop deploying AI as a search bar. See how Vigil transforms enterprise decision-making with simulation, predictive surfacing, and role-specific intelligence.
Request a Demo