More Agents, More Secrets, More Risk

Every AI agent an organization deploys is also a new non-human identity, whether anyone thinks of it that way or not. It authenticates, holds credentials, acts on systems, data, and other services. It takes these actions continuously, at machine speed, with no human standing nearby to notice when something does not adhere to policy.

Most security teams don't think of it that way yet. An agent feels like a feature, a script or a workflow bolted onto an existing product. But to a vault, an IAM system, or the infrastructure it touches, an agent is an identity like any other, except it never logs out, never gets prompted for a second factor, and rarely draws the same scrutiny a human account would.

That distinction is about to matter a lot more than it has in the past.

Agents Are NHIs, and NHIs Run on Secrets

A human logs in with a password and, ideally, MFA. An agent authenticates with an API key, a token, a certificate pulled from a vault at runtime, etc. There's no login screen to pause at, no prompt asking "was this really you?" The credential works, or it doesn't, and once it's issued, nothing is currently in the process that forces anyone to check back in on it.

That's a workable model when a handful of service accounts and workloads are involved. It's a different problem entirely once an organization is standing up hundreds or thousands of autonomous agents, each spinning up its own credentials, each depending on a secret somewhere to do its job.

More agents means more NHIs. More identities means more secrets. And more secrets, issued faster than most security teams can inventory them, means more risk sitting in places nobody is watching.

Vaults Solve a Key Problem

Vaults and secrets managers do exactly what they were built to do, store secrets securely and enforce access policy at the point of retrieval. That job matters, and is even more critical now.

But a vault's job ends the moment a secret is handed out. What happens next, whether it's rotated the way policy says, whether the identity retrieving it is actually what it claims to be, whether it's still even a valid credential six months later, was never really the vault's value proposition to solve. 

A handful of secrets can survive on spot-checks and yearly reviews. Thousands of agent-issued credentials, each rotating and expiring on its own schedule, can't. CyberArk research found that 68% of organizations report they lack identity security controls for AI. Read plainly, that means most companies are already past the point where manual oversight was ever going to hold, whether they've noticed the crack yet or not.

Three Places This Risk Actually Shows Up

Strip away the abstraction, and agent-driven secrets sprawl breaks into three concrete problems.

Secrets sprawl. A new agent usually means a new credential, and every credential can widen the attack surface. A secret reused across several agents or hosts, because issuing it separately felt like extra work, is one compromise away from being a master key with many doors behind it.

Governance drift. A rotation policy is just a document until someone checks whether it's actually happening and corrects it if it is not. Secrets that should rotate and don't, Secrets that rotate, but the workload using them never picks up the new one or even secrets that expire on paper and keep authenticating in production anyway. None of it shows up in a vault's own access logs, because from the vault's point of view, a valid-looking credential just did what valid-looking credentials do.

Accountability gaps. When a secret is retrieved, is that a human, an autonomous agent, or a human quietly riding on an agent's assumed role to get somewhere they shouldn't? A shared or hardcoded secret makes that question almost impossible to answer from the vault alone. And an agent that was never registered as an identity in the first place isn't just ungoverned. It's invisible to current tools.

Governance and Runtime Security Aren't Optional Anymore

Because of what they do, vaults or secret managers will always be necessary, but the recent evolution of the market stresses the need for something built on top of them. Observability that checks whether policy is actually followed in practice, not just configured, and runtime security that can attribute every action to the real identity behind it, in real time, whether that identity is human, agent, or something in between.

A vault that secures the secret, paired with an observability layer that governs the identity and agent behind every action taken with it, isn't two disciplines stacked side by side for good measure. It's enhancing the value of both products. Neither one produces what the other does alone, and together they cover the full path a secret and an agent actually take, from issuance, to retrieval, to whatever happens after.

The number of agents running inside most organizations is only going one direction. Whether the secrets behind them are actually governed is still, for almost everyone, an open question, and not a comfortable one.