Security programmes are still built around a question that no longer describes the estate: who logged in? The average enterprise now runs far more non-human identities than human ones, and those identities do not log in so much as exist, holding keys, certificates, tokens and service account passwords that let applications, pipelines, containers and now AI agents talk to each other. They are created by provisioning scripts, by cloud platforms, by CI/CD tooling and by developers who needed something to work before a deadline. Most of them are never reviewed again.
The result is a category of privileged access that sits outside the review cycles, joiner-mover-leaver processes and password policies designed for staff. Gartner’s commonly cited framing is that machine identities already outnumber human identities by a wide margin, and every credible survey since has shown the governance gap widening rather than closing.
Why the threat is structural rather than incidental
Three properties make non-human identity a distinct problem. First, proliferation is invisible. A new container, a new integration, a new agent each generates credentials, and very few organisations can enumerate what exists across cloud, on-premises and SaaS estates. Second, the credentials are high-value and durable. A service account with broad permissions and a five-year-old password is a stable, quiet path to production data, which is precisely what an attacker wants. Third, detection is weak. Machine-to-machine traffic is the baseline of enterprise networks, so anomalous use of a machine credential looks like normal operations unless the organisation has baselined what each identity should be doing.
Attack paths that start in the stack
The practical exploitation pattern is consistent. An exposed secret in a repository, a build log or a container image gives initial access. From there, the attacker pivots through the machine identity layer rather than the human one, because machine credentials are frequently over-privileged, shared and unmonitored. Software supply chain compromise, CI/CD pipeline abuse and lateral movement through service accounts all rely on the same weakness: the credential proves what it was issued to prove, and nothing about whether the action being taken is appropriate.
What governance looks like when it is applied to machines
Discovery is the entry point, and it has to be continuous rather than annual, because the population changes daily. Each identity needs an owner and a purpose, so that an unexpected one can be treated as an exception rather than a mystery. Secrets should move into a vault and be issued at runtime with short lifetimes, which reduces the value of any single exposure. Privilege should be scoped to the workload’s actual function, not inherited from the team that built it. Rotation and expiry need monitoring that escalates automatically, since the most common failure is a certificate or secret that nobody remembered to renew. And every machine identity should generate an audit record attributable to a specific workload, which is the only way to distinguish an agent doing its job from an agent being driven by someone else.
The agentic acceleration
AI agents sharpen every one of these problems. They need credentials to act, they operate at machine speed, they combine tools in ways their designers did not specify, and they are frequently deployed with borrowed human tokens because that is the fastest route to a working demo. The identity layer is where that behaviour can be constrained, registered, scoped and observed. Organisations that treat NHI management as a hygiene project will find it is now the control plane for autonomy, and the ones that treat it as an afterthought will discover their agent inventory the way most organisations discover their service accounts: during an incident.