Singapore has spent a decade building one of the most connected public sectors in the world. Singpass gives residents a single, trusted digital identity; whole-of-government platforms quietly move data between agencies; citizen services increasingly run on APIs rather than counters. That success has produced an uncomfortable side effect. The number of machine identities operating inside the Singapore government estate now dwarfs the number of human ones, and most of the governance machinery was designed with the humans in mind.

The problem is not that machines are unmanaged. It is that they are managed on a different ledger. Human access is requested, approved, reviewed and revoked through processes that assume a person, a role and a manager. Service accounts, workload identities, API keys, certificates and the credentials that AI agents inherit sit outside those workflows almost by architecture, because no one ever thought to put them inside. When an audit asks how many identities have access to a sensitive dataset, the honest answer is often a human count with an asterisk.

Scale breaks the mental model

Every digital service multiplies identities. A single application might carry a runtime service account, a set of CI/CD credentials, a signing key, a database role and several short-lived tokens, and then spin up more for each environment. In a government context, the multiplication is compounded by inter-agency data flows, where a machine identity on one side may legitimately need access to a system on another. Counting people was always tractable; counting machines requires telemetry, and few agencies have it at estate scale.

Ownership is the quiet failure point

Machine credentials outlive the people who created them. When an engineer leaves, their service account remains, still authenticating, still privileged, and now owned by nobody. Decommissioned projects are a reliable source of orphaned identities, because deleting an account feels riskier than leaving it. Any serious non-human identity programme starts with an owner field that is enforced, not advisory, and with a scheduled review that treats an unowned credential as a finding rather than a formality.

Rotation without discovery is theatre

Agencies that cannot see their full machine identity inventory cannot rotate safely, so rotation quietly gets deferred. The practical sequence is discovery first, then classification by privilege and data sensitivity, then rotation on a schedule the team can actually sustain. Secrets stored in code, in configuration files or in shared vaults with broad access undermine the whole exercise, which is why scanning source repositories and workload configurations belongs in the same programme as entitlement review.

Agents raise the stakes

The next wave of automation makes this harder in a specific way. AI agents act on behalf of a user or a service, often by borrowing existing credentials rather than presenting their own. That collapses the distinction between acting and impersonating, and it makes attribution in logs unreliable. An agent that runs with a human’s token produces an audit trail that cannot tell you whether the person or the process did the thing. Short-lived, scoped, purpose-built credentials for agents, plus a record of which agent acted under which delegation, are becoming a basic requirement rather than an advanced control.

What a national-scale approach looks like

The agencies making progress tend to share a pattern: a central inventory that is treated as authoritative, mandatory ownership metadata, automated expiry for credentials that are not renewed, and policy-as-code so that the same rules apply across environments. They also invest in making the secure path the easy path, because a developer who has to file a ticket to get a scoped credential will hardcode a key instead. For Singapore, whose digital government ambitions depend on trust in the underlying plumbing, closing the machine identity gap is less a security project than a prerequisite for the next phase of the plan.