Essay ยท August 2026

Disconnected Does Not Mean Ungoverned

Shared logins are only the easiest compromise to see. The deeper problem is what happens when identity governance is duplicated at the edge.

There is a screen at the site that nobody logs out of.

It was signed in at the start of the shift, by whoever got there first, and it will stay signed in until the shift ends. Six people will use it. The maintenance contractor will use it. Everyone knows. It is in the audit file, described in careful language, with a compensating control written underneath.

Nobody put it there because they are careless. They put it there because the alternative was that the work stopped.

The defense world has a name for the shape: denied, degraded, intermittent, limited connectivity. DDIL is useful language, but it can make the problem sound exotic. It is not. Aircraft in cruise. Ships. Mines. Substations. Farms in the middle of a field at harvest. Factory floors deliberately air-gapped from the corporate network. Construction sites six months before anyone runs fiber. Rural clinics. Disaster response, where the network is the first thing to go and identity matters most. Add the merely degraded, which is most of the world more often than headquarters wants to admit, and the place with assured connectivity starts to look like the special case.

These places are not waiting for coverage. Some are physically remote, some are moving, some are shielded by a mountain or a hull, and some are disconnected on purpose because a network path is the thing they are defending against. The link is not a gap in the rollout. It is a property of the work.

Identity is the part that breaks first, and it breaks worse than everything else, for a structural reason. Almost every other system can defer. A reading can be cached and shipped later. A ticket can queue. A file can sync tonight. Authentication cannot defer, because it gates the thing that happens now. It is the front door, and a door that has to phone another continent before it opens is not a door.

So people work around it. Here is the actual list, and anyone who has run these sites will recognize all of it.

You leave a session open, because signing in takes the better part of a minute when it works at all. You create a site account that the shift shares, because provisioning each person against a directory they cannot reach is not a plan. You cache credentials locally with a long expiry, and then quietly extend the expiry, because the alternative is a truck roll. You keep a break-glass account with standing privilege, and you keep it because the one time you needed it, it worked. You exempt the site from a second factor, because the second factor needs a signal the site does not have. You stand up a local directory that starts drifting from the central one almost immediately. You put in a satellite link and now the satellite link is the single point of failure for the entire site's ability to do anything. Or you write the passwords down, which at least is honest.

Every one of those is a rational local decision. Together they are the security posture of the site, and none of it appears in the architecture diagram at headquarters.

The shared login is only the easiest one to see. The deeper problem is that authority freezes at the edge. A person keeps whatever the cache last knew about them. A break-glass credential stops being exceptional. A second factor becomes a policy that applies everywhere except the place where the consequence is physical. Revocation becomes a message that may arrive after the action it was supposed to prevent. The site is no longer asking whether this person is allowed to do this now. It is asking whether the last reachable system knew enough to let work continue.

Then the audit arrives, and this is the part I find most interesting, because the audit does not actually disagree with any of it. The finding gets written. The risk gets accepted. A compensating control gets attached: additional logging, a supervisor sign-off sheet, a quarterly review, a camera. And everyone moves on, because the auditor understands the site cannot do the thing the policy requires.

A compensating control is a promise that you will be able to reconstruct what happened later, using something other than identity. It is a substitute for knowing who did it. Sometimes it is a good substitute. Mostly it is a sheet of paper next to a machine.

What that costs is not just a password. It is accountable authority. When six people share an account, the log records the account. When a stale local directory says yes, the log records a decision made by yesterday's view of the world. When a standing emergency credential is used, the log records the exception, not the person whose authority should have been checked. You know a valve was opened, a dose was administered, a door was used, a record was changed, and the honest answer to who was allowed to do it, who actually did it, and whether anyone could have stopped it in time is a shift roster and an inference. Nobody was hacked. The system worked exactly as designed. It just cannot tell you the things you needed it for.

Every workaround on that list is the symptom of a requirement nobody wrote down: identity has to work when the network does not, and it has to be as accountable then as it is on a good day. Duplicating the identity system does not solve that. It moves the authority decision from the center, where the controls are strongest, to the edge, where the pressure to keep working is strongest. Another local directory just makes drift the next compensating control.

This is why security, legal, audit, and compliance teams are right to be suspicious of connected accounts and infrastructure at the edge. A live path into an air-gapped plant is a new path to defend. A replicated directory is a new authority to govern. A synchronized account is another place for stale access to survive. The mistake is assuming the only choices are a live connection back to headquarters or a weaker copy of headquarters on site.

Keep the satellite. Use it for telemetry, updates, logs, revocation bundles, anything it is good at carrying. Just do not make it the thing that decides whether the person standing at the machine has authority. Connectivity should improve freshness. It should not be the root of governance.

The same authority can reach past the login screen to the device itself: a door, a tool, a vehicle, a machine, a reader inside an air gap. The device checks the signed credential locally and decides there, with no account service behind it and no live network call to bless the act. The air gap holds, and the decision is still governed.

So state the requirement properly. The person's identity has to arrive with the person, on something they carry, not be fetched from somewhere while they wait. Their authority has to arrive with it: scoped to the thing being done, signed by whoever is allowed to grant it, and checked at the place where the action happens. The site has to be able to verify that alone, holding nothing beforehand except the issuer's public key, and specifically not holding a copy of the workforce. Every action has to be attributable to a real person, which means the credential proves possession by that person rather than knowledge of a secret six people share. When someone leaves, changes role, or is removed, that has to reach the site and be enforced there, without a live connection, and without the site having to trust the courier. And the site has to be able to say honestly how current its information is, rather than quietly acting on a stale list.

None of that requires ripping out the identity estate you already have. Keep the directory. Keep the IdP. Keep the HR system, entitlement process, privileged-access workflow, and review cadence. They remain the governed source of authority. What changes is where the decision gets carried and checked: the central system issues a portable proof of what it already decided, and one governed identity travels with the person instead of being re-created at every site, with authority still scoped to the action.

The site then needs no local account database and no password cache to keep working. Emergency access can still exist, because in a safety-critical place it has to, but it stops having to be a standing shared account that six people know. The site checks the person, the scope, the issuer, the revocation version, and its own staleness at the moment of action. Better evidence than a shared session, and better governance than another directory drifting at the edge.

A credential that carries its own proof does all of that: issued centrally, revoked centrally, verified locally. We have it running. A government-issued digital ID is presented through the browser to a verifier with outbound network calls disabled. From that we mint per-surface credentials. A disconnected reader accepts them, then refuses them once a signed revocation bundle reaches it by hand.

The reason to build it is not that the workarounds are embarrassing. It is that they were always the correct response to the constraint. Change the constraint and you do not need them.

All essays