Every service behind the lab’s single sign-on gate redirected an unauthenticated visitor to a plain-HTTP address on a raw internal IP and port. The configuration that was supposed to prevent that was set correctly and had been recorded as fixed. It changed nothing, and it took two rounds to find out why.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →
A fix that was recorded and did not work
The first fix, in August, set authentik_host_browser to the public HTTPS name of the identity server. The blueprint recording it was marked fixed. In September a verification of a different gate found the old behavior unchanged: the first top-level navigation of every login still went to plain HTTP on an internal address. It was not caused by the new work, and a control proved it. A service gated since August emitted the identical redirect.
The agent then ruled out the usual explanations one at a time. The setting was in the database with the right value. Restarting the server container changed nothing, so it was not a stale cache. No environment variable overrode it. The setting was applied, correct and ignored.
Why it was left alone, and what that choice was
The only remaining lever was a different setting, the outpost’s general host address, and changing it was not a one-line edit. From the identity host itself the public name was refused by the reverse proxy’s allowlist, so repointing it also meant widening that allowlist and routing the outpost’s own traffic back through the proxy. The identity server gates most of the fleet’s services. A mistake would lock the operator out of nearly all of them.
Leave it filed. The redirect works, and the plaintext hop stays on an internal network toward a port already restricted to the operator’s own device subnets. That is a real weakness but not an active compromise, and the record could not say how much a lockout was worth against it. The human judged it not worth risking sign-on for the whole fleet.
What the source said
Two weeks later the agent read the identity server’s source inside the running container rather than its documentation. authentik_host_browser is read only by the controllers that launch managed outposts, which export it to the outposts they start. The embedded outpost, the one this deployment uses, is not started by that code. It never receives the value and builds its browser-facing URL from the general host address instead. The August fix had set a key the embedded outpost does not consult.
With that, the change became small and safe. The general host address was repointed to the public name, the full configuration was re-read first, and the blueprint was applied without a restart because the outpost picked it up live. Back-channel calls such as token and key fetches stay on the local address, so they were unaffected.
How the fix was checked
The agent probed every forward-auth application exactly as the reverse proxy does, at the outpost directly. Before the change each answered with a redirect to the plain-HTTP internal address. After, each answered with a redirect to the public HTTPS name, with the same client identifiers. The outpost’s provider list read back identical, and OIDC discovery for the services that use it was unchanged. The server log showed no errors in the following ten minutes.
What generalizes
A field that exists in the schema and accepts a value is not evidence that anything reads it. The August record said fixed because the value was set, and the September probe said not fixed because the behavior was unchanged, and only the second one measured the thing that mattered. The fix that finally worked came from reading the code path that builds the URL, not from setting another value and watching for a different result.