Security

The application is not a trust boundary

ObjectStoreViewer performs no authentication and no authorization, and it terminates no TLS. Its standalone port must be reachable only through an authenticated proxy and an operator-managed network boundary.

TRUSTED_USER_HEADER is display-only and never authorizes a request. Never add a deployment that assumes the port, or a forwarded identity header, is safe when directly reachable.

Required deployment shape

1. Expose only the proxy through an Ingress or external Service.
2. Let the proxy reach the internal objectstoreviewer Service on port 3000.
3. The proxy strips any client-supplied identity header and sets X-Forwarded-User only after authentication.
4. Label the proxy pod app.kubernetes.io/component: auth-proxy so the shipped NetworkPolicy admits it.
5. Do not grant users or other namespaces direct access to the viewer Service.

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: objectstoreviewer-auth-proxy-only
spec:
  podSelector:
    matchLabels:
      app.kubernetes.io/name: objectstoreviewer
  policyTypes: [Ingress]
  ingress:
    - from:
        - podSelector:
            matchLabels:
              app.kubernetes.io/component: auth-proxy
      ports:
        - protocol: TCP
          port: 3000

Read-only, three times over

Layer	Guarantee
Interface	domain code sees only list, bounded get/open, and stat/head — there is no write method to call
Build	hack/check-readonly.sh fails on mutation-shaped identifiers in cmd//internal/ and on forbidden actions in the shipped policy templates
Credential	the deployment credential must independently deny writes — the layer that survives a bug in the other two

Use the templates under deploy/policies and scope them to the exact destination prefix. See the policy reference.

Secrets never leave the process

Credentials, mounted secret values, provider authorization headers, signed URLs, Azure SAS parameters, and raw credential-bearing SDK errors must not appear in logs, traces, metrics, HTTP responses, HTML, links, tests, fixtures, or snapshots. Redaction is mandatory at every adapter boundary, and mounted secret contents are held in an opaque type whose String() is [REDACTED].

Consequences you will notice in practice:

• startup errors name a variable and a reason, never a value or a path;
• provider failures are normalized to a small category vocabulary (authentication, authorization, throttled, unavailable, timeout, not_found, safety_limit, …) with no upstream message; and
• request logs use stable route names, never raw URLs, queries, identity headers, or authorization headers.

The unit suite asserts this across configuration, all three provider adapters, the application, and the evidence API—including canary credentials that must never appear anywhere in output. Run it with make test.

Object keys are opaque

Keys are never mapped onto local filesystem paths, path-cleaning is never used as an authorization boundary, and no key or cursor may escape its configured store, destination prefix, and scope. In sidecar mode the scanner is forcibly confined to the single configured Barman server prefix even when the credential could read more.

No raw backup content

Bounded format metadata (Barman backup.info, history files, pgBackRest info/manifests) is parsed internally, but only allowlisted fields are rendered. Arbitrary custom metadata and tags are hidden by default. The Barman catalog retains only allowlisted facts — status/type, WAL anchors, timestamps, sizes, compression, encryption, and structural-artifact evidence — and discards unknown metadata fields and their values.

Raw object download is an opt-in Slice 10 feature that has not shipped; with ALLOW_DOWNLOAD=false no usable download route or link exists.

Runtime hardening

The HTTP server uses a 5-second read-header timeout, 15-second read timeout, 30-second write timeout, 60-second idle timeout, 16 KiB maximum headers, and a 15-second graceful-shutdown deadline. Sensitive responses are non-cacheable and carry restrictive CSP, referrer, framing, and content-type headers.

The image is a statically linked binary in a distroless non-root runtime. The container test runs it under an overridden UID/GID 12345 with a read-only root, only /tmp mounted, all capabilities dropped, and no-new-privileges set.

Sidecar channel

In pgconsole-sidecar mode there is no TCP listener at all. The channel is HTTP/1.1 over a pod-private Unix socket authenticated by a pod-local bearer token: the token authenticates the caller, and the socket path — in a volume mounted only into the two intended containers — authenticates the server boundary. The auth proxy must not carry either volume. See the sidecar profile.

Reporting a vulnerability

Do not open a public issue with reproduction details for a disclosure problem. Report it privately through GitHub Security Advisories, with the redacted evidence described in troubleshooting.

SECURITY.md records the supported versions, the response targets, and which findings are in scope — in particular that the absence of authentication, authorization, and TLS in the application is documented design, not a vulnerability.