Harden the login flow: new session at sign-in, signed ID tokens, same-origin requests, headers

From an audit of the Authentik login, run against a fake OIDC provider:

- Session fixation: the session id wasn't changed at sign-in, so one planted beforehand
  (from another page on the domain) became signed in as the victim. It is regenerated now.
- Requests from another origin: SameSite=Lax doesn't stop another app on the same parent
  domain, and a body-less POST (stop a guest, rotate a token) needs no preflight. Every
  POST/PUT/PATCH/DELETE, and GET /auth/logout, is refused when Sec-Fetch-Site (or, for older
  browsers, Origin) says it came from elsewhere. The agent endpoint is exempt: it uses a token.
- ID token signature: openid-client checks a token's claims but, as OIDC allows for a token
  from the token endpoint, not its signature, so the login rested on HTTPS to Authentik
  alone. The signature is now verified against the provider's published keys (jose, now a
  declared dependency).
- SESSION_SECRET: production refuses to start without one (the fallback is in the source
  code) and warns when it is the .env.example placeholder or short.
- Response headers: frame-ancestors 'none', X-Frame-Options, nosniff, Referrer-Policy; the
  framework name is no longer sent.
- A failed sign-in shows a plain page and is audited (session / login_failed, with address
  and reason, no tokens); sign-in is limited to 20 attempts a minute per address so the
  audit log and session store can't be flooded.
- A failed OIDC discovery (provider down at the first sign-in) was remembered until restart;
  it is retried now.

Tests run the real app against the fake provider; switching each fix off in turn makes
the matching tests fail.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
bobbanandClaude Sonnet 5.5 committed 2026-10-06 23:35:34 +02:00
1 parent 73b86d88de
commit 2dac273b1d
12 files changed
+681 -18

No files matched your search

+12
View File
@@ -232,6 +232,18 @@ relatives), and connect to the address they checked rather than looking the name
checks and scans; a WHOIS server must be public. IDs in a request (a device, a node, a DNS record) are checked against what a real ID
looks like before they go into a call to a connected system. `npm test --workspace server` covers all of this.
**Signing in** — through Authentik (OpenID Connect) with PKCE and a random `state`. The ID token's signature is checked against Authentik's
published keys, not just its contents, so the login doesn't rest on the connection to Authentik alone. A new session id is issued at sign-in. A
request that changes something is refused when it comes from another origin: the session cookie is SameSite=Lax, which doesn't stop another app on
the same parent domain, so the server also checks where each such request came from. A failed sign-in shows a plain message and is written to the
audit log (as `session` / `login_failed`, with the address and a short reason), and sign-in is limited to 20 attempts a minute per address.
Every response carries `frame-ancestors 'none'`, `X-Frame-Options`, `nosniff` and a no-referrer-to-other-sites policy. In production the app
refuses to start without a `SESSION_SECRET`, and warns if it is still the placeholder from `.env.example`.
Things to know: the first account to sign in becomes admin, so sign in yourself first on a new install. A session lasts up to 7 days after last use,
and disabling someone in Authentik does not end it — delete them on the Users page (or revoke the session on the Sessions page) and it ends at once.
If Authentik uses a certificate from an internal CA, point Node at the CA with `NODE_EXTRA_CA_CERTS=/path/to/ca.pem` rather than turning verification
off with `NODE_TLS_REJECT_UNAUTHORIZED=0`.
**Privacy** — a page every signed-in user can open that says what this
installation stores (accounts, sign-in sessions with their IP and browser, the audit
and diagnostic logs, server reports, the secrets tracker, credentials), where data