diff --git a/.planning/phases/12-initial-setup-wizard/12-SECURITY.md b/.planning/phases/12-initial-setup-wizard/12-SECURITY.md
new file mode 100644
index 0000000..93b7e33
--- /dev/null
+++ b/.planning/phases/12-initial-setup-wizard/12-SECURITY.md
@@ -0,0 +1,139 @@
+---
+phase: 12-initial-setup-wizard
+audited: 2026-06-15
+status: secured
+asvs_level: 2
+block_on: high
+register_authored_at_plan_time: true
+threats_total: 15
+threats_closed: 15
+threats_open: 0
+threats_mitigate_verified: 12
+threats_accepted: 3
+supply_chain_checks: 1
+---
+
+# Phase 12 — Initial Setup Wizard: Security Audit
+
+**Audited:** 2026-06-15
+**ASVS Level:** 2
+**block_on:** high
+**Compared against:** main...HEAD
+**Status:** SECURED — 15/15 threats resolved (12 mitigate verified, 3 accept documented)
+
+This audit verifies each declared threat mitigation EXISTS in the implemented code. It does
+not scan for new vulnerabilities beyond the register. Implementation files were not modified.
+Note: T-12-07 and T-12-10 are tracked as one accepted-risk entry per the register grouping,
+so the 15 register rows map to 14 IDs.
+
+## Threat Verification
+
+| Threat ID | Category | Disposition | Status | Evidence |
+|-----------|----------|-------------|--------|----------|
+| T-12-01 | Information Disclosure | mitigate | CLOSED | `scripts/generate-secrets.mjs` emits only via `console.log` (`:31-40`). No `writeFile`/`appendFile`/`fetch`/db import anywhere in file; sole imports are `web-push` and `node:crypto.randomBytes` (`:23-25`). Secrets never persisted. |
+| T-12-02 | Tampering | mitigate | CLOSED | `0002_lethal_millenium_guard.sql` uses `MODIFY COLUMN` (`:1-2`), never DROP; adds `claimed` (`:3`); backfills `UPDATE users SET claimed = true WHERE oidc_iss IS NOT NULL` (`:6`). Migration shipped via drizzle-kit generate (`meta/0002_snapshot.json`, `_journal.json` present), not push. |
+| T-12-03 | Information Disclosure | mitigate | CLOSED | `schema.ts appConfig` (`:306-310`) has no column/key for `vapid_private_key` or `app_password_encryption_key`; explicit PROHIBITION comment (`:300-303`). Grep confirms no app_config insert of secret material across `apps/api/src`. |
+| T-12-04 | Tampering | mitigate | CLOSED | `isSetupLocked()` is the FIRST statement returning 423 in every mutating handler: `/config` (`setup.ts:101-102`), `/validate/db` (`:137-138`), `/validate/oidc` (`:159-160`), `/validate/vapid` (`:199-200`), `/credential` (`:244-245`), `/complete` (`:318-319`). Guard re-reads DB per call, no module-level cache (`setupGuard.ts:26-39`). |
+| T-12-05 | Information Disclosure | mitigate | CLOSED | `noEchoHook` returns `{ error: 'Invalid request' }` only, no Zod detail (`setup.ts:49-53`), wired on `/credential` (`:243`). CredentialValidationError → generic 400 (`:294-296`). No `console.*` of `appPassword`/`c.req.valid` (`:248` explicit no-log comment; only `err.message`/string logged at `:299-302`). Helper `credentialSync.ts:63-67` never logs password. |
+| T-12-06 | Information Disclosure | mitigate | CLOSED | `/validate/vapid` reads both keys ONLY from `process.env` (`setup.ts:202-203`); returns only `{ ok: true }` (`:218`) or generic message (`:220-224`); private key never in any response. No app_config key for VAPID private key (T-12-03 evidence). |
+| T-12-07 / T-12-10 | Spoofing | accept | CLOSED | Accepted risk logged below. Claim query is strictly `isNull(users.oidcIss) AND claimed=false LIMIT 1` (`user.ts:112-116`) — no email match. OIDC reach requires Authelia membership (D-08). Two-person household → one pending row. See residual-risk note RR-1. |
+| T-12-08 | Tampering (SSRF) | mitigate | CLOSED | `configSchema.oidcIssuer` refined `startsWith('https://')` (`setup.ts:63`). `/validate/oidc` discovery fetch uses `AbortSignal.timeout(5000)` (`:175-176`). See finding F-1 (error-message leakage, IN-03) — bounded, non-blocking. |
+| T-12-09 | Tampering | mitigate | CLOSED | `index.ts` mounts `app.route('/api/setup', setupRouter)` (`:49`) BEFORE `app.use('/api/*', devAuthBypass())` (`:54`), `oidcConfigFallbackMiddleware`/`oidcAuthMiddleware()` (`:69-70`). Setup surface never reaches the OIDC 302 guard. |
+| T-12-11 | Elevation of Privilege | mitigate | CLOSED | `shouldBeAdmin = flagRow?.value !== 'true' && Number(count) === 0` (`user.ts:156`). Post-setup logins (setup_complete='true') cannot self-promote. Claim path preserves wizard-set `is_admin`, does not overwrite (`:121-130`, comment `:120`). |
+| T-12-12 | Tampering | mitigate | CLOSED | No `claims.email`/`users.email` in claim path; match is identity-null + claimed-false only (`user.ts:112-116`). Identity upsert keys on `(oidcIss, oidcSub)` (`:78-82`), never email. `deriveDisplayName` uses email only as a display hint, never for identity (`:56-64`). |
+| T-12-13 | Information Disclosure | mitigate | CLOSED | Wizard has no generate-secrets step; Step 2 collects only non-secret VAPID **public** key (`SetupPage.tsx:614-631`, field `vapidPublicKey`). No SESSION_SECRET / encryption key / VAPID private key referenced in `SetupPage.tsx` or `client.ts` payloads (`SetupConfigPayload` is public-only, `client.ts:550-555`). |
+| T-12-14 | Tampering (XSS) | mitigate | CLOSED | No `dangerouslySetInnerHTML` in `SetupPage.tsx` (grep across `apps/pwa/src` shows zero usages — only doc comments elsewhere). All copy and config values rendered as plain-text JSX children. |
+| T-12-15 | Information Disclosure | mitigate | CLOSED | App password input is `type="password"` (`SetupPage.tsx:836`), `autoComplete="new-password"` (`:837`). Server-side `noEchoHook` (T-12-05). Password held only in transient form state, never persisted client-side. |
+| T-12-SC | Tampering (supply chain) | accept | CLOSED | `git diff main...HEAD` of `package.json` adds ONLY a script entry (`generate-secrets`), no `dependencies`/`devDependencies` change; no `apps/**/package.json` or lockfile change in branch diff. web-push/node:crypto/lucide-react/react-query/react-router already present. Accepted risk logged below. |
+
+## Lead Assessment (from 12-REVIEW.md known overlap)
+
+The auditing prompt flagged four prior-review items as leads bearing on the register. Verdicts:
+
+- **WR-04 (appExternalUrl no https refine) → does NOT defeat T-12-08.** T-12-08's scope is the
+  **OIDC issuer** SSRF surface, and `oidcIssuer` IS https-refined (`setup.ts:63`). The SSRF
+  fetch target (`/validate/oidc`) uses `oidc_issuer` only, never `app_external_url`. So the
+  declared T-12-08 mitigation is intact. Separately, `appExternalUrl` is `.url().max(512)` with
+  no scheme refine (`setup.ts:66`); it is consumed only as `OIDC_AUTH_EXTERNAL_URL` to build
+  the redirect base. An `http://` value is an operator-self-inflicted misconfig (OIDC login
+  fails closed at Authelia), not an attacker-controlled open-redirect — the value is operator-
+  supplied during a one-time, lock-gated setup, not a per-request user input. Tracked as
+  hardening F-2, not an open threat.
+
+- **IN-03 (validate/oidc echoes raw network error) → bounded info-disclosure, does NOT defeat
+  T-12-08.** The 5s timeout is present and verified (`setup.ts:175-176`). The error string at
+  `:182-185` can surface internal `ECONNREFUSED <ip>:<port>` to the pre-auth client. T-12-08's
+  declared mitigation (https validation + server-side fetch + 5s timeout) is fully present; the
+  leak is a residual disclosure the register did not call out as in-scope. For a self-hosted,
+  single-operator, lock-gated setup endpoint the exposure window/audience is the operator
+  themselves. Tracked as hardening F-1 (recommended, non-blocking at ASVS L2 for this context).
+
+- **WR-02 (TOCTOU: two concurrent /credential POSTs → two unclaimed admin rows) → weakens the
+  robustness of the "exactly one pending user" assumption but does NOT defeat the accepted
+  T-12-07/T-12-10 spoofing property.** Both racing inserts are performed BY THE OPERATOR during
+  their own setup window; both rows are `isAdmin=true` and represent the operator's own intent.
+  The claim binds the first OIDC login (which still must be an Authelia-authorized member, D-08)
+  to one of two operator-owned rows — it does not let an external/wrong principal claim an
+  identity. The second row becomes an orphaned admin (a correctness/cleanup defect, also WR-01),
+  not a privilege-escalation or spoofing vector. Recorded as residual risk RR-1. Recommend the
+  WR-02 check-before-insert (or DB transaction) fix to restore the single-pending-row invariant.
+
+- **WR-05 (oidcConfigFallbackMiddleware permanently mutates process.env) → no security impact
+  on this register.** The injected values are the non-secret OIDC issuer / client_id / external
+  URL (D-01), the same values that would otherwise be set as env. No secret is written to
+  `process.env` by this path (`middleware.ts:85-87`). The "stale after wizard re-run" behavior
+  is an operability concern, not a confidentiality/integrity threat. No register threat depends
+  on re-reading these post-first-request.
+
+## Accepted Risks Log
+
+- **T-12-07 / T-12-10 — first-login-claims binds the wrong principal (Spoofing).** Accepted
+  per D-08. The claim query matches purely on `oidc_iss IS NULL AND claimed=false` and binds
+  the first OIDC login to the single pending wizard row. Soundness rests on: (a) OIDC reach is
+  gated by Authelia membership (only household members can authenticate at all), and (b) a
+  two-person household has exactly one pending unclaimed row at first login. No email coupling
+  (T-12-12) means a leaked/guessed email cannot influence the binding. Accepted as sound for
+  the two-person, Authelia-fronted deployment. Residual robustness caveat: see RR-1.
+
+- **T-12-SC — supply-chain / dependency install.** Accepted: zero runtime/dev dependencies
+  added this phase (verified by `git diff main...HEAD -- package.json`: only a `scripts`
+  entry added). No new attack surface from third-party packages.
+
+- **(implicit) generate-secrets operator handling.** Per SC-3, secrets are printed to stdout
+  and the operator is responsible for safe handling (paste into docker-compose env). The script
+  itself never persists them (T-12-01). Accepted: operator-custody model is the documented
+  trust boundary.
+
+## Residual Risks (non-blocking, recommended hardening)
+
+- **RR-1 (WR-02):** Concurrent `/credential` POSTs can create a second orphaned unclaimed admin
+  row, weakening the "exactly one pending user" invariant behind the accepted T-12-07/10 risk.
+  Not exploitable for cross-principal spoofing/escalation (both rows are operator-owned), but
+  recommend the check-before-insert / transaction fix from 12-REVIEW WR-02. Combine with WR-01
+  (roll back insert when post-insert re-select returns nothing) to fully close the orphan path.
+- **F-1 (IN-03):** `/validate/oidc` returns raw network error detail (possible internal IP/port)
+  to the pre-auth client. Recommend returning a generic message and logging detail server-side.
+- **F-2 (WR-04):** Add `.refine(startsWith('https://'))` to `appExternalUrl` to fail fast on a
+  mistyped `http://` redirect base. Operability hardening; not attacker-controlled.
+
+## Unregistered Flags
+
+None. No `## Threat Flags` section exists in any 12-*-SUMMARY.md; no new attack surface appeared
+during implementation that lacks a register mapping.
+
+## Files Audited
+
+- apps/api/src/routes/setup.ts
+- apps/api/src/lib/setupGuard.ts
+- apps/api/src/index.ts
+- apps/api/src/auth/user.ts
+- apps/api/src/auth/middleware.ts
+- apps/api/src/broker/credentialSync.ts
+- apps/api/src/db/schema.ts
+- apps/api/src/db/migrations/0002_lethal_millenium_guard.sql
+- scripts/generate-secrets.mjs
+- apps/pwa/src/routes/SetupPage.tsx
+- apps/pwa/src/api/client.ts
+- package.json (supply-chain diff)
+
+_Audited: 2026-06-15 — gsd-security-auditor. Implementation files unchanged (read-only)._