# Open security findings (M1 auth review)

An adversarial review of the M1 auth, session and authorization code produced 28
confirmed findings (~20 after deduplication). The ones **fixed** are listed first
so nobody re-reports them; the rest are open and are the M1 exit criteria.

## Fixed

| # | Issue | Fix |
|---|---|---|
| 1 | `pro.upsertProfile` computed `requiresReReview` and never applied it — a verified plumber could silently become a verified electrician 40 km away | Demotes a `verified` profile to `pending` in the same transaction, writes an audit row, and now counts a **trade change** as material (it did not before). Needed a new `verified -> pending` edge in `VERIFICATION_GRAPH` |
| 2 | `phoneNumber()` registers `/phone-number/request-password-reset`, `/phone-number/reset-password` and `/sign-in/phone-number` **unconditionally** — not gated on `emailAndPassword.enabled: false`. Together they mint a password credential with no SMS sent to the owner, then accept it forever with no OTP | All three served as 404 via `disabledPaths` (checked in `onRequest`, before rate limiting) |
| 3 | `bearer()` accepted the raw plaintext `sessions.token` column as an `Authorization: Bearer` credential — one leaked DB row is a replayable login | Plugin removed. The mobile client it was for does not exist yet |
| 4 | Phone numbers stored exactly as typed, so `+34600111222` and `0034600111222` are two "unique" accounts — defeating the UNIQUE constraint, bans, and duplicate detection | `phoneNumberValidator` pins E.164 on send-otp and sign-in; `toE164`/`isE164` added to `@linkder/shared` |
| 5 | `NEXT_PUBLIC_APP_URL` fell back to `http://localhost:3000`, which drops `Secure` and the `__Secure-` prefix from the production session cookie | Throws at boot in production |

## Open — must close before M1 ships

**Authorization / attribution**

- `resolveSession` discards `session.impersonatedBy`, so every admin action taken
  while impersonating is written to `audit_log` as the **victim**. The `Session`
  type has no field for it, so no procedure can tell. Admin impersonation is
  enabled (`admin()` plugin) and is currently invisible to the application.
  *Fix:* add `impersonatedBy` to `Session`, populate it in `resolveSession`, and
  record it on all three `audit_log` insert sites.
- `user.setRole` is a check-then-write across two connections; a concurrent
  `job.create` permanently strands the job it was meant to protect.
- `deck.swipe` checks `job.status` outside its transaction — a concurrent
  `job.cancel` loses the race and a pending request lands on a cancelled job.
- `deck.undo`'s "already reached the pro" guard is check-then-write: it deletes
  the tombstone for a request that really was sent, and reports success.

**Storage / uploads**

- Government ID and insurance scans go to the **public** bucket;
  `isPrivateKind` is never called from production code.
- Presigned PUT does not pin `Content-Type`, so a caller can store HTML under an
  image key in a public bucket.
- `pro.addMedia`, `pro.addCredential` and `job.create` accept **any** URL the
  caller sends, with no binding to an object that caller actually uploaded —
  which bypasses the presign design entirely.

**Abuse / spend**

- No per-number cap, destination allowlist, or global ceiling on
  `/phone-number/send-otp`. SMS spend is unmetered.
- better-auth rate limiting is per-process in-memory and collapses to one global
  bucket behind a proxy chain — 10 req/min then locks out the whole platform.
  Needs the Redis-backed `rateLimit.customStorage`, not `secondary-storage`
  (which throws unless all five `SecondaryStorage` methods exist).

**Lower severity**

- `user.setEmail` lets any authenticated user claim an arbitrary *unverified*
  address, locking the real owner out of Google sign-up and leaking which
  addresses are registered.
- Open redirect: the `next` query param on the sign-in form goes straight to
  `router.push`.
- `audit_log.ip` trusts the leftmost `X-Forwarded-For` entry verbatim —
  attacker-controlled.
- `findPossibleDuplicates` has no call site; the guarantee it was written to
  provide never executes.
- `DeckClient` renders a frozen `initialCards` prop and invalidates a query
  nothing subscribes to, capping the deck at 20 cards per page load.
- A right swipe after a left swipe creates a live request while the stored swipe
  row still records a rejection.

## Note on the OTP-in-plaintext decision

`apps/web/src/lib/auth.ts` argues the plaintext OTP is acceptable partly because
"anyone who can read that table can already read `session.token`". With
`bearer()` removed that argument is weaker than when it was written — a session
token now requires an httpOnly browser cookie to use, an OTP does not. The
decision still stands (300s window, 3 attempts, rate limited) but revisit it if
OTP lifetime or attempt budget ever increases.