Custom AI Router Assessment

Custom AI Router Assessment (Next.js-class self-hosted gateways)

Companion to security/ai-api-gateway-security (which covers New API/One API forks and is manually authored — do not expect to edit it; extend THIS skill instead when learning applies to custom-stack routers).

1. Fingerprint

Panels hide from HTML links — discover them by following redirect targets. Probe likely paths (/admin, /dashboard, /login, /maintainer) and read each 307's Location; that is how /maintainer/login surfaces even when no page links to it.

2. Route mining (Next.js variant)

curl -s https://TARGET/ | grep -oE '/_next/static/chunks/[a-zA-Z0-9_.-]+\.js'   # collect chunks
# download all, then per chunk:
grep -ohE '"(/api/[a-zA-Z0-9_/$ {}.-]+)"' chunk.js | sort -u                     # routes (~70 typical)
grep -n 'fetch("/api/' chunk.js                                                  # call sites → read ±300 chars for payload shapes

• Re-mine after fetching OTHER pages: the maintainer login page pulls extra chunks (the auth chunk holding POST /api/auth/login {"username","password"}).
• Template-literal routes appear as /api/oauth/${a}/authorize — normalize ${...} to {param} when listing.

3. Auth-gate probing discipline (non-destructive)

• Write-shaped endpoints: send syntactically valid but nonexistent IDs ({"orderId":"zz-nonexistent-probe-9x7q"}) — proves the gate (401 vs processed) without creating artifacts.
• Destructive endpoints (/api/version/shutdown, reset/purge): GET method-probe only, never POST in production. 401 vs 405 answers the gate question.
• Expect and record uniform JSON errors: {"error":"Unauthorized"} (maintainer APIs) vs {"error":"Subscription login required"} (customer APIs) — two distinct auth realms worth mapping separately.

4. Embedded credentials — verification pitfalls

• Firebase web keys in bundles: validate via POST https://identitytoolkit.googleapis.com/v1/accounts:createAuthUri?key=<KEY> with body {"continueUri":"https://TARGET","email":"probe@example.invalid"}.
  • MUST pass ?key= as query param. Omitting it returns a misleading 403 "unregistered callers" that looks like a restriction win.
  • 400 "API key not valid" = revoked/dead → informational, NOT a finding.
• Google OAuth client IDs served publicly (e.g. via /api/subscription/google-client-id) are by design public; judge by redirect-URI allowlist, never report secrecy alone.

5. Login rate-limit & enumeration test

≤10 sequential fake-credential POSTs, ~0.4s apart. Good control looks like: 2×401 then 429 with {"quarantined":true,"retryAfter":900}. Quarantine usually also swallows follow-up username probes (uniform responses ⇒ no user enum). Record that your IP is now quarantined ~15 min — schedule other tests on different endpoints meanwhile.

6. Recurring finding patterns (checklist)

1. Catalog dump unauth — e.g. GET /api/voucher-groups returning hidden items (isVisible:false), prices, stock, sales stats, internal UUIDs. Severity Low; PoC = plain GET asserting hidden items present.
2. Version disclosure — unauth /api/version; also sanity-check feed logic (latest < current = bonus bug).
3. Header gaps — nosniff/XFO/referrer-policy often present; CSP and HSTS typically missing on Cloudflare-fronted Next.js.
4. Sandbox endpoints shipped to prod — e.g. simulate-payment (UI warns "sandbox only"), shutdown. Verify gate unauth; flag residual risk: if a low-role CUSTOMER session can call payment-simulation, escalate toward Critical (needs authenticated test account).
5. No Cache-Control on API responses — defense-in-depth note when responses lack any caching headers.

7. SPA-fallback false-positive elimination

Unknown paths may 307→portal or return the SPA HTML shell with 200/404. Before claiming "exposed file", verify content-type + size differ from baseline HTML (real config = application/json/text; fallback = text/html fixed size).

8. Report format (user preference)

Consolidated markdown per target: Executive Summary table (# / Finding / Severity / Status) up top; each finding = CVSS v3.1 full vector, raw request/response evidence, impact, remediation, paste-ready reproduction block. Include a "defenses confirmed working" section — operators value positives. Deliver as zip: SUMMARY.md + findings/ + pocs/ + evidence/.

Pitfalls

• crt.sh JSON sometimes returns non-JSON (HTML challenge) — handle parse failure gracefully, don't conclude "no certs".
• dig/whois/jq/zip may be absent: DNS via DoH curl 'https://dns.google/resolve?name=X&type=Y'; archives via Python zipfile; JSON via python3 -m json.tool.
• Cloudflare-fronted targets: keep ≤10 req/s sequential during probing to avoid WAF 429 noise contaminating results.

References

• references/custom-router-case.md — full case file: endpoint catalog, verified defenses, payload shapes, PoC recipes from a completed engagement.