js-recon
Analyze JavaScript bundles from any web app for API endpoints, GraphQL operations, hardcoded secrets, internal service hosts, per-environment config maps (prod/staging/QA), and admin/debug surface. Us
SKILL.mdskills/js-recon/SKILL.md
GitHubJS Recon Analyzer
Extract maximum intelligence from frontend JavaScript with zero requests to the target beyond the JS files themselves (they're public static assets).
Workflow
- Collect bundle URLs. Sources, in order of value: page HTML
<script src>+modulepreloadlinks, microfrontend manifests (/micro-frontends/*/manifest.json, import maps), URLs seen in proxy history, wayback CDX hits (archived bundles reveal endpoints the current build dropped), and the app's entry shims — entry files often import bigger chunks (import"./vendor-XXX.js"). Follow imports one level deep; the main logic hides in vendor/route chunks, not the 1KB entry shim. Also check for static data files: some sites ship whole catalogs as JS (data/catalog.jssettingwindow.SOMETHING=[...]— one 5MB file can contain the entire dataset). - Download each bundle once to a local
js/dir (static CDNs are usually not WAF-protected; app hosts might be — see curl_cffi note in bb-pipeline). - Mine with
scripts/mine_js.py js/*.js— extracts, grouped:- REST/GraphQL path strings (quoted and backticked, with
${}→{ID}normalization) - GraphQL
query/mutationoperation names - Secret patterns: AKIA*, AIza*, sk_live, private key blocks, raw JWTs,
pub…Datadog client keys (flag as public-by-design) - Internal hosts:
*.svc.cluster.local,internalUrl/internalHostvalues, RFC1918/hostnames that shouldn't be client-side - String-concatenation bases (e.g.
CUSTOMER_BASE_PATHconstants) — grep 200 chars of context to find the appended suffixes
- REST/GraphQL path strings (quoted and backticked, with
- Build the dictionary: normalize to
{ID}-parameterized paths, group by service prefix, count occurrences. This becomes the endpoint test list for Phase 4 — not a fuzz wordlist.
Deep passes beyond the extractor (do these manually on big finds)
- Config object mining: search for
ApiUrl:",baseUrl:,authUrl,kycApiUrl,gatewayassignments — one bundle's config block often maps the whole per-region/per-tier topology (prod/staging/qa/sandbox hosts + feature toggles). Staging/QA hosts found this way are new attack surface; resolve + fingerprint them. - Auth-model reconstruction: find the header-builder function (e.g.
const Ke=()=>{...Authorization:Bearer...}) — does it add auth only-if-authenticated? Then anonymous calls are expected somewhere: find which resources the anonymous flow touches (lead vs signup tracks) — that's the pre-auth attack surface. - URL-concat sites: for each API base var, locate
.concat(and template-literal uses to recover the full endpoint shapes (path params, query params like?searchTerm=). - Turnstile/proof gates: note where the app attaches bot-proof headers (
cf-turnstile-token) — those endpoints are script-hostile; plan the browser-pass for them instead of curl.
Reading the results (what history says matters)
- Version numbers: endpoints appear per TLD/environment with different versions — probing
v{n-1}/v{n+1}on each service occasionally exposes legacy versions (expect mostly 404; log it and move on). - GraphQL operation names are an API map without auth: mutations reveal state-changing surface (e.g.
updateCompanyDetails,uploadCompanyDocuments) — the authz test plan writes itself. - i18n keys and component names (
addressesEditAddress,OrderDetails) reveal features whose backing endpoints you should hunt in route chunks. - Env-var remnants (
VITE_*) leak BFF path maps and feature flags. - Zero secrets is the common case on mature programs — say so plainly and move on; the path dictionary is the real prize.
Rules
- Never execute or eval extracted code; strings are data only.
- Treat extracted endpoints as leads — every one still needs an authorized test before it means anything.
- If the target's app host blocks scripted downloads (Cloudflare JA3), use
curl_cffiwithimpersonate="chrome"/"edge", or have the user's browser fetch and save the files.