attack-surface-mapping — Claude Code skill for app recon
attack-surface-mapping draws a testable attack surface from one authorized URL or application. Use it when you see only a login page and need to know what to test.
MIT 1,000 installs ★ 2,677 updated 2026-09-13
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a claude-codeWhat attack-surface-mapping does
The skill takes one target and one application, pulls out hosts, APIs, keys, and the object graph from what the app already exposes. Output is a portrait, host/API inventory, key table, response-class labels, and the object graph. It is the fast path: reconnaissance first, testing lives in other skills.
One hard rule: no directory brute or payload spray at the start. Write a three-to-five-sentence portrait first: who uses the app, core objects, which fields hold money or privilege, what an unauthenticated caller can touch. Cannot write it? Capture one real page's traffic first.
Field patterns live in SURFACE_PATTERNS.md. A login page is a shell; the business plane hides behind it: service=, redirect_uri=, env.js, baseURL, apiHost. Frontend inventory includes scripts, sourcemaps, /api/ paths, keys, hidden routes. Empty inventory plus "I will brute paths next" is a failed mapping.
What it helps with
- find business hosts behind a login page
- build API and key inventory from frontend JS
- label responses as login-gate, differential, or unauth exception
- grow the object graph from responses
- tell alive hosts from dead ones without brute force
Example prompts for the agent
- Map the attack surface for this URL; I only see a login page
- Build a JS inventory: paths, keys, hidden routes — then classify responses
- Walk the object graph from responses, starting with list and detail
How to install attack-surface-mapping
1Skills CLI (Vercel)
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a claude-code2Manually: put the files in place
~/.claude/skills/attack-surface-mapping/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/attack-surface-mapping/
1Skills CLI (Vercel)
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a codex2Manually: put the files in place
.agents/skills/attack-surface-mapping/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a cursor2Manually: put the files in place
.cursor/skills/attack-surface-mapping/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install yaklang/hack-skills skills/attack-surface-mapping2Manually: put the files in place
.github/skills/attack-surface-mapping/SKILL.md1Skills CLI (Vercel)
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a gemini-cli2Platform CLI
gemini skills install https://github.com/yaklang/hack-skills --path skills/attack-surface-mappingManually: put the files in place
.devin/skills/attack-surface-mapping/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add yaklang/hack-skills --skill attack-surface-mapping -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about attack-surface-mapping
How do I install attack-surface-mapping in Claude Code?
Run npx skills add yaklang/hack-skills --skill attack-surface-mapping -a claude-code. The skill file lands at ~/.claude/skills/attack-surface-mapping/SKILL.md.
What does attack-surface-mapping do?
From one authorized URL or application it builds a portrait, host and API inventory, key table, response classes, and the object graph. It prepares testing, it does not test.
Why can't I start with directory brute force?
The skill forbids brute and payload spray until the portrait and inventory are ready. An empty inventory plus a plan to brute paths counts as a failed mapping.
Source text
SKILL.md · MIT
--- name: attack-surface-mapping description: >- Draw a testable attack surface from one authorized target URL or one application. Use when the user says 攻击面, 供给面, 画攻击面, map the surface, application recon, find the business host, JS inventory, or when the only visible page is login. Derive hosts, APIs, keys, and the object graph from what the app already exposes. Do not open with directory brute or payload spray. Use when the user runs /attack-surface-mapping. --- # Attack Surface Mapping Given **one target** and **one application**, draw the surface from what that application already exposes. Then stop. Testing lives in other skills. This skill is the fast path. Success is a portrait plus a host/API inventory plus a key table plus response-class labels plus an object graph. A probe count is not success. Field patterns for where the rest of the surface actually lives: [SURFACE_PATTERNS.md](./SURFACE_PATTERNS.md). ## When - A new URL, a new app, or "I only see a login page" - Need to know **what** to test before loading injection / auth / upload skills - The agent is about to brute directories, spray quotes, or expand to unrelated hosts Do **not** use this skill to expand an organization-wide host universe. Map the current application cluster. Finish it. Then, if scope allows, take the next cluster. Authorization and destruction bounds: [hack](../hack/SKILL.md) start gate. Stay in scope. ## What must exist (not a file tree) Before any vulnerability skill, these facts must be retrievable. Persist them in whatever the local workspace already uses for notes and evidence — an existing task folder, a proxy project, session notes, a ticket. Do not invent a new directory layout when one is already in play. | Fact | Keep | |---|---| | Portrait | 3–5 sentences, not an essay | | Hosts | Business hosts, gateways, API domains this app already named | | Endpoints | Method, path, params, auth required? | | Keys | Signing salt, ciphertext id + frontend pubkey, hidden/admin route, hardcoded demo account. Record `none` per row if absent | | Response class | Login-gate / differential / unauthenticated exception | | Object graph | list → detail → attachment / export / approval | Requests, diffs, and screenshots stay where they were captured when that store is already the working set. Empty inventory plus "I will brute paths next" is a failed mapping. ## Fast path ``` Portrait → find the business plane (login is a shell) → inventory from the app (JS / traffic / docs), keys not just paths → classify responses → grow the object graph from responses → same-skin / same-gate collapse → surface-done → hand off ``` Do not insert directory brute, full-template scanning, or password spraying into this loop. ### 1. Portrait (mandatory, before any request spray) Three to five sentences: - Who uses this (consumer / merchant / operator) - Core objects (order, ticket, coupon, document, tenant) - Which fields hold money, privilege, or state - What an unauthenticated caller can already touch Cannot write it → capture one real page's traffic first. Do not scan into a blank portrait. Mini-program, native app, GraphQL, WebSocket, batch export, agent-with-tools, template preview, file convert, command RPC: treat as **this site's surface**, not a sidenote. ### 2. Find the business plane A login page is a shell. The surface is the post-login business host or the **same-host gateway behind the form**. Seeing a login page is not a reason to change assets. Find the plane without logging in: - Query: `service=` / `redirect_uri=` / `callback=` / `returnUrl=` / `jumpUrl=` / `next=` - Client: `env.js` / `baseURL` / `apiHost` / `/prod-api` / `VUE_APP_*` / `REACT_APP_*` - Transport: 302 `Location`, `X-Frame-Options: ALLOW-FROM` - Naming: same-product `api` / `admin` / `gateway` / product host Someone else's SSO / CAS / OAuth page: do not audit the identity product. Follow it back to **this** product's business plane. Criterion: the login page's owner is not this business. Alive vs dead: | Treat as alive | Treat as dead | |---|---| | 401, 403, login wall, admin challenge | Timeout, parking page, no business response | | Management console challenge | Default CDN / empty static shell with no script | Alive ≠ grind the form. Captcha OCR, slider farms, and login-box dictionaries are not mapping. Form checks that are in-scope for mapping (once, then stop): empty password, skip-password step, extra fields on the login API (`tenant` / `corpId` / `moduleId`), business paths already visible next to the form (list / detail / stats). Username and password boxes are not business parameters. ### 3. Inventory from the application **Frontend present:** open a business page → collect scripts (including async chunks and sourcemaps) → extract APIs **and keys** → capture traffic to fill gaps → persist endpoints and keys in the local evidence store. JS extracts more than `/api/` paths. For each row, write the value or `none`: | Extract | Why it is surface | |---|---| | `/api/` paths, RPC cmd numbers, GraphQL operations | Endpoint list | | Signing salt, hardcoded key, `sign` that does not need a cookie | Replay without a session | | Ciphertext id + frontend public key (`modulus` / JSEncrypt) | Neighbor-id is encryptable | | Hidden / admin routes in the router, unpublished chunks | APIs the UI never shows | | Hardcoded demo account, test tenant, experience entry | A key, not a login-form dictionary | | Command / template / expression / file-convert / RPC-with-exec / agent tools | Execution plane; do not invent params if none exist | **No frontend / JS blocked:** Swagger, OpenAPI, captured traffic, HTML inline, known gateway prefixes. Do not idle waiting for a full JS dump. **Frontend present but inventory empty → directory brute** is forbidden. Docs and debug planes, if this app already linked them: [api-recon-and-docs](../api-recon-and-docs/SKILL.md). Exposed VCS / backups: [insecure-source-code-management](../insecure-source-code-management/SKILL.md). Both are *this cluster's* extra planes, not a new search. ### 4. Classify responses before any probe | Response | Class | Mapping action | |---|---|---| | One sentence "please log in" / `NotLogin`, no list / total / roster / detail | Login gate | Do not mark as an injection surface | | Has list / total / business fields, even `total=0` | Differential | Keep on the test list | | Error, 500, timeout, or length/timing off baseline | Differential | Keep; unstable diff → stop after one or two compares | | Missing a parameter dumps a roster or detail | Unauthenticated exception | Highest-priority unauth surface | An empty list is "structure returned, count is 0", not a login gate. ### 5. Object graph is surface Identifiers in any field name (`id` / `userId` / `tenantId` / `fileKey` / `openid` / ciphertext PK) are surface. Walk, and write it down: ``` list → detail → attachment / export / preview / approval ``` Parent authorized, child often not. After a list passes, the next surface is the attachment, not a new host. Anything a response newly names — id, download URL, token, internal host, role field — goes onto **this site's** queue immediately. Do not drop it when changing pages. ### 6. Collapse same-skin and same-gate Fingerprint-same (title, skeleton, build hash, API prefix, login chain): pick 2–3 representatives. Siblings only get a glance: new path / new port / another app / another `jump` / `service=` / `moduleId` business plane. No new plane → do not open a full matrix. **Same-gate** (all four): 1. Business paths already extracted 2. Unauthenticated exceptions already probed 3. Remaining business endpoints return the **same login-code family** 4. No unauthenticated other-subject data, and no missing-param dump Then stop this gate. Siblings with the same `baseURL` + same code: glance for new paths or a code change. Auth endpoints (issue session / reset / rebind / ticket-swap) still go on the list if present; they are not "please log in, so skip". Same host is not same-skin. Extra paths on the same host always expand. No business script, default server page, or leftover behind the same login-code family: falsify once and leave. Do not pad with a full matrix. ### 7. Surface-done (then hand off) Mapping is done when: 1. Portrait exists, or this is recorded as a shell with no business object 2. Endpoint inventory exists (full or degraded); keys have values or `none` — in the local evidence store, not a prescribed path 3. Response classes labeled; unauthenticated exceptions ticked 4. Object graph recorded, or recorded as unlistable 5. Same-skin / same-gate leftovers glanced, not rematrixed 6. Auth endpoints listed if the inventory has issue-session / reset / rebind / ticket-swap / 2FA — listed, not yet exploited Then load [hack](../hack/SKILL.md) for effort order and the matching category skill. Do not start testing inside this file. ## Handoff | Surface you drew | Load | |---|---| | Unauthenticated other-subject data, object ids | [auth-sec](../auth-sec/SKILL.md), [idor-broken-object-authorization](../idor-broken-object-authorization/SKILL.md) | | Issue-session / reset / rebind / ticket-swap | [authbypass-authentication-flaws](../authbypass-authentication-flaws/SKILL.md) | | Differential filters, URL-fetch params, templates | [injection-checking](../injection-checking/SKILL.md) | | Upload / preview / convert | [upload-insecure-files](../upload-insecure-files/SKILL.md) | | Money / coupon / stock / approval | [business-logic-vuln](../business-logic-vuln/SKILL.md) | | REST / GraphQL / gateway docs | [api-sec](../api-sec/SKILL.md) | | Public middleware admin | [unauthorized-access-common-services](../unauthorized-access-common-services/SKILL.md) | ## Anti-patterns - Opening with directory brute or full-template scanning - Grinding captcha / default passwords on login HTML - Seeing a login page and changing assets - Script extracted only paths; salts, ciphertext ids, hidden routes, demo accounts unread - Treating 401 / 403 / login wall as dead - Pouring a full matrix into the same login-code family - Expanding to a new cluster while this application's live surface is unfinished - Calling mapping "done" because the homepage returned 200 - Minting a new notes tree when the workspace or proxy project already holds the evidence
© yaklang · MIT · yaklang/hack-skills