api-security
Skill39 starsChanged 34 days agoMITView on GitHub
API security testing for REST/GraphQL/WebSocket/SOAP: discovery, authz, BOLA/IDOR, JWT/OAuth, rate-limit, and CI-oriented validation workflows.
Files
API 安全测试
覆盖 REST / GraphQL / WebSocket / SOAP 全协议 10 阶段方法论,从发现到 CI/CD 集成
适用场景
- REST API 安全测试(OpenAPI/Swagger 驱动或盲测)
- GraphQL 安全审计(内省、批查询、别名过载)
- WebSocket 安全测试
- JWT / OAuth 2.0 认证测试
- BOLA/IDOR/BFLA 授权漏洞检测
- API 限速绕过与 DoS 测试
10 阶段测试流程
Phase 1: API 发现与侦察
主动发现:
□ Vespasian: 无头浏览器爬取 → 自动生成 OpenAPI 3.0 / GraphQL SDL 规范
□ Entropy --discover: 从 robots.txt + JS 文件提取端点
□ Kiterunner / ffuf: 爆破未文档化的端点路径
□ 检查常见路径: /swagger.json, /openapi.json, /graphql, /api-docs
GraphQL 内省(三级尝试):
1. 标准内省查询
2. 精简查询(绕过 WAF 全量封禁)
3. 仅查 __schema { types { name } }(最小探测)Phase 2: 认证测试
JWT 分析(jwt_tool / Burp):
□ alg:none 攻击: 修改头部为 "alg":"none",清空签名
□ 密钥混淆: RS256 公钥 → HS256 对称密钥
□ 弱 HMAC 密钥爆破: jwt_tool -C -d wordlist.txt
□ 过期/声明篡改: 修改 exp/iat/sub/role 声明
□ kid 注入: ../../etc/passwd → HMAC 签名绕过
OAuth 2.0:
□ redirect_uri 操控 → 授权码泄漏
□ CSRF via state 参数缺失
□ Token 在 Referer 头泄漏
□ PKCE 缺失检测
GraphQL 认证:
□ mutation 通过 GET 请求绕过认证(CSRF)
□ 批查询认证绕过Phase 3: 授权测试(BOLA/IDOR/BFLA)
BOLA(对象级授权绕过):
□ 遍历数字 ID: /user/1 → /user/2 → /user/3
□ 遍历 UUID
□ 遍历用户名/邮箱
□ Burp Autorize: 双会话重放对比
BFLA(功能级授权绕过):
□ 普通用户执行管理员 API
□ HTTP 方法切换: GET → PUT → PATCH → DELETE
□ API 版本降级: /v2/admin → /v1/admin
□ 批量操作注入: {"users": [1,2,3]} → {"users": [1,2,3,admin_id]}
工具: Burp Autorize, AuthMatrix, Entropy (malicious_insider persona)Phase 4: GraphQL 专项
内省泄漏 → 信息暴露检测
别名过载 → 100+ 别名 DoS
批查询 → 10+ 同时查询 DoS
字段重复 → __typename × 500
指令过载 → 递归 @skip/@include
循环查询 → 深度嵌套内省递归
字段建议 → 错误消息信息泄漏
GraphiQL/Playground 暴露 → IDE 公开风险
GET 突变 → CSRF 风险
追踪/调试模式 → 元数据泄漏
工具: FireTail, Escape DAST, api.sh (Phases 1-3)Phase 5: REST 输入验证
□ HTTP 方法切换: GET→POST→PUT→DELETE→OPTIONS→PATCH
□ Content-Type 篡改: JSON→XML→multipart
□ NoSQL 注入: {"username": {"$gt": ""}}
□ SSRF via URL 参数: webhook URL/头像 URL/导入 URL
□ XXE in XML 端点
□ 参数污染: /api?role=user&role=admin
□ 批量赋值: 向请求体添加 is_admin: truePhase 6: 业务逻辑与差分测试
□ Entropy compare: diff v1 vs v2 API → 状态码变化/字段删除/延迟回归
□ 多角色工作流测试: admin/user/readonly 权限矩阵
□ 优惠券/积分/价格操控
□ 竞态条件: 并发请求测试 TOCTOUPhase 7: WebSocket 测试
□ 端点发现
□ 消息注入(注入 payload、原型污染)
□ 超大消息处理
□ 类型混淆
□ 跨站点 WebSocket 劫持(CSWH)Phase 8: 限速与 DoS
□ 限速绕过 via 头部: X-Forwarded-For, X-Real-IP
□ 路径变体: /api/ → /api → /Api/ → /API/
□ Slowloris 低带宽耗尽
□ GraphQL 批查询深度嵌套 DoS
□ IP 轮换测试(ProxyCat 代理池)Phase 9: 数据暴露
□ 响应过度暴露: 对比 API 返回 vs UI 展示
□ 分页枚举: ?page=1&limit=10000
□ 错误消息信息泄漏: 堆栈跟踪/内部路径/SQL 错误
□ GraphQL 嵌套遍历访问越权数据
□ OpenAPI 规范暴露敏感端点Phase 10: CI/CD 集成
□ Entropy --ci --watch: spec 变更时自动重跑
□ Escape DAST: 按严重度阈值自动阻断构建
□ 发现持久化为回归测试
□ StackHawk(开发者优先、ZAP 内核)工具链
| 工具 | 用途 | 获取 |
| Vespasian | 流量 → OpenAPI/GraphQL 规范 | GitHub: praetorian-inc/vespasian |
| Entropy | LLM 生成攻击场景,5 personas | GitHub: arjinexe/entropy-chaos |
| Escape DAST | 业务逻辑安全测试 | escape.tech |
| api.sh | 8 阶段全协议攻击管道 | GitHub: Sharon-Needles/api |
| FireTail | GraphQL 12 专项测试 | firetail.ai |
| jwt_tool | JWT 全面测试 | GitHub: ticarpi/jwt_tool |
| Burp Autorize | 双会话授权对比 | Burp BApp Store |
参考
references/rest-graphql-testing.md— REST + GraphQL 深度测试references/jwt-oauth-testing.md— JWT + OAuth 安全测试
Other versions26 projects use a version of this.
Word for word · 0
None found.
Edited · 6
- zhaoxuya520/reverse-skill-private★ 40k85% same
- zhaoxuya520/reverse-skill★ 38k85% same
- 2233admin/reverse-skill-evolver★ 2694% same
- lza6/claw-code-tingfeng★ 585% same
- DorianGallo/reverse-skill★ 2100% same
- lushun2014/lushibo-skills★ 0100% same
Wrote their own version · 20
- hardw00t/ai-security-arsenal★ 101same name
- jefflester/claude-skills-supercharged★ 43same name
- jim60105/copilot-prompt★ 22same name
- awarexone/AXguard★ 17same name
- GoldenWing-360/claude-security-skills★ 16same name
- YepAPI/skills★ 8same name
- langbyyi/websec-skills★ 4same name
- Kur1sulab/blackbox★ 3same name
- LeahyCC/claude-skills★ 3same name
- Sanskar015/claude-skills-supercharged★ 3same name
10 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository85 other files it gives its agents.
- apk-reverseSkill
- attack-chainSkill
- binary-diffSkill
- cairnSkill
- code-sentinel-ctfSkill
- code-sentinel-engagementSkill
- code-sentinel-gpt56Skill
- code-sentinel-matrixSkill
- code-sentinel-opsSkill
- code-sentinel-payloadSkill
- code-sentinel-ptesSkill
- code-sentinelSkill
- cyberstrikeSkill
- diagram-generatorSkill
- docs-generatorSkill
- edr-bypass-reSkill
- firmware-pentestSkill
- gpt56-instruct-routerSkill
- gpt56-reverse-pentestSkill
- ida-reverseSkill
- js-reverseSkill
- llm-securitySkill
- malware-analysisSkill
- mobile-competitive-game-securitySkill
- mobile-reverseSkill
- patch-diff-exploitSkill
- pdfSkill
- pentest-toolsSkill
- src-hunterSkill
- playwright-interactiveSkill
- playwrightSkill
- pojia-anticheatSkill
- pojia-auto-opsSkill
- pojia-evasionSkill
- pojia-exploitSkill
- pojia-game-hackSkill
- pojia-game-redteamSkill
- pojia-game-securitySkill
- pojia-labSkill
- pojia-license-securitySkill
- pojia-malwareSkill
- pojia-mature-contentSkill
- pojia-memorySkill
- pojia-pentestSkill
- pojia-reverseSkill
- pojia-social-engSkill
- pojia-unlimitedSkill
- pwn-chainSkill
- radare2Skill
- range-research-scopeSkill
- recover-license-native-e2eSkill
- reverse-engineeringSkill
- screenshotSkill
- security-best-practicesSkill
- security-ownership-mapSkill
- security-threat-modelSkill
- shadowtrace-anti-debugSkill
- shadowtrace-artifact-triageSkill
- shadowtrace-binary-analysisSkill
- shadowtrace-binary-protectionSkill
- The first 60 shown.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

