bughunter-ai - Autonomous Bug Bounty Hunting Framework powered by

项目说明

<p align="center">
<img src="docs/images/bughunter-banner.png" alt="BugHunter AI" width="800" />
</p>

<h1 align="center">BugHunter AI</h1>

<p align="center">
<strong>Autonomous Bug Bounty Hunting Framework Powered by Claude Code + PAI</strong>
</p>

<p align="center">
<em>28 specialized AI agents. 8 orchestrated workflows. State-machine orchestration. 51 skills. Zero human input required.</em>
</p>

<p align="center">
<a href="#quick-start">Quick Start</a> &bull;
<a href="#full-setup-guide">Full Setup</a> &bull;
<a href="#pai-superpowers">Superpowers</a> &bull;
<a href="#architecture">Architecture</a> &bull;
<a href="#agents">Agents</a> &bull;
<a href="#sample-prompts">Prompts</a> &bull;
<a href="#contributing">Contributing</a>
</p>

<p align="center">
<img src=" alt="Claude Code" />
<img src=" alt="PAI v3.0" />
<img src=" alt="28 Agents" />
<img src=" alt="51 Skills" />
<img src=" alt="Bun" />
<img src=" alt="Burp MCP" />
<img src=" alt="MIT License" />
</p>

---

What is BugHunter AI?

BugHunter AI is an autonomous bug bounty hunting framework that turns Claude Code into an elite security researcher. Built on top of the PAI (Personal AI Infrastructure) system, it combines 20 specialized agents, 51 skills, 13 expert AI agents, 20 lifecycle hooks, and a full offensi toolchain into a single command:

What happens next:
1. A state machine initializes and tracks 10 phases of the hunt
2. Credentials are loaded from an encrypted vault (ne)
3. The app is profiled — flows mapped, tech stack fingerprinted, trust boundaries identified
4. Hypothesis-dri deploy in parallel — each with a specific attack theory
5. Findings are reported in real-time with C
6. A professional bug bounty report is generated automatically

---

Why BugHunter AI?

| Manual Bug Bounty | BugHunter AI |
|---|---|
| Hours of recon before first payload | Autonomous recon → profiling → attack in minutes |
| Forget where you left off between sessions | State machine checkpoints every phase — --resume anytime |
| Credentials in plaintext notes | Encrypted vault with 1Password integration |
| Run same tools blindly on every target | Hypothesis-driven: agents attack WHERE the AppProfile says bugs live |
| Test one thing at a time | 5 agents run in parallel, findings shared in real-time |
| Miss AI/LLM | Dedicated LLMSecurityAgent with OWASP LLM Top 10 |
| Medium findings silently dropped | Mediums archived for attack chain correlation |
| No memory between hunts | Cross-session learning — gets smarter with every engagement |
| Single tool / single domain | 51 skills co web, mobile, API, cloud, network, binary, AI |

---

The Full Stack: PAI + Superpowers + BugHunter

BugHunter AI is not just a skill — it's part of a complete offensi infrastructure. Here's what you get when you install the full setup:

PAI (Personal AI Infrastructure) v3.0

The foundation layer. PAI is a general problem-solving system that provides:

Superpowers Plugin

Claude Code's official superpowers plugin adds enhanced capabilities:

Security Skill Stack (16 Skills)

| Skill | Co |
|-------|----------|
| BugBountyFramework | Autonomous 10-phase hunt orchestration |
| WebAssessment | OWASP WSTG v5, OWASP Top 10 |
| SecurityHub | Master security command center with intelligent skill routing |
| Offensi | Kill chain tracking, adapti |
| APISecurityTesting | REST, GraphQL, gRPC, BOLA, BFLA, mass assignment |
| MobileSecurity | OWASP MASTG v2, Frida, Objection, MobSF |
| NetworkSecurity | AD attacks, BloodHound, Kerberoasting, pi |
| CloudSecurity | AWS/Azure/GCP, IAM escalation, Pacu, ScoutSuite |
| ExploitDev | Heap exploitation, ROP chains, safe-linking bypass |
| Re | Ghidra, radare2, binary analysis |
| MalwareAnalysis | Static/dynamic analysis, YARA, IOC extraction |
| PromptInjection | OWASP LLM Top 10, MITRE ATLAS |
| | C, AFL++ fuzzing, CodeQL |
| SASTOrchestration | Semgrep, CodeQL, custom rules, taint tracking |
| SCASecurity | SBOM, supply chain, dependency confusion |
| ThreatModeling | STRIDE, PASTA, attack trees |
| Recon | Subdomain enum, asset discovery, Shodan |
| RedTeam | 32-agent ad analysis |

Expert Agent Army (13 Agents)

| Agent | Specialty |
|-------|-----------|
| Pentester | Offensi specialist — assessment, exploitation |
| Engineer | Elite principal engineer — TDD, strategic planning |
| Architect | System design — constitutional principles, feature specs |
| Algorithm | PAI Algorithm expert — ISC creation and e |
| Designer | UX/UI specialist — Figma, shadcn/ui |
| Artist | content — Flux, GPT-Image-1, prompt engineering |
| QATester | Browser-automation — Gate 4 completion gates |
| Intern | 176-IQ generalist — multi-faceted problem solving |
| ClaudeResearcher | Multi-query academic research WebSearch |
| CodexResearcher | Technical archaeology — O3, GPT-5-Codex consultation |
| GeminiResearcher | Multi-perspecti Gemini |
| GrokResearcher | Contrarian analysis via xAI Grok |
| PerplexityResearcher | In journalism |

MCP Servers (3 Core + 5 Recommended)

Core (Pre-configured):
| Server | Purpose |
|--------|---------|
| Burp Suite MCP | Proxy traffic analysis, scope sync, Collaborator, HAR export |
| Filesystem MCP | Direct file system access |
| GitHub MCP | Repository operations, PR management |

Recommended Security MCPs:
| Server | Purpose | Install |
|--------|---------|---------|
| Shodan | Internet asset search | npm install -g /mcp-server |
| | Malware/IOC analysis | npx -y @/mcp-server |
| CVE/NVD | database | Custom Python server |
| Nuclei | scanning | Custom Node server |
| GitHub Security Ad | C | MCP |

Plugins (4 Active)

| Plugin | Purpose |
|--------|---------|
| superpowers | Enhanced Claude Code capabilities |
| claude-mem | Persistent cross-session memory |
| swift-lsp | Swift language ser |
| ui-ux-pro-max | Ad UI/UX design capabilities |

Lifecycle Hooks (20 Hooks)

| Hook | Trigger | Purpose |
|------|---------|---------|
| Security | Pre: Bash, Edit, Write, Read | all tool calls for security |
| | Pre: Bash | Voice notification gate |
| AgentExecutionGuard | Pre: Task | Guards agent execution |
| SkillGuard | Pre: Skill | skill in |
| SetQuestionTab | Pre: AskUserQuestion | Terminal tab management |
| AlgorithmTracker | Post: Bash, Task* | Tracks algorithm phase progression |
| QuestionAnswered | Post: AskUserQuestion | Tracks Q&A flow |
| WorkCompletionLearning | SessionEnd | Captures learnings from work |
| SessionSummary | SessionEnd | Generates session summaries |
| RelationshipMemory | SessionEnd | Builds relationship context |
| UpdateCounts | SessionEnd | Updates system statistics |
| IntegrityCheck | SessionEnd | system integrity |
| RatingCapture | UserPromptSubmit | Captures quality ratings |
| AutoWorkCreation | UserPromptSubmit | Auto-creates work tracking |
| UpdateTabTitle | UserPromptSubmit | Dynamic terminal tab titles |
| SessionAutoName | UserPromptSubmit | Auto-names sessions |
| StartupGreeting | SessionStart | Displays PAI banner |
| LoadContext | SessionStart | Loads session context |
| Check | SessionStart | checks |
| StopOrchestrator | Stop | Clean shutdown handling |

---

Architecture

---

Quick Start

For the minimal BugHunter-only install:

For the full setup with PAI, Superpowers, and all security skills, see the Full Setup Guide.

---

Full Setup Guide

See SETUP.md for the complete, step-by-step guide to replicate the full infrastructure:

1. Install prerequisites (Claude Code, Bun, Go tools, Burp Suite)
2. Install PAI v3.0 (the foundation)
3. Configure your identity (DA + Principal)
4. Install plugins (Superpowers, claude-mem, ui-ux-pro-max)
5. Configure MCP servers (Burp, Filesystem, GitHub)
6. Install BugHunter AI skill
7. Set up optional security MCPs (Shodan, , NVD)
8. Configure notifications (ntfy, Discord, Twilio)
9. the installation

---

Hunt Modes

| Mode | C | Finding Target | Best For |
|------|----------------|----------------|----------|
| bounty (default) | >= 8.0 | 10 | Bug bounty programs — only critical/high findings |
| pentest | >= 4.0 | 20 | Penetration tests — comprehensi |
| comprehensive | >= 0.0 | 50 | Full security audits — e documented |

---

Agents

BugHunter deploys 20 specialized agents, each an expert in one class:

| Agent | Focus | Key Techniques |
|-------|-------|----------------|
| AppRe | Application understanding | Flow mapping, tech fingerprinting, trust boundary analysis |
| LLMSecurityAgent | AI/LLM | OWASP LLM Top 10, prompt injection, RAG poisoning |
| XSSAgent | Cross-site scripting | Reflected, stored, DOM-based, mutation XSS |
| SQLiAgent | SQL injection | Union, blind, time-based, second-order SQLi |
| SSRFAgent | Ser request forgery | Cloud metadata, internal services, protocol smuggling |
| IDORAgent | Insecure direct object refs | Horizontal/ pri escalation, UUID prediction |
| AuthAgent | Authentication bypass | JWT attacks, session fixation, OAuth flaws, MFA bypass |
| APIAgent | API security | BOLA, mass assignment, rate limiting, GraphQL introspection |
| CORSAgent | CORS misconfiguration | Origin reflection, null origin, wildcard subdomains |
| FileUploadAgent | File upload attacks | Content-type bypass, polyglot files, path tra |
| XXEAgent | XML external entities | Blind XXE, OOB data exfiltration, SSRF via XXE |
| RCEAgent | Remote code execution | Command injection, SSTI, deserialization, SSRF→RCE |
| BusinessLogicAgent | Business logic flaws | Race conditions, price manipulation, workflow bypass |
| MobileAgent | Android/iOS security | SSL pinning bypass, exported components, insecure storage |
| WindowsAgent | Windows/AD attacks | Kerberoasting, NTLM relay, pri escalation |
| ReconAgent | Reconnaissance | Subdomain enum, port scanning, tech fingerprinting |
| Re | Binary analysis | Static analysis, dynamic analysis, identification |
| ExploitDe | Exploit de | PoC creation, payload crafting, reliability testing |
| DesktopAppAgent | Desktop app security | Electron, .NET, Java app testing |
| GraphQLAgent | GraphQL security | Introspection, batch abuse, nested DoS, relay IDOR, subscription hijack |
| WebSocketAgent | WebSocket security | CSWSH, message injection, origin bypass, auth hijacking |
| CSRFAgent | CSRF exploitation | Token/SameSite/Referer bypass, content-type tricks, CORS+CSRF chains |
| CachePoisoningAgent | Web cache poisoning | Unkeyed header injection, cache deception, CDN-specific bypasses |
| HTTPSmugglingAgent | HTTP request smuggling | CL.TE, TE.CL, H2.CL desync, response queue poisoning |
| SubdomainTakeo | Subdomain takeover | Dangling DNS, cloud service fingerprinting, cookie/CSP impact |
| RaceConditionAgent | Race conditions | HTTP/2 single-packet attack, limit bypass, double-spend, TOCTOU |
| PrototypePollutionAgent | Prototype pollution | Client-side PP→XSS gadgets, ser PP→RCE, AST injection |
| LLMAgent | Legacy LLM testing | Basic prompt testing (superseded by LLMSecurityAgent) |

How Agents Work

Agents don't blindly scan. They receive specific hypotheses from the AppProfile:

This means 90% less noise and 10x faster confirmation.

---

Orchestrator Workflows

The hunt orchestrator classifies your target and dispatches the appropriate workflow. Each workflow defines its own phases, agent dispatch order, parallelism, and gate conditions.

| Workflow | Trigger | Phases | Agents | Lines |
|----------|---------|--------|--------|-------|
| W_HUNT_WEB | Web application URL | 10 | 21 agents (parallel) | 2,018 |
| W_HUNT_API | API endpoint / Swagger / GraphQL | 9 | 10 agents | 487 |
| W_HUNT_LLM | AI/LLM app (chatbot, RAG, copilot) | 13 | 8 agents | 688 |
| W_HUNT_MOBILE | APK / IPA file | 10 | 7 agents | 1,037 |
| W_HUNT_NETWORK | IP range / CIDR / AD target | 9 | 5 agents | 814 |
| W_HUNT_CLOUD | AWS / Azure / GCP en | 10 | 5 agents | 607 |
| W_HUNT_THICK_CLIENT | Electron / .NET / Ja app | 10 | 6 agents | 874 |
| W_RECON | Standalone recon request | 10 | 2 agents | 703 |

How Workflows Work

Each phase has gate conditions — the workflow only ad when the gate criteria are met. Agents within a phase run in parallel for maximum speed.

---

Tool Chain

| Tool | File | Purpose |
|------|------|---------|
| Hunt Orchestrator | Tools/hunt-orchestrator.ts | State machine — phase tracking, checkpointing, resume, dashboard |
| Credential Vault | Tools/credential-vault.ts | Encrypted credential storage, 1Password, env vars, auto-redact |
| Auth Manager | Tools/auth-manager.ts | B2C/OAuth/SAML automation, session persistence, health checks |
| Burp Bridge | Tools/burp-bridge.ts | Burp Suite REST API bridge — scope sync, HAR export, Collaborator |
| Browser Harness | Tools/playwright-harness.ts | Browser automation CLI (primary) / Playwright CLI (fallback) |
| Appium Harness | Tools/appium-harness.ts | Mobile app testing — Android/iOS through proxy |

---

Sample Prompts

Your First Hunt

Hunt with Stored Credentials

Pentest Mode (Find More)

Hunt an AI Application

Resume a Hunt

Full Power Hunt

Use Security Skills Directly

See examples/sample-prompts.md for more.

---

Li

Check hunt progress anytime:

---

Directory Structure

---

How It Differs from Other Tools

| Feature | BugHunter AI | Nuclei/Burp Scanner | Manual Testing |
|---------|-------------|---------------------|----------------|
| Intelligence | Understands the app first, then attacks | Signature matching | Human expertise |
| Context | Remembers across sessions | Stateless | Notes/memory |
| Hypothesis-driven | Tests specific theories | Tests e | Depends on researcher |
| AI/LLM testing | First-class OWASP LLM Top 10 | Not supported | Rare expertise |
| Parallel agents | 5 specialized agents simultaneously | Single scanner | One person |
| State machine | Checkpoints, resume, ne progress | Run from scratch | Bookmarks/notes |
| Credential security | Encrypted vault + 1Password | Config files | Plaintext notes |
| Multi-domain | Web + Mobile + API + Cloud + Network + Binary | Single domain | Limited scope |
| Algorithm | 7-phase structured reasoning (PAI) | None | Informal methodology |
| Skills ecosystem | 51 skills, 16 security-focused | Plugin-based | Tool-dependent |

---

Responsible Use

This framework is designed for authorized security testing only:

BugHunter AI enforces scope: The framework includes hard scope enforcement that blocks testing out-of-scope targets. Configure your scope before hunting.

The maintainers are not responsible for misuse. Always follow your program's rules of engagement.

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Ideas for contributions:

---

License

MIT License. See LICENSE for details.

---

Acknowledgements

---

<p align="center">
<strong>Built with Claude Code + PAI by <a href="
</p>

<p align="center">
<em>If BugHunter AI helps you find bugs, give it a star!</em>
</p>

项目信息

开发语言：TypeScript

Star 数：82

技术要点

该项目属于安全软件方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Autonomous Bug Bounty Hunting Framework powered by Claude Code. 20 AI agents, state-machine orchestration, Burp Suite MCP, credential vault, LLM security track. Type 'hunt target.com' and let AI find the bugs.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。