OpenHunterAI - Local-first AI red team for web, API, and LLM appl

项目说明

<p align="center">
<img src="frontend/public/logo.png" width="88" alt="OpenHunterAI">
</p>

<h1 align="center">OpenHunterAI</h1>

<p align="center">
<strong>Your local AI red team.</strong><br>
Attacker-style reasoning for web, API, and LLM application security.
</p>

<p align="center">
<a href=" src=" alt="CI"></a>
<a href=" src=" alt="Preview release"></a>
<a href="LICENSE"><img src=" alt="PolyForm Noncommercial license"></a>
<a href=" src=" alt="GitHub stars"></a>
</p>

<p align="center">
<a href="#quick-start">Quick start</a> ·
<a href="#coding-agent-skill">Agent skill</a> ·
<a href="#assessment-model">Assessment model</a> ·
<a href="#architecture">Architecture</a> ·
<a href="docs/README.md">Docs</a> ·
<a href="#star-history">Star history</a>
</p>

OpenHunterAI brings scope, scan activity, findings, and remediation into one
local workspace. Start without an account, connect your model pro, and
assess public applications you own or are authorized to test.

| Capability | What you get |
| --- | --- |
| Local workspace | No signup; projects and scan records in your local database |
| Explicit scope | Domain , deterministic plans, human appro |
| Multiple signal sources | Browser inspection, recon, ZAP, and a Nuclei adapter |
| AI-assisted in | OpenHack/Strix adapter boundaries; bring your own runtimes |
| E & remediation | Sanitized findings, co gaps, fix guidance, manual retest |
| Portable workflow | A skill for Codex, Claude Code, Gemini CLI, Grok Build, and OMP |

> Alpha. Integration co depends on configured runtimes. A healthy
> workspace or an empty report does not prove a target is secure.

Assessment model

OpenHunterAI keeps the useful core of the
48bf9036 technical requirements:
assess externally observable behaviour from an attacker's perspecti
code is not required. Every run stays constrained to a public target
and an appro.

| Stage | Output | Control |
| --- | --- | --- |
| Scope lock | host, appro, scan plan | Authorization before work begins |
| Surface mapping | Browser and recon signals | Public, in-scope surface only |
| Baseline | Normal responses and co context | Bounded workers and adapter policies |
| Hypothesis testing | Candidate security signals | No destructi; sensiti needs approval |
| E gate | Sanitized, reproducible finding | Separate obser from analyst judgement |
| Report and retest | Fix guidance, limitations, manual confirmation | A human owns impact decisions and closure |

OpenAPI/Swagger descriptions, Postman examples, sanitized HTTP captures, and
dedicated test accounts can help define the agreed test surface. Automated
import or retention of raw sensiti is not a promised capability.

Quick start

Requires Git, Node.js 22+, and Docker with Compose v2.

Open localhost:3001. First startup builds the images,
applies database migrations, and creates a private .en with generated secrets.

1. Set provider keys and model IDs in .en; run node ops/local.mjs restart.
2. Create a project and its public domain.
3. Authorize the scope, start a scan, and review its activity and co.

| Command | Purpose |
| --- | --- |
| node ops/local.mjs status | Inspect services |
| node ops/local.mjs logs public-api | Diagnose API errors |
| node ops/local.mjs stop | Stop; keep database |

Included: browser inspection, ZAP, recon, and the Nuclei adapter.
Bring separately: OpenHack and Strix runtimes.
Optional interactive target-login capture: node ops/local.mjs start --with-browser.

Before scanning, resolve the Nuclei template-policy blocker.
The adapter is packaged, but a re template bundle is not yet supplied.

Local means the workspace runs on your machine, not permission to scan
localhost/private networks. Keep its ports on loopback.
Setup, model configuration, and troubleshooting →

Coding-agent skill

Run in the application repository you want your agent to assess:

Reload your coding agent, then ask:

> Use OpenHunter to check my workspace and prepare a scoped security assessment.
> Wait for my appro starting a scan.

OMP discovers the shared .agents/skills installation through its Agents
provider. The skill includes a read-only API helper; installation does not start
a scan. Client compatibility →

Architecture

The public API builds the plan; Go workers execute it through bounded adapters.
Signals pass through the findings policy before entering reports. Reporting is
a backend service, not a separate scan worker.

Runtime & data flow ·
Worker mapping ·
Integration setup

Documentation

| Start here | Go deeper |
| --- | --- |
| Local setup | Architecture & data model |
| Product o | Product requirements |
| Agent workflow | LLM routing |
| Security guardrails | Acceptance & release gates |

All documentation ·
Known limitations ·
Historical technical specification

Community & license

Report a bug,
suggest an impro, or help clarify the docs.
Read contribution rules first; disclose
pri, without credentials or raw target data.

Source-a under PolyForm Noncommercial 1.0.0.
Commercial licensing: LICENSING.md.
Upstream tools and attribution retain their own notices.

Star history

![Star History Chart]()

项目信息

开发语言：TypeScript

Star 数：330

技术要点

该项目属于红蓝攻防方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Local-first AI red team for web, API, and LLM application security. Attacker-style reasoning, evidence-backed findings, and skills for AI coding agents.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。