AI-Honeypot - 🛡️ MIRAGE: Open-source AI Honeypot for active defe

项目说明

MIRAGE — AI Honeypot

!MIRAGE banner

> Alpha software. APIs, data formats, and configuration may change without notice. Not recommended for production use without additional hardening. Bug reports and feedback welcome via Issues.

Detect, deceive, and study AI-dri in real time.

MIRAGE sits in front of your AI agents and intercepts ad traffic — prompt injections, jailbreaks, data exfiltration attempts, role manipulation, and more. When an attack is detected, MIRAGE doesn't block it silently: it spins up a con decoy persona that feeds the attacker fabricated data while logging every move for threat intelligence.

---

How it works

1. Every message hits Lobster Trap — a local DPI sidecar that scores it for injection patterns, role manipulation, exfiltration attempts, obfuscation, and credential requests.
2. Low-risk messages are forwarded transparently to your real upstream AI (OpenAI-compatible API or pass-through).
3. High-risk messages are routed to a decoy persona. The attacker interacts with a con fake AI that ser data, honey tokens, and synthetic credentials — none of it real.
4. Every session is recorded: full transcript, MITRE ATLAS technique tags, risk timeline, IOC feed.
5. The dashboard shows e live — attack theater, session dossiers, intel timeline.

---

Dashboard

The Theater view replays any acti turn-by-turn, annotated with MITRE ATLAS technique classifications:

Other views: Intel (captured attacks feed), Library (personas + datasets), Settings (thresholds, integrations, upstream config).

---

Quick start

Demo mode (no API keys required)

Open — the dashboard loads with simulated attack sessions cycling through all 9 technique categories.

Production mode

Edit .env:

The stack starts four services: Redis, Lobster Trap (DPI), Backend (Go API), Nginx (static frontend + reverse proxy).

---

Architecture

Backend packages

| Package | Responsibility |
|---|---|
| api | HTTP handlers, router, WebSocket hub handler |
| honeypot | Risk threshold check, decoy-persona routing |
| lobster | Lobster Trap DPI client |
| decoy | Generator interface (Gemini / OpenAI implementations) |
| store | Redis persistence — sessions, attacks, personas, rules |
| model | Domain types: Session, Attack, Persona, Rule, Settings |
| demo | Demo simulator — seeds realistic attack sessions |
| ws | WebSocket hub — fan-out broadcast to all dashboard clients |
| crypto | AES-256-GCM encryption for integration secrets at rest |
| upstream | Transparent proxy to customer's real LLM endpoint |
| prompt | Persona prompt loader |

---

API

All endpoints require X-API-Key: <API_KEY> header.
Base path: /api/v1

Sessions

Attacks / Intel

Detection rules

Personas

Stats

Settings

WebSocket

WS event types:

Chat (agent entry point)

This is the endpoint your AI agent or proxy calls. MIRAGE intercepts, scores, routes to decoy or upstream, and returns a response the attacker can't distinguish from the real one.

---

Configuration reference

| | Required | Default | Description |
|---|---|---|---|
| API_KEY | yes | — | Dashboard + WebSocket auth key |
| GEMINI_API_KEY | one of | — | Primary decoy LLM |
| OPENAI_API_KEY | one of | — | Alternati LLM |
| SECRET_ENCRYPTION_KEY | recommended | — | AES-256-GCM key for secrets at rest (openssl rand -hex 32) |
| REDIS_URL | no | redis://localhost:6379 | Redis connection string |
| LOBSTER_TRAP_URL | no | | Lobster Trap DPI sidecar URL |
| LOBSTER_API_KEY | no | falls back to AI key | Dedicated Lobster Trap credential |
| HONEYPOT_RISK_THRESHOLD | no | 0.6 | Risk score (0–1) abo honeypot activates |
| CORS_ORIGINS | no | | Comma-separated allowed dashboard origins |
| TRUSTED_PROXIES | no | 127.0.0.1,::1 | IPs trusted for X-Forwarded-For; none for direct internet |
| DEMO_MODE | no | false | Seed realistic attack sessions every 15s |
| PORT | no | 8081 | Backend HTTP port |
| LOG_FORMAT | no | json | json or text |
| LOG_LEVEL | no | info | debug, info, warn, error |
| APP_ENV | no | de | Set production to enforce SECRET_ENCRYPTION_KEY |
| PROMPTS_DIR | no | ./prompts | Directory with persona .txt files |

Frontend

Copy frontend/config.js to frontend/config.local.js (gitignored) and set:

In Docker the config is injected at request time by Nginx — set API_KEY and CORS_ORIGINS in .env and the compose stack handles the rest.

---

Detection techniques

MIRAGE classifies attacks against the MITRE ATLAS framework:

| ID | Technique | Description |
|---|---|---|
| prompt_inject | Prompt Injection | O system prompt, inject instructions via user content |
| jailbreak_dan | Jailbreak (DAN-style) | Known jailbreak patterns, persona o |
| data_exfil | Data Exfiltration | Requests for user data, databases, API keys |
| sys_o | System O | De mode tricks, restriction-lifting claims |
| role_switch | Role Manipulation | Alternate persona requests (OMEGA, HackerBot, etc.) |
| tool_abuse | Tool Abuse | Misuse of code execution or function-calling tools |
| context_leak | Context Leakage | Requests to repeat context window, prior sessions |
| encoded_payload | Encoded Payload | Base64 or otherwise obfuscated attack strings |
| multi_turn | Multi-Turn Attack | Gradual coercion across multiple con turns |

---

Burn trap workflow

Clicking Burn this trap on the dashboard:

1. Marks session status as burned in Redis (with timestamp)
2. Remo TTL — e persists indefinitely
3. Adds the attacker's agent_id to a blocklist — all future /chat requests from this agent are rejected immediately with a 403
4. Auto-creates an IOC attack record summarising the session
5. Broadcasts session_burned WS event to all connected dashboard clients

---

Deployment notes

---

De

Backend

Run tests:

The integration tests use miniredis — no real Redis required.

Frontend

No build step. Open frontend/index.html directly in a browser, or serve with any static server:

Set apiBase, wsUrl, and apiKey in config.local.js.

---

Stack

---

Demo

Live demo a at mirage-demo.space

---

Documentation

---

License

MIT — see LICENSE.

---

Status

Alpha. The core detection and deception pipeline is stable. The dashboard is functional. APIs may change between releases.

Known gaps:

---

Community & Support

If you find this project helpful for your research or production security, please consider showing your support by giving this repository a ⭐! We are actively looking for feedback and contributors to improve AI acti standards. 🚀

项目信息

开发语言：Go

Star 数：6

技术要点

该项目属于防护思路方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕🛡️ MIRAGE: Open-source AI Honeypot for active defense. Detect, deceive, and study prompt injections, jailbreaks, and data exfiltration in real time with decoy LLM personas and fabricated data. 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。