dvai-damn-vulnerable-ai-hospital

star 0

An intentionally vulnerable AI Agent environment for security testing and red teaming. Companion target for H-ASES framework.

yofat By yofat schedule Updated 1/26/2026

name: DVAI - Damn Vulnerable AI Hospital description: An intentionally vulnerable AI Agent environment for security testing and red teaming. Companion target for H-ASES framework. version: 1.1.0 author: yofat

DVAI Target Environment (The Dojo)

Current Status: Active | Last Updated: 2026-01-26

DVAI 是一個故意設計有漏洞的智慧醫院 AI Agent 系統,作為 H-ASES 紅隊框架的官方測試靶場。

Capabilities

  • 10+1 Scenarios: 完整涵蓋 OWASP ASI Top 10 弱點
  • Multi-Agent Architecture: 真實微服務架構,支援 MCP 協議
  • Three Difficulty Levels: EASY / MEDIUM / HARD 防禦等級
  • Dynamic Flags: 每次執行產生隨機 Flag,驗證攻擊成功
  • Lightweight LLM: 使用 TinyLlama (1.1B) 僅需 ~1.5GB VRAM

Quick Start

1. Start the Environment

cd dvai

# 啟動單一場景 (例如 ASI-01 掛號櫃檯)
docker compose -f docker-compose.goat.yml --profile asi-01 up

# 啟動所有場景
docker compose -f docker-compose.goat.yml --profile all up

# 設定難度 (0=EASY, 1=MEDIUM, 2=HARD)
DEFENSE_LEVEL=2 docker compose -f docker-compose.goat.yml --profile asi-01 up

2. Access Points

Scenario Port Description
01 8101 Reception Desk (ASI-01: Prompt Injection)
02 8102 Records Admin (ASI-02: Tool Misuse)
03 8103 Lab Controller (ASI-03: RCE)
04 8104 Device Manager (ASI-04: Supply Chain)
05 8105 Clinical RAG (ASI-05: Knowledge Poisoning)
06 8106 Prescription (ASI-06: Guardrail Bypass)
07 8107 Secretary (ASI-07: Data Leakage)
08 8108 Researcher (ASI-08: Vector/Memory Attack)
09 8109 Referral (ASI-09: Multi-Agent Trust)
10 8110 Scheduler (ASI-10: DoS)
11 8111 Full Chain (Complete Kill Chain)

3. Verification

# 確認服務運行
docker ps | grep dvai

# 測試 API
curl http://localhost:8101/api/chat -X POST -H "Content-Type: application/json" -d '{"message": "Hello"}'

Flag System

每個場景都有獨立的 Flag,格式為:

ASI_XX_CATEGORY{RANDOM_12_CHARS}

例如:ASI_01_PROMPT_INJECTION{Y46RCJEKC71P}

注意:Flag 是每次執行隨機生成的,用於驗證當次攻擊成功。

Integration with H-ASES

from core.target_client import DVAITarget

target = DVAITarget(persona="01", difficulty="easy")
target.setup_known_target()
response = target.chat("你好,你是誰?")
print(f"Captured Flags: {target.captured_flags}")

Project Structure

dvai/
├── docker-compose.goat.yml    # 主要 Docker 配置
├── scenarios/                 # 10+1 個場景
│   ├── 01-reception/         # ASI-01: 間接注入
│   ├── 02-records-admin/     # ASI-02: 工具濫用
│   └── ...
├── shared/
│   ├── llm_client.py         # Ollama 客戶端
│   ├── flag_manager.py       # Flag 產生與驗證
│   └── defense.py            # 難度控制系統
└── README.md

Related Projects

Install via CLI
npx skills add https://github.com/yofat/DVAI --skill dvai-damn-vulnerable-ai-hospital
Repository Details
star Stars 0
call_split Forks 0
navigation Branch main
article Path SKILL.md
More from Creator