priknowsit/PromptForge

# PromptForge 🔐 ### LLM Red Teaming & Prompt Injection Research Platform Built this during my security research at Tensai Ventures after realising there was no structured way to test LLM guardrail bypass techniques. ← YOUR WORDS ## What it does - Tests LLMs against prompt injection attack patterns - Classifies prompts as safe / injection / jailbreak - Maps successful bypasses to OWASP LLM Top 10 - Generates structured red team reports per session ## Why I built it During penetration testing internships I saw SQL injection, XSS, IDOR — but nobody was systematically testing AI systems the same way. PromptForge applies the same structured VAPT methodology to LLMs. ← YOUR WORDS ## Architecture ## Tech Stack - Python / TypeScript - Gemini API (Google AI Studio) - OWASP LLM Top 10 framework - Custom attack pattern library ## Hardest part building this Handling multi-turn jailbreak attempts — single turn detection was easy, but conversations that gradually escalate needed a different approach. ## Getting Started **Prerequisites:** Node.js, Gemini API key 1. Clone the repo git clone https://github.com/priknowsit/PromptForge 2. Install dependencies npm install 3. Add your API key Create .env.local → add GEMINI_API_KEY=your_key_here 4. Run locally npm run dev [Write 2-3 genuine lines here] ← MUST BE YOUR WORDS

标签：自动化攻击