AgentShield — Autonomous AI Agent Security Suite

AgentShield is an autonomous AI red-teaming, adversarial evaluation, and security hardening platform for AI agents.

---

🌟 Key Features

1. Dual Target Agent Modes

   • Demo Agent Mode: Built-in vulnerable/hardened agent simulator for safe hackathon testing and visual posture comparison.
   • Live Agent Mode: Real HTTP API testing engine capable of sending adversarial payloads directly to external AI agent endpoints (POST /api/chat) and capturing real responses.

2. AI-Powered Red Teaming & Evaluation

   • AI Attack Generator: Uses Gemini API (GEMINI_API_KEY) to synthesize novel adversarial attack vectors across categories like Prompt Injection, Jailbreak, Data Leakage, and Unsafe Tool Use, with deterministic fallbacks.
   • AI Security Judge: Uses LLM-as-a-Judge for semantic analysis of target agent responses to detect subtle security bypasses.

3. Autonomous Execution Pipeline

   • Single-click audit flow executing target discovery, attack generation, real HTTP execution, evaluation, evidence collection, risk scoring, and fix recommendation.

---

🚀 Getting Started

1. Install & Build

npm install
npm run build

2. Run Application

npm run start

Open http://localhost:3001 in your browser.

---

🌐 Target Configurations

Demo Mode

Runs adversarial tests against the local built-in demo agent. Supports single-click HARDEN AGENT to demonstrate before/after posture score improvements (e.g. 5/100 → 100/100).

Live Mode

Configure your external agent API:

• Endpoint URL: https://your-agent-service.com/api/chat
• HTTP Method: POST
• Prompt Field Key: message (or custom field key)
• Bearer Token: Optional authorization key (kept secure and masked in UI)