MIRAGE
AI honeypot that deceives attackers instead of blocking them
Start new thread
Victoria Fonareva

MIRAGE - AI honeypot that deceives attackers instead of blocking them

by•
Most AI security tools silently block attacks. MIRAGE does the opposite — when it detects a prompt injection, jailbreak, or data exfiltration attempt, it routes the attacker to a convincing decoy persona that feeds them fabricated data and honey tokens. Nothing real leaks. Every move is logged and tagged with MITRE ATLAS techniques for threat intelligence. Built with Go + Redis + Gemini/OpenAI. Docker Compose, no build step. Demo mode works without any API keys.

Add a comment

Replies

Best
Victoria Fonareva
Maker
📌
Hey PH! 👋 I built MIRAGE at a hackathon focused on AI agent security. The idea came from a simple question: why do we only block attacks? Traditional honeypots deceive attackers to study them — why not apply that to AI? So instead of dropping suspicious requests, MIRAGE spins up a fake AI persona that acts convincingly real, serves fabricated data, and logs everything. The attacker thinks they're succeeding. They're not. Demo mode runs without any API keys — just docker compose up and you'll see live simulated attack sessions. Feedback very welcome, still alpha!