📌 Overview

An open-source, AI-powered agentic honeypot that detects scam intent, autonomously engages scammers, extracts high-value intelligence, and reports structured results — without ever revealing detection.

Online scams are no longer static scripts. Modern fraudsters adapt dynamically, manipulate victims over multiple turns, and exploit trust, urgency, and emotion. Agentic Honey-Pot is designed to fight back using the same weapon scammers rely on: adaptive conversation intelligence.

Unlike rule-based filters or single-turn classifiers, this system deploys an autonomous AI agent that:

• Detects scam intent early
• Assumes a believable human persona
• Engages scammers across multi-turn conversations
• Extracts actionable intelligence in real time
• Produces structured, machine-readable threat data

• Scam research & threat intelligence
• Fraud prevention pipelines
• Security analytics platforms
• AI agent demonstrations & hackathons
• Academic / applied security research

Digital fraud (bank scams, UPI fraud, phishing, fake offers, impersonation) evolves faster than traditional defenses.

Conventional systems fail because they:

• Rely on static rules or keyword matching
• Break down in multi-turn conversations
• Cannot reason about manipulation strategies
• Are easily probed and bypassed by scammers

This project introduces an agentic honeypot that actively engages scammers, adapts in real time, and converts deception attempts into intelligence.

• Detect scam or fraudulent intent
• Activate an autonomous AI agent upon detection
• Maintain a consistent, human-like conversational persona
• Handle long, multi-turn interactions
• Extract structured scam intelligence
• Expose a secure public REST API
• Authenticate requests using API keys

graph TD
    A[Incoming Message] --> B[Scam Intent Detection]
    B -->|Positive| C[Agent Activation]
    C --> D[Persona & Memory Manager]
    D --> E[Multi-Turn Engagement]
    E --> F[Intelligence Extraction Engine]
    F --> G[Structured API Response]
    G --> H[Final Callback ]

• Non-revealing: The agent never exposes detection or intent
• Adaptive: Responses evolve based on scammer behavior
• Deceptive-by-design: Engagement appears natural and human
• Extractive-first: Every interaction maximizes intelligence gain

• Python 3.9+
• Git
• API key (environment variable)

git clone https://github.com/ksinghs/agentic-honeypot.git
cd agentic-honeypot
pip install -r requirements.txt

uvicorn app.main:app --reload

POST /api/honeypot/message

x-api-key: YOUR_SECRET_API_KEY
Content-Type: application/json

{
  "session_id": "abc-123",
  "message": "Your account is blocked. Click here to verify"
}

{
  "agent_reply": "Oh no, that sounds serious. Can you tell me which account this is about?",
  "scam_confidence": 0.94,
  "extracted_indicators": []
}

The AI agent is designed to:

• Sustain multi-turn conversations
• Adapt language, tone, and strategy dynamically
• Preserve a consistent persona
• Self-correct when inconsistencies arise
• Avoid exposing detection or analysis
• Maximize intelligence extraction before disengagement

The system extracts and normalizes:

• 🏦 Bank account numbers
• 💳 UPI IDs
• 🔗 Phishing URLs
• 📞 Phone numbers
• ⚠️ Scam-related keywords & phrases
• 🧠 Behavioral and manipulation patterns

All intelligence is returned in structured JSON with evidence traces.

• Scam intent detection
• Multi-turn agent engagement
• Intelligence extraction pipeline
• Intelligence analytics dashboard
• Multi-language support
• Scam network correlation
• Threat actor profiling

Contributions are welcome.

• Fork the repository
• Create a feature branch
• Commit changes with clear messages
• Submit a pull request

Security-related contributions should include threat models where applicable.

This project is licensed under the MIT License.

Kundan Singh AI • Data • Security • Agentic Systems