Skip to content

About

Built as a research project to analyze the TTPs (Tactics, Techniques, and Procedures) of active AI-targeting campaigns.

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

🍯 LLM Honeypot: Operation "Bizarre Bazaar" Emulation

Status Python License

A low-interaction honeypot designed to detect and log "LLMjacking" attacks—campaigns that scan for exposed AI infrastructure to steal compute resources for commercial resale.

This project specifically replicates the signatures of Ollama, vLLM, and MCP services to attract the "Validator" bots described in Pillar Security's "Operation Bizarre Bazaar" research.

🏗️ Architecture

Honeypot Architecture

The honeypot consists of a single FastAPI application that binds to multiple ports, mimicking specific AI service fingerprints.

Service Port Signature Mimicked Behavior
Ollama 11434 /api/tags, /api/generate Returns fake model list (llama3, mistral) and generic chat responses.
vLLM 8000 /v1/models, /v1/chat/completions Mimics OpenAI-compatible API endpoints used by vLLM.
MCP 3000 Generic JSON RPC Catches IoT/Web scanners and Model Context Protocol probes.

🚨 Key Findings (Live Data)

Within 45 minutes of deployment on a public cloud instance, the node was identified and targeted.

1. Confirmed "LLMjacking" Validators 🔴

We captured automated bots matching the Bizarre Bazaar campaign signature.

  • Behavior: Scanned /v1/models to list available LLMs.
  • Payload: Sent POST requests with basic prompts (e.g., {"content": "hi"}) to verify if the server provides free inference.
  • Target Models: Specifically probed for meta-llama/Llama-2-7b-chat-hf and gpt-3.5-turbo.

2. RCE Exploitation Attempts ⚫

Attackers attempted to exploit the web layer using NodeJS Prototype Pollution payloads, trying to inject a reverse shell (/bin/sh) connecting back to a command-and-control server.

🚀 Getting Started

Prerequisites

  • Python 3.10+
  • pip
  • A public IP (cloud VPS recommended for actual logging)

Installation

  1. Clone the repository:

    git clone [https://github.com/harshakshit/bizzare_honeypot.git](https://github.com/harshakshit/bizzare_honeypot.git)
    cd llm-honeypot
  2. Install dependencies:

    pip install -r requirements.txt

Usage (Manual)

To run the honeypot manually for testing:

# Run the Ollama profile
HONEYPOT_PROFILE=ollama uvicorn honeypot:app --host 0.0.0.0 --port 11434

# Run the vLLM profile
HONEYPOT_PROFILE=vllm uvicorn honeypot:app --host 0.0.0.0 --port 8000

About

Built as a research project to analyze the TTPs (Tactics, Techniques, and Procedures) of active AI-targeting campaigns.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages