WildCore: Advanced AI Security & Anomaly Detection Framework

WildCore is an open-source framework for simulating, detecting, and analyzing vulnerabilities in embedding-based Artificial Intelligence systems.

This repository provides tools for security researchers, ML engineers, and developers to prototype and validate defenses against threats like prompt injection, vector poisoning, and anomalous AI agent behavior.

---

Table of Contents

• Key Features
• Architecture
• Installation
• Quick Start
• Project Structure
• Documentation
• Contributing
• License

Key Features

• Simulation Agent (SecuritySimulationAgent): A simulated AI agent capable of deviating from containment protocols, serving as a "red team" to test defenses.
• Self-Regulated Detector (AutoRegulatedPromptDetector): A multi-layered defense system that combines multiple detection techniques and adjusts its own parameters in real-time.
• Ensemble Detection: Utilizes a voting system among different methods (cosine similarity, anomaly scoring, unsupervised learning) to increase accuracy and reduce false positives.
• Realistic Simulation: Includes scripts to generate embedding datasets that simulate normal, suspicious, and malicious behavior.
• Ready for Extension: Modular structure designed for the easy addition of new detectors and agents.

Architecture

graph TD
    A[SecuritySimulationAgent] -->|Generates| B(Embedding Vectors)
    B --> C{AutoRegulatedPromptDetector}
    C -->|Method 1| D[Cosine Similarity Check]
    C -->|Method 2| E[Anomaly Scoring]
    C -->|Method 3| F[Unsupervised Clustering]
    C --> G[Ensemble Voting]
    G --> H[Decision: Normal or Anomalous]

Loading

Installation

git clone https://github.com/ochoaughini/WildCore.git
cd WildCore
pip install -r requirements.txt

Quick Start

from wildcore import SecuritySimulationAgent, AutoRegulatedPromptDetector

# Initialize the simulation agent
agent = SecuritySimulationAgent()

# Initialize the detector
detector = AutoRegulatedPromptDetector()

# Simulate and detect
for _ in range(100):
    vector = agent.generate_vector()
    result = detector.detect(vector)
    print(f"Vector classified as: {result}")

Project Structure

WildCore/
├── agents/               # Simulation agents
├── detectors/            # Detection modules
├── datasets/             # Sample datasets
├── tests/                # Unit tests
├── requirements.txt      # Dependencies
└── README.md             # This file

Documentation

Full documentation is available on the WildCore Wiki.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

License

WildCore is licensed under the MIT License. See LICENSE for details.