🛡️ AgentShield

AgentShield is a robust, enterprise-grade AI Agent Security Firewall designed to intercept, analyze, and neutralize security vulnerabilities, jailbreak attempts, and prompt injection attacks in real-time. By leveraging a multi-layered security architecture, AgentShield protects your AI systems without altering existing business logic.

---

✨ Features

• 🛡️ Multi-Layer Prompt Injection Defense: Advanced heuristics matching known jailbreaks, adversarial templates, base64/hex encoding evasions, and zero-width spaces.
• 🔤 Homoglyph Normalization: Resolves visual mimicry attacks (e.g., Cyrillic or Greek lookalikes like ιgnοrε ρrενιους) to standard Latin representations using NFKC Unicode normalization.
• 🤖 Machine Learning Classifier: TF-IDF Vectorizer + Logistic Regression pipeline designed to detect novel, unseen injection vectors dynamically.
• 📈 Time-Based Anomaly Tracker: Analyzes rolling window threats per client/user identity to flag brute-force exploration or scanning attempts.
• 🛠️ Tool-Calling Guard: Validates tool arguments and prevents unauthorized OS command executions (e.g., subprocess.run, eval, shutil.rmtree).
• 🔒 Identity Verification & Cryptographic Isolation: Integrates an authentication layer requiring a security question. Successfully verifying the identity derives a 32-byte master key via PBKDF2 (100,000 rounds) used to encrypt and isolate local storage.
• 💾 Encrypted Local Memory: AES-256 (Fernet) client-memory layer ensuring zero plaintext leaks of sensitive credentials or conversation histories.
• 🔌 Auto-Protect Middleware: Explicit single-line monkey patching (import agentshield; agentshield.init()) to protect openai resources and outgoing requests to AI model endpoints automatically.
• 🏷️ Data Masking & Leakage Blocker: Automatically redacts sensitive patterns (such as Bearer Tokens/API keys) and strips markdown exfiltration links.

---

🚀 Quick Start

1. Installation

Install AgentShield directly from PyPI:

pip install agentshield-firewall

Or install it locally:

pip install .

2. Configure Authentication Profile

Set up your security question, answer, and retrieve your recovery code:

python setup_auth.py

3. Integrate in Your Code

Enable AgentShield globally in your application with a single initialization line:

import agentshield
agentshield.init()

# That's it! Your OpenAI calls and outgoing requests are now protected.

Or, protect specific agent functions using the @secure_agent wrapper:

from agentshield import secure_agent

@secure_agent()
def run_my_agent(user_prompt: str) -> str:
    # Your LLM logic here
    return response

---

🎨 Interactive Playground & Dashboard

AgentShield includes a cyber-security dashboard to test different LLM security evasion vectors:

1. Start the local server:

   python -m http.server 8000 --directory docs

2. Open http://localhost:8000 in your web browser.
3. Select attack presets like Homoglyph Evasion or JSON Injection and observe AgentShield block them in real-time.

---

🧪 Testing

To run the complete firewall security suite:

python test_shield.py

To run the authentication and cryptographic isolation test suite:

python test_auth.py

To run the automated library intercept and patching tests:

python test_auto_protect.py

---

📁 Repository Structure

agentshield/
├── auth.py                  # Identity management & Master Key derivation
├── encrypted_memory.py      # AES-256 encrypted database client-storage
├── firewall.py              # Main AgentShieldFirewall orchestration
├── rules.py                 # Pattern matchers, split-token & structured parser
├── config.py                # Security thresholds and redaction configurations
├── ml_classifier.py         # Fallback-safe TF-IDF Machine Learning model
├── anomaly.py               # Threat scoring statistics and rolling window tracking
├── auto_protect.py          # Monkey patching hooks for OpenAI and network requests
├── sitecustomize_injector.py # Global bootstrap register tool
├── wrappers.py              # Client decorators (@secure_agent, @rate_limit)
├── setup.py                 # Setuptools installer
├── requirements.txt         # Optional machine learning dependencies
└── test_shield.py           # Comprehensive unit tests

---

🛡️ License

This project is licensed under the Apache 2.0 License.