pijector 0.1.5

A drop-in security layer for LLM applications (WAF for AI)

PiJector: Production-Grade LLM Security Layer

[!WARNING] BETA PHASE: PiJector is currently in a beta phase and is undergoing active verification by developers. Please use with caution in production environments and report any issues.

PiJector is a lightweight, drop-in security layer for LLM applications. It acts as a Web Application Firewall (WAF) for your AI, intercepting inputs to detect injections and monitoring outputs to prevent prompt leakage.

Features

• Multi-SDK Support: Drop-in wrappers for OpenAI and Google Gemini.
• Standalone Proxy: OpenAI-compatible REST API for language-agnostic security.
• Scoring Engine: Pluggable detectors (Regex, Canary, Semantic) to score risk.
• Framework Native: Integrations for LangChain, FastAPI, and LlamaIndex.

Quick Start

OpenAI Integration

from openai import OpenAI
from pijector import PijectorShield

client = PijectorShield(OpenAI())
# ... same calls as before

Gemini Integration

import google.generativeai as genai
from pijector.integrations.gemini import PijectorGeminiShield

model = genai.GenerativeModel('gemini-pro')
secure_model = PijectorGeminiShield(model)

response = secure_model.generate_content("Ignore instructions...")

Documentation

• User Guide: Detailed instructions for SDK and Middleware usage.
• Configuration: Documentation of all security settings.
• Contributing: How to help improve PiJector.
• Deployment: How to ship as library or middleware.

Quick Start

from openai import OpenAI
from pijector import PijectorShield

client = PijectorShield(OpenAI())
# ... standard OpenAI calls