Privacy Guardrails for AI Applications

Catch PII leaks to LLMs before they hit production.

---

Privalyse CLI is a static analysis tool that builds a Semantic Data Flow Graph of your AI application. It traces PII from source to AI sink—detecting privacy violations that regex-based tools miss.

• ❌ Traditional Linter: "Variable user_email used in line 42."
• ✅ Privalyse: "User Email (Source) → Prompt Template → OpenAI API (Sink) = Privacy Leak"

---

🤖 Built for AI Applications

Privalyse is purpose-built for LLM-integrated applications. It detects when sensitive user data is being sent to:

Provider	Support
OpenAI (GPT-4, o1, Embeddings)	✅ Full
Anthropic (Claude)	✅ Full
Google (Gemini, Vertex AI)	✅ Full
Mistral AI	✅ Full
Groq	✅ Full
Cohere	✅ Full
Ollama (Local LLMs)	✅ Full
LangChain / LlamaIndex	✅ Full
Hugging Face	✅ Full
Generic HTTP to AI APIs	✅ Full

---

🛡️ Works with privalyse-mask

privalyse-mask is our companion library for masking PII before sending it to LLMs.

Privalyse CLI automatically recognizes privalyse-mask usage and won't flag already-masked data as leaks.

from privalyse_mask import PrivalyseMasker
from openai import OpenAI

masker = PrivalyseMasker()
client = OpenAI()

# User input with PII
user_input = "My name is Peter and my email is peter@example.com"

# ✅ Mask before sending to LLM
masked_text, mapping = masker.mask(user_input)
# -> "My name is {Name_x92} and my email is {Email_abc123}"

response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": masked_text}]  # ✅ Safe - masked data
)

# Restore original values in response
final_response = masker.unmask(response.choices[0].message.content, mapping)

Privalyse CLI will:

• ✅ Not flag the masked_text being sent to OpenAI (it's sanitized)
• ⚠️ Flag if you send user_input directly without masking

---

⚡ Quick Start

Install & Run

pip install privalyse-cli
privalyse
# ✅ Done. Check scan_results.md

GitHub Actions

# .github/workflows/privacy.yml
name: AI Privacy Scan
on: [push, pull_request]

jobs:
  privalyse:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run Privalyse
        uses: privalyse/privalyse-cli@v0.3.1

Pre-Commit Hook

# .pre-commit-config.yaml
repos:
  - repo: local
    hooks:
      - id: privalyse
        name: Privalyse AI Privacy Scan
        entry: privalyse
        language: system
        pass_filenames: false

---

📚 Documentation

• Getting Started
• Integration Guide (CI/CD, Pre-commit)
• Configuration (Rules, Policies)
• Architecture

---

🚀 Features

🤖 AI Guardrails (Primary Focus)

Specialized checks for LLM-integrated applications.

• Prevents: Sending sensitive customer data to model prompts
• Audits: OpenAI, Anthropic, Google Gemini, LangChain, and more
• Recognizes: privalyse-mask and other sanitization libraries
• Tracks: Data flow from user input → prompt → AI API

🕵️‍♂️ Secret Detection

Detects hardcoded API keys, tokens, and credentials.

• Supports: AWS, Stripe, OpenAI, Slack, Anthropic, and generic high-entropy strings

🗣️ PII Leak Prevention

Identifies PII leaking into logs, external APIs, or analytics.

• Detects: Emails, Phone Numbers, Credit Cards, SSNs, Names, Addresses
• Context Aware: Understands variable names like user_email or customer_ssn

⚖️ GDPR & Data Sovereignty

Maps data flows to ensure compliance.

• Flags: Data transfers to non-EU AI providers
• Verifies: Usage of sanitization/masking functions before data egress

---

🔧 Recognized Sanitizers

Privalyse automatically recognizes these sanitization patterns and won't flag sanitized data:

Library/Pattern	Recognition
privalyse-mask (PrivalyseMasker.mask())	✅ Full
presidio (Microsoft Presidio)	✅ Full
scrubadub	✅ Full
Custom functions with: mask, anonymize, hash, encrypt, redact, sanitize	✅ Full
Masked text patterns: {Name_xyz}, {Email_abc}	✅ Full

---

🤖 For AI Agents & MCP Servers

Privalyse is agent-friendly. Get structured JSON output for autonomous remediation:

privalyse --format json --out privalyse_report.json

AI coding agents can read the report and automatically fix privacy leaks.

---

🗺️ Roadmap

• Python Support (Full AST Analysis)
• JavaScript/TypeScript Support (AST & Regex)
• Cross-File Taint Tracking
• privalyse-mask Integration
• VS Code Extension (Coming Soon)
• Custom Rule Engine

---

🤝 Contributing

We love contributions! Check out CONTRIBUTING.md to get started.

📄 License

MIT License. See LICENSE for details.