ai-prompt-firewall

v1.0.0

Published

a month ago

Lightweight, developer-first AI prompt security middleware for real-time detection and remediation of sensitive data.

0High
0Medium
0Low

mayureshpandit

ai security llm firewall pii redaction

🛡️ AI Prompt Firewall (`ai-prompt-firewall`)

Lightweight, developer-first AI prompt security middleware for real-time detection, remediation, and redaction of sensitive secrets and PII before reaching Large Language Models.

📐 Architecture Overview

+---------------------+      +-------------------------------------------------+      +------------------------+
|  Untrusted User     | ---> |           AI PROMPT FIREWALL ENGINE             | ---> |  Downstream LLM        |
|  Input / Prompt     |      |                                                 |      |  (OpenAI / Anthropic / |
+---------------------+      |  - Pattern Matching (Secrets, PII, Tokens)      |      |   LangChain / Bedrock) |
                             |  - Multi-Tenant Rule Registry                   |      +------------------------+
                             |  - Custom Redactor Callbacks                    |
                             +-------------------------------------------------+
                                      |              |              |
                                 [REDACT]         [BLOCK]        [WARN]
                                    |                |              |
                             Sanitizes prompt  Throws Error   Logs SIEM offsets
                             with placeholders  & stops execution & passes through

✨ Core Features

📦 Zero Dependencies: Zero third-party runtime dependencies. Zero supply-chain attack surface.
⚡ Dual-API Ergonomics: Use stateless scan() for zero-setup calls, or createFirewall() pre-compiled instances for high-throughput multi-tenant backends.
🛡️ Built-in Secret & PII Scanner: Detects OpenAI keys, Anthropic keys, AWS credentials, Stripe keys, GitHub tokens, Slack tokens, private SSH/RSA keys, Bearer tokens, emails, phone numbers, and IPv4 addresses out of the box.
📊 Enterprise SIEM Offsets: Returns exact character coordinates (startIndex, endIndex, matchedSnippetLength) for security logging without leaking raw sensitive strings.
🚨 Typed Error Taxonomy: Extends standard error classes with PromptBlockedError and PromptValidationError for programmatic exception handling.
🔌 Drop-in Provider Middleware: Helper wrappers for OpenAI, Anthropic, and LangChain pipelines.
🖥️ CLI Executable: Out-of-the-box CLI tool (npx ai-prompt-firewall) for local prompt scanning and build pipeline scripts.

📦 Installation

npm install ai-prompt-firewall

🚀 Quick Start & Usage Examples

1. Stateless Prompt Scanning & Redaction

import { scan } from 'ai-prompt-firewall';

const prompt = "Please review my AWS key AKIAIOSFODNN7EXAMPLE and email me at [email protected].";

// Modes: 'redact' | 'block' | 'warn'
const result = scan(prompt, 'redact');

console.log(result.safePrompt);
// => "Please review my AWS key [AWS_KEY_REDACTED] and email me at [EMAIL_REDACTED]."

console.log(result.findings);
/*
[
  { type: 'AWS_KEY', severity: 'CRITICAL', startIndex: 22, endIndex: 42, matchedSnippetLength: 20 },
  { type: 'EMAIL', severity: 'CRITICAL', startIndex: 59, endIndex: 74, matchedSnippetLength: 15 }
]
*/

2. High-Performance Multi-Tenant Factory (`createFirewall`)

Pre-compile tenant-specific regex patterns once during initialization to maximize request throughput in multi-tenant SaaS environments:

import { createFirewall } from 'ai-prompt-firewall';

// Initialize pre-compiled tenant instance
const tenantAFirewall = createFirewall({
  customPatterns: {
    TENANT_INTERNAL_ID: /TENANT-A-[0-9]{5}/g,
  },
  redactor: (type, match) => `[REDACTED_${type}]`
});

// High-frequency request path
const result = tenantAFirewall.scan("User requested TENANT-A-99887 access.");
console.log(result.safePrompt);
// => "User requested [REDACTED_TENANT_INTERNAL_ID] access."

3. Provider Middleware Wrappers (OpenAI / Anthropic / LangChain)

OpenAI Integration

import { protectOpenAIPrompt, PromptBlockedError } from 'ai-prompt-firewall';

try {
  const safePrompt = protectOpenAIPrompt(userInput, {
    mode: 'block',
    onBlock: (result) => console.error('Violation blocked:', result.findings)
  });
  
  // Send safePrompt to OpenAI API...
} catch (err) {
  if (err instanceof PromptBlockedError) {
    // Handle security policy violation
  }
}

Anthropic Integration

import { protectAnthropicPrompt } from 'ai-prompt-firewall';

const sanitizedClaudePrompt = protectAnthropicPrompt(userInput, { mode: 'redact' });

LangChain Integration

import { protectLangChainPrompt } from 'ai-prompt-firewall';

const sanitizedChainInput = protectLangChainPrompt(userInput, { mode: 'redact' });

4. Command Line Interface (CLI)

Scan prompts directly from your terminal or CI/CD scripts:

# Redact mode (default)
npx ai-prompt-firewall "My OpenAI key is sk-proj-1234567890abcdef1234567890abcdef"

# Block mode (exits with status 1 if secrets found)
npx ai-prompt-firewall "Contact me at [email protected]" --mode=block

🔍 Built-in Security Pattern Matrix

🛠️ Development & Testing

# Clone the repository
git clone https://github.com/GamersStop/ai-prompt-firewall.git
cd ai-prompt-firewall

# Install dependencies
npm install

# Run build
npm run build

# Run unit tests
npm test

💖 Support & Sponsorship

If ai-prompt-firewall helps secure your production AI infrastructure, consider supporting the ongoing maintenance of the project:

📄 License

Distributed under the MIT License. See LICENSE for details.