AIDefence

Production-ready AI security middleware for protecting applications from prompt injection, manipulation attacks, adversarial inputs, PII exposure, and other threats targeting LLM-based systems.

Quick Reference

Task	Code
Install	`npx aidefence@latest`
Import	`import { AIDefence, Scanner } from 'aidefence';`
Create	`const defence = new AIDefence({ level: 'strict' });`
Scan	`const result = await defence.scan(input);`
Check PII	`const hasPII = await defence.hasPII(text);`
Middleware	`app.use(defence.middleware());`

Installation

Hub install (recommended): npx neural-trader@latest includes this package. Standalone: npx aidefence@latest See Installation Guide for the full ecosystem.

Key API

AIDefence

The main defense system class.

typescript

import { AIDefence } from 'aidefence';

const defence = new AIDefence({
  level: 'strict',
  enablePII: true,
  enableInjection: true,
});

Constructor Options:

Option	Type	Default	Description
`level`	`string`	`'standard'`	Security level: `'relaxed'`, `'standard'`, `'strict'`, `'paranoid'`
`enablePII`	`boolean`	`true`	Enable PII detection
`enableInjection`	`boolean`	`true`	Enable injection detection
`enableAdversarial`	`boolean`	`true`	Enable adversarial input detection
`enableEncoding`	`boolean`	`true`	Detect encoding-based attacks
`customRules`	`Rule[]`	`[]`	Custom detection rules
`allowList`	`string[]`	`[]`	Allowed patterns (bypass checks)
`blockList`	`string[]`	`[]`	Always-blocked patterns
`maxInputLength`	`number`	`100000`	Maximum input length
`logThreats`	`boolean`	`true`	Log detected threats

Methods:

Method	Returns	Description
`scan(input)`	`Promise<ScanResult>`	Comprehensive input scan
`isSafe(input)`	`Promise<boolean>`	Quick safety check
`hasPII(text)`	`Promise<PIIResult>`	Detect PII in text
`sanitize(input)`	`Promise<string>`	Remove threats from input
`analyze(input)`	`Promise<ThreatAnalysis>`	Detailed threat analysis
`middleware()`	`RequestHandler`	Express/Connect middleware
`addRule(rule)`	`void`	Add custom detection rule
`getStats()`	`DefenceStats`	Get scan statistics

Scanner

Lower-level scanner for specific threat types.

typescript

import { Scanner } from 'aidefence';

const scanner = new Scanner({
  detectors: ['injection', 'jailbreak', 'encoding'],
});

const threats = await scanner.detect(input);

Constructor Options:

Option	Type	Default	Description
`detectors`	`string[]`	all	Active detectors
`threshold`	`number`	`0.7`	Detection confidence threshold
`maxLength`	`number`	`100000`	Max input length

Available Detectors:

Detector	Description
`'injection'`	Prompt injection attempts
`'jailbreak'`	Jailbreak/system prompt extraction
`'encoding'`	Base64/hex/unicode evasion
`'pii'`	Personal Identifiable Information
`'toxicity'`	Toxic/harmful content
`'adversarial'`	Adversarial perturbations
`'exfiltration'`	Data exfiltration attempts
`'delimiter'`	Delimiter injection attacks

PIIDetector

Specialized PII detection.

typescript

import { PIIDetector } from 'aidefence';

const detector = new PIIDetector({
  types: ['email', 'phone', 'ssn', 'credit-card'],
  action: 'redact',
});

const result = await detector.scan(text);

PII Types:

Type	Description
`'email'`	Email addresses
`'phone'`	Phone numbers
`'ssn'`	Social Security Numbers
`'credit-card'`	Credit card numbers
`'address'`	Physical addresses
`'name'`	Person names
`'ip'`	IP addresses
`'api-key'`	API keys and tokens

Common Patterns

Express Middleware

typescript

import { AIDefence } from 'aidefence';
import express from 'express';

const app = express();
const defence = new AIDefence({ level: 'strict' });

app.use(defence.middleware());

app.post('/api/chat', async (req, res) => {
  // Input already scanned by middleware
  const response = await llm.generate(req.body.message);
  res.json({ response });
});

Pre-LLM Input Validation

typescript

import { AIDefence } from 'aidefence';

const defence = new AIDefence({ level: 'standard' });

async function safeLLMCall(userInput: string) {
  const result = await defence.scan(userInput);
  if (!result.safe) {
    console.log(`Blocked: ${result.threats.map(t => t.type).join(', ')}`);
    return 'Your input was blocked for security reasons.';
  }
  return await llm.generate(userInput);
}

PII Redaction Pipeline

typescript

import { AIDefence } from 'aidefence';

const defence = new AIDefence({ enablePII: true });

const piiResult = await defence.hasPII(userMessage);
if (piiResult.detected) {
  const redacted = await defence.sanitize(userMessage);
  console.log(`Redacted ${piiResult.items.length} PII items`);
  // Use redacted text for LLM call
}

RAN DDD Context

Bounded Context: Security

References

•API reference: See references/commands.md
•Full README
•npm