Introduction

As AI applications move into production, security becomes paramount. KalGuard is Infrarix's real-time AI security layer that scans, detects, and redacts sensitive information before it reaches — or leaves — your AI models.

Whether you're building a customer-facing chatbot or an internal AI assistant, KalGuard ensures compliance with GDPR, HIPAA, and SOC 2 requirements without adding significant latency to your pipeline.

The Problem

AI applications face unique security challenges that traditional WAFs and input sanitizers don't address:

Prompt injection attacks: Malicious inputs that manipulate LLM behavior
PII leakage: Users accidentally submit SSNs, credit cards, or health data
Data exfiltration: LLM responses that inadvertently expose sensitive training data
Compliance violations: Failure to redact PII before logging or storage
Jailbreak attempts: Inputs designed to bypass safety guardrails

How KalGuard Works

KalGuard operates as a transparent middleware layer in your AI pipeline with three processing stages:

1. Input Scanning

Every request is analyzed in real-time using pattern matching, ML classifiers, and contextual analysis. KalGuard identifies 50+ PII entity types including emails, phone numbers, SSNs, credit card numbers, API keys, and health records.

2. Threat Detection

Advanced classifiers detect prompt injection patterns, jailbreak attempts, and adversarial inputs. KalGuard maintains a continuously updated threat signature database.

3. Redaction & Response

Detected PII is redacted in-place with configurable replacement tokens. Threats are blocked or flagged based on your security policy. All actions are logged for audit purposes.

Quick Start

curl -X POST https://api.infrarix.com/v1/kalguard/scan \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "My SSN is 123-45-6789 and email is john@example.com",
    "actions": ["redact"],
    "entities": ["ssn", "email"]
  }'

Response:

{
  "redacted": "My SSN is [SSN_REDACTED] and email is [EMAIL_REDACTED]",
  "entities_found": [
    { "type": "ssn", "confidence": 0.99, "start": 10, "end": 21 },
    { "type": "email", "confidence": 0.98, "start": 32, "end": 48 }
  ],
  "threats": [],
  "latency_ms": 4.7
}

SDK Integration

TypeScript

import { KalGuard } from '@infrarix/kalguard'

const guard = new KalGuard({ apiKey: process.env.INFRARIX_KEY })

// Scan before sending to LLM
const result = await guard.scan(userInput, {
  actions: ['redact', 'block_threats'],
  entities: ['all'],
})

if (result.threats.length > 0) {
  return { error: 'Input blocked for security' }
}

// Safe to send redacted text to your LLM
const llmResponse = await openai.chat(result.redacted)

What KalGuard Detects

Category	Entity Types	Accuracy
Personal	Name, DOB, Address, Phone	99.2%
Financial	Credit Card, Bank Account, SSN	99.7%
Health	Medical Record, Insurance ID	98.9%
Technical	API Keys, Passwords, Tokens	99.5%
Threats	Prompt Injection, Jailbreak, XSS	99.1%

Compliance Support

KalGuard helps you meet requirements for:

GDPR: Automatic PII detection and redaction for EU data subjects
HIPAA: PHI detection and compliant handling of health information
SOC 2: Comprehensive audit logging of all security events
CCPA: Consumer data identification and deletion support
PCI DSS: Credit card data detection and masking

Performance

Metric	Value
Median scan latency	<8ms
Detection accuracy	99.7%
False positive rate	<0.1%
Entity types supported	50+
Max input size	100KB

Frequently Asked Questions

Does KalGuard store scanned data?

No. All scanning happens in-memory and data is never persisted. Audit logs contain only metadata, not the original content.

Can I customize detection rules?

Yes. Enterprise plans support custom regex patterns, entity definitions, and threat signatures.

Does KalGuard work with streaming responses?

Yes. KalGuard supports streaming scan mode for real-time redaction of LLM output streams.

Get Started

Secure your AI pipeline today. Explore KalGuard on our Products page or check out the comparison with VigilanceAI.