Skip to main content

​
What is Guardian?

Guardian is Continum’s pre-LLM protection layer that detects and redacts PII before it reaches your LLM provider. It runs in under 100ms, making it suitable for production use.

​
Why Guardian?

​
The Problem

Users might accidentally send PII to your AI application: 
User: "My email is john@example.com and SSN is 123-45-6789"
  ↓
LLM receives PII (❌ compliance violation)
  ↓
LLM might include PII in response
  ↓
PII logged in LLM provider's systems

​
The Solution

Guardian intercepts and redacts PII before the LLM sees it: 
User: "My email is john@example.com and SSN is 123-45-6789"
  ↓
Guardian detects PII (< 100ms)
  ↓
Redacted: "My email is [EMAIL_REDACTED] and SSN is [SSN_REDACTED]"
  ↓
LLM receives clean input ✅

​
How It Works

​
1. Pattern Matching

Guardian uses regex patterns to detect common PII types:
  • Email addresses
  • US SSN (123-45-6789)
  • Credit cards (4111-1111-1111-1111)
  • Phone numbers (+1-555-123-4567)
  • UK/US Passports
  • Health/Medical IDs
  • IP addresses
  • Physical addresses

​
2. Entity Detection

For complex PII (names, locations), Guardian uses lightweight ML models:
  • Named Entity Recognition (NER)
  • Context-aware detection
  • Confidence scoring

​
3. Redaction Strategies

Guardian offers multiple redaction strategies: Full redaction: 
"john@example.com" → "[EMAIL_REDACTED]"
Partial redaction (preserves context): 
"john@example.com" → "j****@example.com"
"123-45-6789" → "***-**-6789"
Hashed redaction (consistent across requests): 
"john@example.com" → "[EMAIL_a3f9c2]"

​
Configuration

​
Enable Guardian

import { Continum } from '@continum/sdk';

const continum = new Continum({
  continumKey: process.env.CONTINUM_KEY,
  openaiKey: process.env.OPENAI_API_KEY,
  guardianEnabled: true,  // Enable Guardian
  guardianAction: 'REDACT_AND_CONTINUE'  // or 'BLOCK_ON_DETECT'
});

​
Guardian Actions

Configure what happens when PII is detected: REDACT_AND_CONTINUE (default):
  • Redacts PII from prompt
  • Continues with LLM call
  • User gets response with redacted input
BLOCK_ON_DETECT:
  • Blocks LLM call entirely
  • Returns error to user
  • No LLM call made
ALLOW_ALL:
  • Logs detection but doesn’t redact
  • Useful for monitoring without blocking

​
Sandbox Configuration

Set Guardian action per sandbox: 
// In dashboard or via API
{
  "name": "Strict PII Protection",
  "slug": "pii_strict",
  "sandboxType": "PII_DETECTION",
  "guardianAction": "BLOCK_ON_DETECT",  // Block if PII detected
  "alertThreshold": "MEDIUM"
}

​
API Endpoint

​
POST /guardian/scan

Scan a prompt for PII before sending to LLM: 
const response = await fetch('https://api.continum.co/guardian/scan', {
  method: 'POST',
  headers: {
    'x-continum-key': process.env.CONTINUM_KEY,
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    sandbox: 'pii_strict',
    provider: 'openai',
    model: 'gpt-4o',
    systemPrompt: 'You are a helpful assistant',
    userInput: 'My email is john@example.com'
  })
});

const result = await response.json();
/*
{
  "action": "REDACT",
  "violations": ["EMAIL"],
  "reasoning": "Email j****@example.com detected",
  "cleanPrompt": "My email is [EMAIL_REDACTED]",
  "confidence": 0.98,
  "detectedEntities": [
    {
      "type": "EMAIL",
      "originalValue": "john@example.com",
      "redactedValue": "j****@example.com",
      "start": 12,
      "end": 28,
      "confidence": 0.98
    }
  ],
  "durationMs": 87
}
*/

​
SDK Integration

Guardian is automatically integrated when enabled: 
const continum = new Continum({
  continumKey: process.env.CONTINUM_KEY,
  openaiKey: process.env.OPENAI_API_KEY,
  guardianEnabled: true
});

// Guardian runs automatically before LLM call
const response = await continum.llm.openai.gpt_4o.chat({
  messages: [
    { role: 'user', content: 'My SSN is 123-45-6789' }
  ]
});

// If guardianAction is REDACT_AND_CONTINUE:
// - Guardian detects SSN
// - Redacts to "My SSN is [SSN_REDACTED]"
// - Sends redacted prompt to OpenAI
// - Returns response to user

// If guardianAction is BLOCK_ON_DETECT:
// - Guardian detects SSN
// - Throws error: "PII detected, request blocked"
// - No LLM call made

​
Performance

Guardian is optimized for production use:
OperationLatencyNotes
Pattern matching10-30msRegex-based detection
Entity detection50-80msML-based NER
Total Guardian overhead< 100msAcceptable for production

​
Detected PII Types

Guardian detects the following PII categories:

​
Personal Identifiers

  • Full names (with context)
  • Email addresses
  • Phone numbers
  • Physical addresses

​
Government IDs

  • US Social Security Numbers (SSN)
  • UK/US Passport numbers
  • Driver’s license numbers
  • National ID numbers

​
Financial Data

  • Credit/debit card numbers
  • CVV codes
  • Bank account numbers
  • IBAN codes
  • Cryptocurrency wallet addresses

​
Health Data

  • Medical record numbers (MRN)
  • Health insurance IDs
  • Prescription numbers
  • Biometric identifiers

​
Technical Identifiers

  • IP addresses (when linked to individuals)
  • Device IDs
  • Session tokens
  • API keys

​
Guardian Scans in Dashboard

All Guardian scans are logged and visible in the dashboard:
  • Detected PII types
  • Redaction applied
  • Action taken (ALLOW, REDACT, BLOCK)
  • Confidence scores
  • Duration

​
Best Practices

​
When to Enable Guardian

✅ Enable Guardian when:
  • Users can input free-form text
  • Application handles sensitive data
  • Compliance requires pre-LLM protection
  • GDPR/CCPA/HIPAA applies
❌ Skip Guardian when:
  • Input is fully controlled (no user input)
  • Performance is critical (< 100ms not acceptable)
  • Post-hoc auditing is sufficient

​
Choosing Guardian Action

REDACT_AND_CONTINUE:
  • Best for most use cases
  • Maintains user experience
  • Protects LLM from PII
  • Logs violations for monitoring
BLOCK_ON_DETECT:
  • Maximum protection
  • Use for highly sensitive applications
  • May frustrate users if false positives
  • Consider showing helpful error message
ALLOW_ALL:
  • Monitoring mode
  • Use during testing/rollout
  • Logs detections without blocking
  • Upgrade to REDACT after validation

​
Limitations

​
False Positives

Guardian may flag non-PII as PII: 
"Call me at extension 1234" → Detected as phone number
"My favorite number is 123-45-6789" → Detected as SSN
Mitigation: Use confidence thresholds and context-aware detection.

​
False Negatives

Guardian may miss obfuscated PII: 
"My email is john at example dot com" → Not detected
"SSN: one two three four five six seven eight nine" → Not detected
Mitigation: Combine Guardian with post-LLM Mirror auditing.

​
Language Support

Guardian is optimized for English. Other languages may have reduced accuracy.

​
Next Steps

Mirror

Learn about post-LLM auditing

Sandbox

Configure sandbox types

API Reference

Guardian API documentation

PII Detection

Comprehensive PII detection guide