Data Leakage Prevention

Data Leakage Prevention (DLP) identifies and protects sensitive data like personally identifiable information (PII), credentials, and proprietary content.

The Risk

LLMs can leak sensitive data in two ways:

1. Training Data Leakage

Models may memorize and regurgitate sensitive data from their training set.

2. Context Leakage

When processing documents or conversation history, models may expose sensitive information in their responses.

Glitch DLP detects sensitive data in both inputs (preventing submission) and outputs (preventing exposure).

PII Categories

Glitch provides comprehensive PII detection with both signature-based (fast, pattern matching) and LLM-based (contextual analysis) detectors.

Signature-Based Detectors (Fast Path)

Detector Type	Description	Example	Latency
pii/email	Email addresses	user@example.com	~11µs
pii/credit_card	Credit/debit card numbers (Visa, MC, Amex, Discover)	4111-1111-1111-1111	~11µs
pii/us_social_security_number	US Social Security Numbers	123-45-6789	~11µs
pii/phone_number	Phone numbers (US, international)	+1 (555) 123-4567	~11µs
pii/ip_address	IPv4 and IPv6 addresses	192.168.1.1, 2001:db8::1	~11µs
pii/iban_code	International Bank Account Numbers	DE89370400440532013000	~11µs

LLM-Based Detectors (Deep Analysis)

Detector Type	Description	Example	Latency
pii/address	Physical mailing addresses	123 Main St, Apt 4B, NYC 10001	~50-100ms
pii/name	Personal names with identifying context	Dr. John Smith, patient ID 12345	~50-100ms

Tip

Signature-based detectors are extremely fast (~11µs) and have high precision for structured data like credit cards and SSNs. LLM-based detectors handle context-dependent data like addresses and names that are harder to detect with patterns.

Configuration

Basic PII Protection

{
  "output_detectors": [
    { "detector_type": "pii/email", "threshold": "L2", "action": "block" },
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/us_social_security_number", "threshold": "L1", "action": "block" }
  ]
}

Full DLP Policy

{
  "input_detectors": [
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/us_social_security_number", "threshold": "L1", "action": "block" }
  ],
  "output_detectors": [
    { "detector_type": "pii/email", "threshold": "L2", "action": "block" },
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/us_social_security_number", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/phone_number", "threshold": "L2", "action": "log" },
    { "detector_type": "pii/ip_address", "threshold": "L2", "action": "log" },
    { "detector_type": "pii/iban_code", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/address", "threshold": "L3", "action": "log" },
    { "detector_type": "pii/name", "threshold": "L3", "action": "log" }
  ]
}

Detection Examples

Credit Card

Output: "Your card ending in 4532-8901-2345-6789 has been charged."

Detection: pii/credit_card
Confidence: 0.98
Action: BLOCKED

Note: Credit card patterns have high confidence due to
Luhn checksum validation.

Email

Output: "Contact john.doe@company.com for support."

Detection: pii/email
Confidence: 0.95
Action: BLOCKED

SSN

Output: "SSN: 123-45-6789"

Detection: pii/us_social_security_number
Confidence: 0.97
Action: BLOCKED

Phone

Output: "Call us at +1 (555) 123-4567"

Detection: pii/phone_number
Confidence: 0.85
Action: LOGGED (if configured)

Threshold Recommendations

Detector	Recommended Level	Notes
pii/credit_card	L1	High-precision pattern matching card prefixes
pii/us_social_security_number	L1	Strict pattern with separators, low false positives
pii/iban_code	L1	Strict international bank account format
pii/email	L2	Email-like patterns can be benign
pii/phone_number	L2-L3	Many number patterns look like phones
pii/ip_address	L2	Version numbers can look like IPs
pii/address	L3	Addresses have high variance, LLM-analyzed
pii/name	L3-L4	Names are highly contextual, LLM-analyzed

Tip

Credit cards, SSNs, and IBANs can use L1 (strict) because their patterns are highly specific. Names and addresses use LLM analysis and benefit from lower thresholds or log-only actions since they’re context-dependent.

Input vs. Output Protection

Input Protection

Prevents users from submitting sensitive data:

{
  "input_detectors": [
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" }
  ]
}

Use cases:

• Prevent accidental PII submission
• Compliance with data handling policies
• Reduce liability from processing sensitive data

Output Protection

Catches sensitive data in LLM responses:

{
  "output_detectors": [
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/email", "threshold": "L2", "action": "block" }
  ]
}

Use cases:

• Prevent training data leakage
• Protect against prompt injection data extraction
• Compliance with privacy regulations (GDPR, CCPA)

Allow Lists for Known Data

Sometimes legitimate data looks like PII. Use allow lists:

{
  "allow_list": {
    "entries": [
      "support@yourcompany.com",
      "sales@yourcompany.com"
    ],
    "match_type": "exact"
  },
  "output_detectors": [
    { "detector_type": "pii/email", "threshold": "L2", "action": "block" }
  ]
}

This blocks email addresses except your company’s support emails.

Actions and Responses

Blocked Response

HTTP/1.1 403 Forbidden
X-Risk-Blocked: true
X-Risk-Categories: pii/credit_card
X-Risk-Confidence: 0.98

{
  "error": {
    "message": "Response blocked: sensitive data detected",
    "type": "data_leakage_prevention",
    "code": "pii_detected"
  }
}

Logged Response

HTTP/1.1 200 OK
X-Risk-Blocked: false
X-Risk-Categories: pii/email
X-Risk-Confidence: 0.85

Content is delivered but logged for review.

Best Practices

1. Always Protect High-Sensitivity Data

Credit cards and SSNs should always be blocked:

{
  "output_detectors": [
    { "detector_type": "pii/credit_card", "threshold": "L1", "action": "block" },
    { "detector_type": "pii/us_social_security_number", "threshold": "L1", "action": "block" }
  ]
}

2. Log Before Blocking (Low-Confidence)

For lower-confidence detections, start with logging:

{
  "output_detectors": [
    { "detector_type": "pii/name", "threshold": "L3", "action": "log" }
  ]
}

Review logged content to tune your policy.

3. Combine with Custom Detectors

Add patterns for domain-specific sensitive data:

{
  "custom_detectors": [
    {
      "name": "employee_id",
      "pattern": "EMP-\\d{6}",
      "action": "block",
      "description": "Internal employee IDs"
    }
  ]
}

Compliance Considerations

DLP helps with:

Regulation	Relevant Data Types
GDPR	All PII (names, emails, addresses)
CCPA	California resident PII
PCI-DSS	Credit card numbers
HIPAA	Health information (custom detectors)

Caution

Glitch DLP is one layer of a compliance strategy. Always consult legal/compliance experts for your specific requirements.

Next Steps

• Allow & Deny Lists — Customize detection rules
• Custom Detectors — Add domain-specific patterns
• Malicious Links — Protect against harmful URLs