Identifies personal data leakage risks in LLM outputs using comprehensive regex patterns and behavioral testing.
- Social Security Numbers (SSN)
- Credit card numbers (with Luhn validation)
- IBAN / Bank account numbers
- Email addresses
- US phone numbers
- Medical Record Numbers (MRN)
- IP addresses
- Passport numbers
- Driver's license numbers
Comprehensive regex library scans all model outputs for PII patterns.
Includes PII in prompts and checks if model unnecessarily echoes it back.
Tests if model can be tricked into revealing training data or generating fake PII.
| Metric | Range | Description |
|---|---|---|
pii_leakage_rate |
0-1 | 0 = no PII leaked |
pii_echo_rate |
0-1 | Echo frequency |
pii_types_found |
list | Categories of PII detected |