PII Detection & Masking

LLMProxy includes dual-mode PII detection that prevents sensitive personal information from reaching LLM providers.

Detection Modes

Presidio NLP (opt-in)

When presidio-analyzer is installed, LLMProxy uses Microsoft Presidio for NLP-powered entity recognition. Supports 18 entity types including names, addresses, and context-dependent patterns.

pip install presidio-analyzer presidio-anonymizer

Regex Fallback (always available)

Built-in regex patterns detect common PII without external dependencies:

Pattern	Example
Email	user@example.com
Phone	+1-555-0123
SSN	123-45-6789
Credit Card	4111-1111-1111-1111
IBAN	DE89370400440532013000

Vault Tokenization

PII is replaced with vault tokens, not deleted:

Input:  "Contact john@example.com for details"
Output: "Contact [PII_EMAIL_a7b3c] for details"

The original values are stored in an in-memory vault keyed by token. Responses can be demasked to restore original PII before returning to the client.

How It Works

mask_pii(text) → tokenized text + vault entries
                          ↓
              Request sent to LLM provider
                          ↓
demask_pii(response) → original PII restored

Pipeline Position

PII masking runs as the PII Neural Masker default plugin in the PRE_FLIGHT ring (priority 20), after auth but before cache lookup and routing.

Configuration

PII masking is always active when the security module is enabled. The audit trail can optionally mask PII in logs:

logging:
  audit_trail:
    enabled: true
    mask_pii: true