Implementation:Protectai Llm guard Sensitive

Knowledge Sources	LLM Guard LLM Guard Documentation
Domains	NLP, Data_Privacy, Output_Validation
Last Updated	2026-02-14 12:00 GMT

Overview

Concrete tool for detecting and optionally redacting PII in LLM outputs using NER models and Presidio, provided by the LLM Guard library.

Description

The Sensitive class is an output scanner that detects personally identifiable information in LLM-generated text. It uses the same detection pipeline as the Anonymize input scanner (DeBERTa NER + regex patterns + Presidio AnalyzerEngine) but is designed for output validation. When redact is enabled, detected entities are replaced using Presidio's AnonymizerEngine.

Usage

Import this scanner to detect PII leakage in LLM outputs. Use with redact=True to automatically remove sensitive data before returning responses to users.

Code Reference

Source Location

Repository: llm-guard
File: llm_guard/output_scanners/sensitive.py
Lines: L29-116

Signature

class Sensitive(Scanner):
    def __init__(
        self,
        *,
        entity_types: list[str] | None = None,
        regex_patterns: list[DefaultRegexPatterns | RegexPatternsReuse] | None = None,
        redact: bool = False,
        recognizer_conf: NERConfig | None = None,
        threshold: float = 0.5,
        use_onnx: bool = False,
        language: str = "en",
    ) -> None:
        """
        Args:
            entity_types: PII types to detect. Default: all standard types.
            regex_patterns: Custom regex patterns. Default: None.
            redact: Redact detected PII in output. Default: False.
            recognizer_conf: NER model config. Default: DEBERTA_AI4PRIVACY_v2_CONF.
            threshold: Minimum confidence score. Default: 0.5.
            use_onnx: Use ONNX runtime. Default: False.
            language: Detection language ("en" or "zh"). Default: "en".
        """

    def scan(self, prompt: str, output: str) -> tuple[str, bool, float]:
        """
        Detect PII in LLM output and optionally redact.

        Returns:
            - Output (original or redacted if redact=True)
            - False if PII detected, True if clean
            - Risk score based on highest detection confidence
        """

Import

from llm_guard.output_scanners import Sensitive

I/O Contract

Inputs

Name	Type	Required	Description
entity_types	list[str]	No	PII types to detect (default: all standard types)
regex_patterns	list	No	Custom regex patterns (default: None)
redact	bool	No	Redact detected PII (default: False)
recognizer_conf	NERConfig	No	NER model config (default: DEBERTA_AI4PRIVACY_v2_CONF)
threshold	float	No	Confidence threshold (default: 0.5)
use_onnx	bool	No	Use ONNX runtime (default: False)
language	str	No	Detection language (default: "en")
prompt	str	Yes (scan)	Original prompt
output	str	Yes (scan)	LLM output to check

Outputs

Name	Type	Description
output	str	Original output or redacted version
is_valid	bool	False if PII detected above threshold
risk_score	float	Highest detection confidence, normalized

Usage Examples

Basic Detection

from llm_guard.output_scanners import Sensitive

scanner = Sensitive(threshold=0.5)

prompt = "Tell me about the CEO"
output = "The CEO is John Smith, reachable at john@company.com"
_, is_valid, score = scanner.scan(prompt, output)
# is_valid: False (PII detected)

With Redaction

from llm_guard.output_scanners import Sensitive

scanner = Sensitive(redact=True, threshold=0.5)

prompt = "Tell me about the CEO"
output = "The CEO is John Smith, reachable at john@company.com"
redacted, is_valid, score = scanner.scan(prompt, output)
# redacted: "The CEO is <PERSON>, reachable at <EMAIL_ADDRESS>"

Related Pages

Implements Principle

Principle:Protectai_Llm_guard_Sensitive_Data_Detection

Requires Environment

Uses Heuristic

Heuristic:Protectai_Llm_guard_ONNX_Runtime_Optimization

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment