Crate iron_safety

Expand description

PII detection and output sanitization for multi-agent systems.

Provides real-time scanning of LLM outputs to detect and redact personally identifiable information (PII) before responses reach end users. Essential for compliance with data protection regulations (GDPR, CCPA, HIPAA).

§Purpose

This crate provides safety guardrails for LLM output:

Real-time PII detection in streaming responses
Automatic redaction of sensitive data patterns
Configurable pattern matching for different PII types
Low-latency scanning suitable for production use

§Supported PII Types

Currently detects and redacts:

Email addresses: Standard RFC-compliant email patterns
Phone numbers: US format (XXX-XXX-XXXX)

Additional patterns can be added by extending PiiDetector.

§Key Types

PiiDetector - Main detector with regex-based pattern matching

§Public API

§Basic Usage

use iron_safety::PiiDetector;

let detector = PiiDetector::new()?;

// Check if text contains PII
let text = "Contact me at user@example.com or 555-123-4567";
if detector.check(text) {
  println!("⚠️ PII detected!");
}

// Redact PII from text
let safe_text = detector.redact(text);
// "Contact me at [EMAIL_REDACTED] or [PHONE_REDACTED]"

§Streaming Integration

use iron_safety::PiiDetector;

async fn stream_llm_response(detector: &PiiDetector) -> Result<(), Box<dyn std::error::Error>> {
  let mut buffer = String::new();

  // Simulate streaming chunks
  for chunk in &["Hello ", "user@", "example.com", "!"] {
    buffer.push_str(chunk);

    // Check accumulated buffer for PII
    if detector.check(&buffer) {
      // Redact before sending to client
      let safe_chunk = detector.redact(&buffer);
      send_to_client(&safe_chunk).await?;
      buffer.clear();
    }
  }

  Ok(())
}

async fn send_to_client(text: &str) -> Result<(), Box<dyn std::error::Error>> {
  // Implementation...
}

§Integration with Analytics

use iron_safety::PiiDetector;

fn process_response(text: &str, detector: &PiiDetector, store: &EventStore) {
  if detector.check(text) {
    // Log PII detection for compliance audit
    // store.record_pii_detection(agent_id, timestamp);

    // Redact before returning
    let safe = detector.redact(text);
    // ... send safe text
  }
}

§Feature Flags

enabled - Enable PII detection (disabled for minimal builds)

§Performance

PII detection uses compiled regex patterns with the following characteristics:

Pattern compilation: One-time cost at initialization
Check operation: O(n) where n = text length, typically <100μs for 1KB
Redact operation: O(n) with string allocation

For streaming, check small chunks frequently rather than buffering entire responses.

§Limitations

Current implementation has known limitations:

Pattern-based only: No ML-based detection for context-aware PII
US-centric: Phone patterns only match US format (XXX-XXX-XXXX)
No international support: Email/phone patterns dont cover all locales
False positives possible: Aggressive matching may flag non-PII

For production use in regulated environments, consider:

Adding locale-specific patterns
Implementing allowlists for known-safe patterns
Combining with ML-based PII detection services
Regular auditing of redaction logs

Structs§

PiiDetector: PII detector with configurable patterns