One of the most common risks with AI applications is PII leakage. Users paste emails, phone numbers, and other personal data into prompts without realizing it gets sent to third-party AI providers like OpenAI or Anthropic.
What counts as PII?
In the context of AI prompts, PII includes:
- Email addresses
- Phone numbers
- Social security numbers
- Credit card numbers
- Physical addresses
- Names (when combined with other identifiers)
Detection approaches
Pattern matching (regex)
The simplest approach uses regular expressions to detect known PII patterns:
- Emails: standard email regex
- Phone numbers: US format patterns (xxx-xxx-xxxx, etc.)
- SSNs: xxx-xx-xxxx pattern
This is what SignalVault uses for its contains_pii rule type.
Named entity recognition (NER)
More sophisticated approaches use ML models to identify PII, but these add latency and complexity. For real-time guardrails, regex patterns provide the best tradeoff of speed vs. accuracy.
What to do when PII is detected
You have several options:
- Block the request entirely — safest, but disrupts the user experience
- Redact the PII — replace with placeholders like [EMAIL], [PHONE]
- Warn — log the detection but allow the request through
The right choice depends on your compliance requirements. For SOC2 and GDPR, blocking or redacting is typically required.
Implementation with SignalVault
Create a PII detection rule in your app's Rules tab:
- Type: PII Detection
- Action: Block or Redact
- Environment: All environments
- Patterns: email, phone, ssn
Every request will be checked automatically, and violations appear in your dashboard.