Redaction in action
Type or paste a prompt below. The real-time scanner highlights sensitive content and shows what gets redacted before it reaches the AI service.
Detection patterns
| Pattern | Category | Regex | Example match |
|---|---|---|---|
| SSN | PII | \d{3}-\d{2}-\d{4} | 412-55-8901 |
| API Key | Credentials | sk-[a-zA-Z0-9-]+ | sk-prod-a8f2e1b9 |
| AWS Key | Credentials | AKIA[A-Z0-9]{12,} | AKIA3EXAMPLE7KEY |
| IP:Port | Network | \d+\.\d+\.\d+\.\d+:\d+ | 10.0.5.23:4444 |
| PII | [a-z.]+@[a-z]+\.[a-z]+ | j.smith@corp.com | |
| Person name | PII | NLP entity recognition | John Smith |
Real Workforce AI Security uses ML-based NER (Named Entity Recognition) — far more sophisticated than regex.
These patterns are a simplified demo of the same principle.
Loading...
Think Deeper
Try this:
A redacted prompt loses critical context and the AI gives a useless response. How do you handle this?
This is the utility vs security trade-off. Options: 1. Use Ask action instead — let the user justify the sensitive data. 2. Replace with synthetic data (fake names, placeholder IPs) that preserves structure. 3. Use an on-prem AI for prompts requiring sensitive context. The best solution depends on the use case.
Cybersecurity tie-in: Redaction preserves the utility of the AI interaction
while removing the risk. The AI still receives the structure of the prompt — just not the
sensitive values. This is the same principle as data masking in database security.