Technical Whitepaper & User Guide
Prompt PII Masker & Restorer Technical Architecture Guide
When sending prompts containing client names, phone numbers, emails, or government IDs to LLMs like ChatGPT or Claude, organizations face serious compliance risks under PIPA and GDPR. SafeAI eliminates this risk via 100% client-side memory pseudonymization with zero cloud upload.
1. Regex Tokenization & Reversible Session Unmasking
SafeAI scans sensitive strings directly inside the V8 JavaScript engine, replacing them with randomized tokens like [USER_1] and [PHONE_1]. A temporary reverse mapping is stored in local memory, enabling 1-click restoration once the AI response arrives. No central database or third-party server is ever involved.
2. Recommended Anonymization Workflow by Scenario
| Business Scenario | Leakage Risk Factor | SafeAI Recommended Action |
|---|---|---|
| Customer Support & CS Logs | Customer names, contact numbers, addresses | Paste chat log → Click 'Run PII Masking' → Unmask after AI analysis |
| HR & Contract Legal Review | National ID numbers, bank accounts, executive names | Redact ID and account numbers before asking LLM for clause analysis |
| Healthcare / Medical Summaries | Patient names, disease codes, medical records | Replace patient identifiers with anonymized tokens for HIPAA safety |
| IT Infrastructure Incident Logs | Internal IPs, hostnames, employee email addresses | Mask corporate network IPs and domain names prior to debugging queries |
3. Enterprise Compliance & Troubleshooting FAQ
-
✓
Zero Cloud Retention: SafeAI operates entirely within client RAM. Closing or refreshing the tab permanently wipes the in-memory mapping table.
-
✓
LLM Context Preservation: Anonymized tokens ([USER_1], [PHONE_1]) maintain syntactic entity distinction so LLMs produce contextually accurate replies.
-
✓
Bidirectional Restoration: Once the AI outputs the response, paste it back and click 'Restore Original Entities' to reconstitute original names in 1 click.