PII Protection Levels by LLM

This guide explains the level of PII protection you can achieve when combining PII Firewall with different AI services and LLMs.

Key Insight: Protection Level Depends on the Route, Not the LLM

The protection level is determined by how you route text to the LLM, not which LLM you use.

Route	Protection Level	Does the LLM see raw PII?
Chrome Extension	✅ Full protection	❌ Never
API / SDK (developer integration)	✅ Full protection	❌ Never
SEva pipeline	✅ Full protection	❌ Never
Claude Desktop + MCP (CLAUDE.md rule)	🟡 High protection	⚠️ Briefly (Claude judges, then masks)
Claude Desktop + session instruction	🟡 High protection	⚠️ Briefly on every message
No protection (direct AI input)	❌ No protection	✅ Sees everything

PII Protection by LLM — Comparison Table

LLM / Service	Chrome Ext	API/SDK	MCP	Full PII FW	Notes
Claude.ai (Anthropic)	✅	✅	⚠️	✅ via Ext/API	Claude Desktop MCP: sees briefly
ChatGPT (OpenAI)	✅	✅	❌	✅ via Ext/API	—
Gemini (Google)	✅	✅	❌	✅ via Ext/API	—
Microsoft Copilot	🟡	✅	❌	✅ via API	Web extension support unconfirmed. Full support via Azure OpenAI API
Perplexity AI	🟡	✅	❌	✅ via API	Web extension support unconfirmed
Local LLM (Ollama / LM Studio)	❌	✅	❌	✅	Data stays local by design. API integration adds an extra safety layer
Mistral AI	🟡	✅	❌	✅ via API	—
Llama (Meta, self-hosted)	❌	✅	❌	✅	Low risk by default (local execution)

Legend: ✅ Supported / Full protection ／ 🟡 Conditional / Unconfirmed ／ ❌ Not supported

How Each Route Works

✅ Chrome Extension (Full Protection)

User types message
      ↓
Extension captures text before sending
      ↓
PII Firewall core engine masks locally
      ↓
Only masked text is sent to the AI service
      ↓
Extension captures AI response
      ↓
PII Firewall restores tokens to original values

The LLM never receives raw PII. Works identically for Claude.ai, ChatGPT, and Gemini.

✅ API / SDK Integration (Full Protection)

Your application
      ↓ call mask_pii()
Masked text
      ↓
Call LLM API (Claude / OpenAI / Gemini / etc.)
      ↓
Response (with tokens)
      ↓ call restoreAll()
Restored response displayed to user

Works with any LLM API — just mask before calling and restore after.

⚠️ Claude Desktop + MCP (High Protection, but Sees Briefly)

User message
      ↓
Claude receives it ← ⚠️ Sees raw PII for a moment
      ↓
Calls mask_pii tool (via CLAUDE.md rule or session instruction)
      ↓
Generates response using only masked text

Because Claude decides when to call mask_pii, it reads the original message first. This is high protection in practice, but not strictly "complete" protection.

Chrome Extension — Supported Services

Service	Mask before send	Restore after receive	Injection detection
ChatGPT (chat.openai.com)	✅	✅	✅
Claude.ai	✅	✅	✅
Gemini (gemini.google.com)	✅	✅	✅
Gmail (AI reply assistance)	✅	✅	✅
Outlook Web (AI assistance)	✅	✅	✅

Other services

For AI services not listed above (Microsoft Copilot, Perplexity, etc.), equivalent protection is available via the API/SDK route. Contact us about adding Chrome extension support for additional services.

Summary: 3 Ways to Achieve Full Protection

1. Install the Chrome Extension (easiest — for individual users)
→ Best for daily use of ChatGPT / Claude.ai / Gemini

2. Integrate via API / SDK (for developers)
→ Works with any LLM. Pattern: mask → call LLM → restore

3. Build a pipeline like SEva (for enterprise)
→ System-side automatic masking before any LLM — including Claude — sees the message

PII Protection Levels by LLM ​

Key Insight: Protection Level Depends on the Route, Not the LLM ​

PII Protection by LLM — Comparison Table ​

How Each Route Works ​

✅ Chrome Extension (Full Protection) ​

✅ API / SDK Integration (Full Protection) ​

⚠️ Claude Desktop + MCP (High Protection, but Sees Briefly) ​

Chrome Extension — Supported Services ​

Summary: 3 Ways to Achieve Full Protection ​