Openclaw Pii Anonymizer

Privacy pipeline for OpenClaw - Hybrid regex + Qwen2.5 LLM to scrub PII (names/emails/SSNs/phones/wallets/IPs/paths) before external AI processing. Script works perfectly; auto-hook interception needs debugging.

OpenClaw PII Anonymizer v2.0

Status: ⚠️ Partially Working

• ✅ Script works perfectly (manual invocation)
• ❌ Auto-hook interception needs debugging

Hybrid regex + Qwen2.5:3b LLM to scrub PII before external AI calls.

Quick Start

bash

# 1. Install Ollama model
ollama pull qwen2.5:3b

# 2. Test the script
cd ~/.openclaw/workspace/skills/openclaw-pii-anonymizer
bash privacy-anonymize-v2.sh "My name is John Doe, SSN 123-45-6789"
# Output: My name is [NAME], SSN [SSN]

What It Does

Replaces PII with tokens:

• Names → [NAME]
• SSNs → [SSN]
• Emails → [EMAIL]
• Phones → [PHONE]
• Wallets → [WALLET]
• IPs → [IP]
• Paths → [PATH]

Two-layer approach:

1. Regex (fast, <1ms) - Structured PII (SSN, email, phone, etc.)
2. Qwen2.5:3b (2-3s) - Contextual names (zero hallucination)

Usage

Manual (Working Now)

bash

# In scripts/workflows
ANONYMIZED=$(bash privacy-anonymize-v2.sh "$USER_INPUT")
echo "$ANONYMIZED" | external-api-call

Automatic Hook (TODO)

Hook installed at ~/.openclaw/workspace/hooks/pii-shield/ but doesn't fire on messages yet. Debugging needed.

Requirements

• Ollama running at http://localhost:11434
• Model: qwen2.5:3b (1.9GB) - Better instruction-following than phi3:mini
• RAM: 16GB recommended (6GB minimum but tight)
• Dependencies: bash, curl, jq, sed

Why Qwen2.5:3b?

Tested alternatives:

• ❌ phi3:mini - Hallucinates extra content, too chatty
• ✅ qwen2.5:3b - Zero hallucination, task-focused, smaller (1.9GB vs 2.2GB)
• Alternative: llama3.2:3b (similar performance)

Performance

• Regex layer: <1ms
• LLM layer: 2-3s (only runs if names detected)
• Optimization: Skips LLM for short messages or already-anonymized text

Known Issues

1. Hook system - message:preprocessed event doesn't fire (needs investigation)
2. Auto-interception - Messages not automatically scrubbed yet
3. Re-contextualization - Not implemented (responses stay anonymized)

For Production

Consider NemoClaw for production deployments:

• Built-in PII handling at architecture level
• Enterprise-grade from Nvidia
• No hook debugging needed

This skill: Development/testing, manual workflows
NemoClaw: Production with real customer PII

Testing

bash

# Test 1: Structured PII
bash privacy-anonymize-v2.sh "SSN 123-45-6789, email [e-mail]"
# Expected: SSN [SSN], email [EMAIL]

# Test 2: Names
bash privacy-anonymize-v2.sh "Hi, I'm Alice Johnson"
# Expected: Hi, I'm [NAME]

# Test 3: Complex
bash privacy-anonymize-v2.sh "John Smith ([e-mail]), SSN 987-65-4321, wallet 0x1234567890abcdef1234567890abcdef12345678"
# Expected: [NAME] ([EMAIL]), SSN [SSN], wallet [WALLET]

Files

• privacy-anonymize-v2.sh - Main script (hybrid approach)
• privacy-anonymize.sh - Old v1 (phi3:mini, deprecated)
• hooks/pii-shield/ - Auto-interception hook (needs debugging)
• README.md - Full documentation

Configuration

bash

export OLLAMA_URL=http://localhost:11434
export OLLAMA_MODEL=qwen2.5:3b

Roadmap

• Fix hook system for auto-interception
• Re-contextualization (restore real names in responses)
• Expanded regex patterns (international formats)
• Async LLM (non-blocking)
• Caching for repeated phrases

Version

v2.0 (March 17, 2026)

• Hybrid regex + Qwen2.5:3b
• Script works perfectly
• Hook needs debugging

v1.0.2 (March 1, 2026)

• phi3:mini based
• Hallucination issues

---

License: MIT
Author: Solmas (Seth Blakely)
Homepage: https://github.com/solmas/openclaw-pii-anonymizer