LangChain PII Protection — MCP Integration Guide
LangChain vulnerability class: PII leakage in AI model outputs (documented in OWASP Top 10 for LLM Applications) CVSS 9.3 vulnerability. Inject anonym.legal MCP server as PII sanitization layer. Detect & redact PII before LLM processing. Secure Claude, GPT-4, Llama 2 pipelines.
The vulnerability class: PII leakage in AI model outputs (documented in OWASP Top 10 for LLM Applications) Problem
🚨 Vulnerability
LangChain agent tools accept user input directly into vector search, memory, and LLM context. PII like SSN, email, credit card exposed to model training.
📊 Impact
CVSS 9.3 (Critical). Affects all versions <5.2.0. Customers share financial data, health records, credentials with LLM providers.
✅ Solution
Inject MCP server before agent execution. Sanitize PII automatically. Deploy anonymize, analyze, decrypt as MCP tools.
MCP Server Integration
1. Get the MCP Server URL
# Nothing to install or start: the server is hosted. # URL: https://anonym.legal/mcp # Auth: Authorization: Bearer YOUR_API_KEY # MCP access is included in Pro and Business.
2. Configure LangChain to Use MCP
from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent
# Connect to the hosted MCP server
client = MultiServerMCPClient({
"anonym-legal": {
"transport": "streamable_http",
"url": "https://anonym.legal/mcp",
"headers": {"Authorization": "Bearer YOUR_API_KEY"},
}
})
# Load the anonym_legal_* tools
tools = await client.get_tools()
# Create agent with PII protection
agent = create_react_agent(llm, tools)
3. Protect Pipeline
user_input = "My name is John, SSN 123-45-6789"
# Before: VULNERABLE to vulnerability class: PII leakage in AI model outputs (documented in OWASP Top 10 for LLM Applications)
# response = agent.invoke({"input": user_input})
# After: PROTECTED by MCP layer
result = await agent.ainvoke({
"messages": [("user", f"Anonymize: {user_input}")]
})
# Use the anonymized text downstream
cleaned = result["messages"][-1].content
MCP Tools Available
🔍 anonym_legal_analyze_text
Detect PII entities in text. Returns entity types, positions, confidence scores. Use for classification.
🔒 anonym_legal_anonymize_text
Replace PII with redactions, masks, hashes, encryption. Safe for LLM input. Preserve text structure.
🔓 anonym_legal_detokenize_text
Restore original values in tokenized text, using the session_id from a tokenize call. Use in retrieval chains. Bearer token auth.
Use Case: Customer Support Chatbot
Vulnerability Chain
Customer: "My order #12345 shipped to 123 Main St, email is jane@ex.com" # BEFORE: vulnerability class: PII leakage in AI model outputs (documented in OWASP Top 10 for LLM Applications) (VULNERABLE) agent.invoke(user_input) # → LangChain stores "jane@ex.com" in vector DB # → Email sent to OpenAI for embedding # → Stored in LLM training logs
Protection Applied
Customer: "My order #12345 shipped to 123 Main St, email is jane@ex.com"
# AFTER: MCP Protection
mcp_output = anonymize_tool.invoke({
"text": user_input,
"method": "mask"
})
# → "My order #12345 shipped to [ADDRESS], email is [EMAIL]"
# Now safe to send to LLM
response = llm.predict(text=mcp_output)
Watch the API In Action
See PII detection and anonymization via REST API and MCP Server
Secure Your LangChain Pipeline
Patch vulnerability class: PII leakage in AI model outputs (documented in OWASP Top 10 for LLM Applications). Deploy MCP server. Inject PII protection before LLM. Secure Claude, GPT-4, Llama pipelines.
Set Up MCP ServerAlso from anonym.legal