How to redact PII from MCP before LLM invocation?

 

Step 1: Understand the Importance of Redacting PII

 

To effectively redact PII (Personally Identifiable Information) from MCP (Model Context Protocol), it is crucial to first understand what PII entails. PII can include names, addresses, phone numbers, social security numbers, and other data that can identify an individual. Redacting PII is important to ensure privacy and security when invoking a Large Language Model (LLM) such as Claude.

 

Step 2: Identify PII within the MCP Context

 

Before invoking the LLM, developers must identify and isolate any PII present in the MCP components, specifically:

 

• System Instructions: Check if any PII is inadvertently included.
• User Profile: Specifically scrutinize this section for user-related PII.
• Document Context: Ensure no documents contain PII.
• Active Tasks/Goals: Verify that tasks and goals do not include PII.
• Tool Access Logs: These should not log or expose PII.

 

Step 3: Implement a Redaction Mechanism

 

To effectively redact PII, implement a system that identifies and masks this information in your data processing pipeline. Here is a basic example using Python regular expressions to redact PII:

import re

def redact_pii(text):
    # Redact email addresses
    text = re.sub(r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+.[A-Z|a-z]{2,}\b', '[REDACTED]', text)
    
    # Redact phone numbers
    text = re.sub(r'\b\d{3}[-.\s]??\d{2}??[-.\s]??\d{4}\b', '[REDACTED]', text)
    
    # Redact social security numbers
    text = re.sub(r'\b\d{3}-\d{2}-\d{4}\b', '[REDACTED]', text)
    
    # Add more patterns as necessary
    return text

Test the redaction function
example_text = "Contact me at [email protected] or 123-456-7890."
redactedtext = redactpii(example_text)
print(redacted_text)

 

Step 4: Integrate Redaction in MCP Workflow

 

Ensure the redaction function is integrated into your MCP setup so that any context before being sent to the LLM is processed:

def processmcpcontext(mcp_context):
    # Redact PII from each relevant part
    mcpcontext['System Instructions'] = redactpii(mcp_context.get('System Instructions', ''))
    mcpcontext['User Profile'] = redactpii(mcp_context.get('User Profile', ''))
    mcpcontext['Document Context'] = redactpii(mcp_context.get('Document Context', ''))
    mcpcontext['Active Tasks'] = redactpii(mcp_context.get('Active Tasks', ''))
    # Ensure any logs are free of PII
    return mcp_context

Example MCP context
mcp_context = {
    'System Instructions': "You are a helpful assistant for John Doe.",
    'User Profile': "John Doe, call 555-555-5555",
    'Document Context': "Please review John's document.",
    'Active Tasks': "Help John with his tasks."
}

Process context before LLM invocation
processedcontext = processmcpcontext(mcpcontext)

 

Step 5: Verify Redaction and Testing

 

Once you’ve implemented the redaction, thoroughly test it to ensure no PII slips through. Verification can involve:

 

• Unit Testing: Create tests to ensure redaction is applied correctly.
• Peer Review: Have other developers review the redaction logic.
• Continuous Monitoring: Set up logs to monitor for any PII exposure post-redaction.

 

Step 6: Maintain and Update Redaction Logic

 

As new types of PII or changes in your MCP requirements arise, update your redaction logic accordingly. Regular audits and updates ensure ongoing compliance and protection of user data.