Add data anonymization for LLM prompt content

[MayaL1M]

Summary

The agent already extracts prompt text from outbound LLM requests, but there is no mechanism to detect or mask sensitive data (PII, credentials, etc.) in that content before it reaches the LLM provider.

Proposed direction

• scan extracted prompt content for PII (names, emails, phone numbers, etc.)
• mask or anonymize before the request is forwarded

What to clarify

• which PII categories to cover initially
• how much extra latency is acceptable

If this does not match the intended direction, please leave comments to clarify. Thanks :)

[jundorok]

which PII categories to cover initially

• I think names, emails, phone numbers could be good start for PII. In future, we might have to expand from PII to enterprise data. For example, user might want to mask all data starting with "6E" or ends with "_MAX". But lets figure this out later.

how much extra latency is acceptable

• its good as long as its at least 1% better than using regex or current industrial method