InBox Wall Sanitizes Hidden Email Content Before AI Processing

Summary

InBox Wall is a local-first Python application that places a read-only security layer between an email inbox and an AI assistant. It connects through IMAP without changing mailbox state, inlines CSS, parses HTML, and scans text nodes for hidden rendering tricks, invisible Unicode, homograph characters, and known prompt-injection phrases. The project uses inheritance-aware style analysis, a strict allow-list of visible HTML tags, background-aware checks for white-on-white text, and bounded phrase matching designed to avoid regular-expression denial-of-service risks. Clean text is stored in a local SQLite database and exposed through a REST API and Model Context Protocol server, while blocked content and original hidden text remain in a separate vault that AI tools cannot access. The design fails closed: scanner errors, invalid input, CSS inlining failures, and files over 5 MB produce a LOCKED_AWAY result with a blank body. It also limits classified text to 2,000 characters, purges quarantined originals after 30 days, and protects local files with restrictive permissions. A dashboard provides visible markers for invisible characters, threat-score explanations, review actions, and an immutable audit log. The project does not scan attachments or image text, analyze phishing URLs, or detect purely persuasive social engineering, and it is intended as defense in depth alongside existing email filters and human review. The repository includes a pytest benchmark based on whether injected tokens leak into clean text, but says its sample corpus should be expanded for production use.