Agent Invisible Text Monitor

Capture text AI agents read but users can't see to prevent prompt injection

Problem

A Show HN featured an MCP server that finds text an agent reads but the user can't see (e.g., white text, zero-width characters). Such hidden text can be used by malicious sites to inject instructions and manipulate agent behavior without the user's knowledge.

Solution

A browser extension or proxy middleware that, when an AI agent reads a webpage or document, extracts and highlights all invisible text, zero-width characters, and hidden HTML elements, with a risk score. Supports interception or warnings before agent execution.

Users

Developers using AI agents to browse the web, security researchers, privacy-conscious power users.

MVP

Parse web page text layer; detect invisible characters and CSS-hidden elements; highlight and generate reports; simple rule engine for suspicious instructions; integrate via MCP or API into agent workflows.

Revenue

Open-source base + enterprise version with custom rules, team policies, and audit logs.

Why now

AI agents increasingly browse autonomously, prompt injection is a real threat, and user-side visibility tools are lacking.