Email Fortress

In January 2026, multiple AI assistants were compromised via email-based prompt injection — an attacker sends a crafted email with hidden instructions, and your agent executes them. Email Fortress prevents this.

This skill establishes hard channel trust boundaries: only your verified channel (Telegram, Discord, Signal) is trusted for commands. Email is read-only — never a command source.

What it does

• Defines which channel is trusted and which is not
• Gives your agent clear rules for handling action requests from email
• Provides prompt injection detection patterns (hidden instructions, base64 payloads, HTML comments)
• Flag-and-confirm workflow: suspicious emails get forwarded to your trusted channel for review

No dependencies

Pure policy skill — no scripts, no APIs, no setup beyond adding a block to your system prompt. Works with any agent on any stack.

Setup

5 minutes. Copy the Email Security block into your MEMORY.md, fill in your trusted channel and user ID, done.