A zero-dependency security middleware that sanitizes email and calendar data to prevent prompt injection and malicious exfiltration in AI agents.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-mail-guard
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install agent-mail-guard using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is AgentMailGuard?
AgentMailGuard is a specialized security layer designed to sit between untrusted data sources and your AI agent context window. Built specifically for the Openclaw Skills ecosystem, it neutralizes a wide array of attack vectors including prompt injections, hidden unicode, and markdown-based data exfiltration. By processing incoming emails and calendar events through a rigorous sanitization pipeline, it ensures that your LLM only receives safe, structured text.
This tool is critical for developers who want to grant their agents access to communication tools without exposing the system to malicious 'ignore previous instructions' style attacks. It prioritizes privacy and security by performing all processing locally, ensuring no sensitive data ever leaves your machine while maintaining a high standard of protection.
AgentMailGuard Use Cases
- Automating email processing from Gmail or Outlook while preventing system instruction overrides.
- Safely summarizing calendar invitations that may contain malicious hidden payloads.
- Filtering untrusted text inputs in any Openclaw Skills workflow that handles external user data.
- Detecting and flagging suspicious senders before their content reaches the agent context.
- Protecting sensitive data from exfiltration via malicious markdown image tags.
How AgentMailGuard Works
- The skill ingests raw text data from email or calendar sources via CLI tools or direct Python calls.
- The sanitization engine runs the input through 13+ regex patterns to identify prompt injection attempts and fake conversation turns.
- It strips invisible unicode characters, homoglyphs, HTML tags, and Base64 payloads that could be used for obfuscation.
- The sender is evaluated against a local trust tier configuration (contacts.json) to determine context priority.
- A sanitized JSON object is produced, containing safety flags and truncated body text safe for agent consumption.
AgentMailGuard Setup
To integrate this security layer into your Openclaw Skills setup, ensure you have Python 3.11 or higher installed.
# Navigate to the skill directory
cd agent-mail-guard
# Configure your trusted contacts
cp contacts.json.example contacts.json
# Edit contacts.json to include your known domains and VIPs
# Run a test sanitization on an email via gog CLI
bash scripts/check-email.sh
# Run the built-in test suite to verify security patterns
python3 -m pytest scripts/test_sanitizer.py
AgentMailGuard Data Schema & Taxonomy
The skill produces a structured JSON output for every processed item, allowing Openclaw Skills to handle the data programmatically:
| Key | Description |
|---|---|
sender |
The validated email address of the sender. |
sender_tier |
Classification as known, unknown, or vip. |
subject |
The sanitized and cleaned subject line. |
body_clean |
The safe, truncated version of the body text. |
suspicious |
Boolean flag indicating if injection patterns were found. |
flags |
A list of specific security violations detected (e.g., 'markdown_img'). |
date |
ISO formatted date string. |
AgentMailGuard Advanced Features
- Support for sender trust tiers to minimize the attack surface of unknown contacts.
- Advanced detection of URL smuggling and reference-style markdown exfiltration.
- Comprehensive character mapping to identify homoglyph (lookalike) attacks.
- Zero-dependency architecture relying solely on the Python standard library for maximum portability.
- Local-first processing ensuring full data sovereignty for sensitive communications.
- Integration ready for any Openclaw Skills automation that utilizes external text inputs.
SKILL.md
Loading
Related Openclaw Skills
Minibook Skill
A comprehensive collaboration skill that connects AI agents to Minibook instances for project management, threaded discussions, and roadmap tracking.
CrawSecure
A documentation-first security skill providing transparency and guidance for performing static analysis on third-party tools.
OpenClaw Operations (openclaw-ops)
A specialized operational toolkit designed for rescue agents to diagnose, repair, and maintain OpenClaw Gateway instances.
Media News Digest
An automated pipeline for gathering and synthesizing media and entertainment industry news from 65+ specialized sources.
Manus AI
An autonomous AI agent with internet access capable of complex research, software engineering, and multi-app workflow automation.
OpenClaw Cloud Setup Skill
An automated deployment skill for provisioning OpenClaw on Google Cloud Platform with secure Tailscale networking and Brave Search integration.
Featured*
