OpenClaw Integration Design

OpenClaw Integration Design

Date: 2026-02-12 Version: v3.3.0 Status: Approved

Overview

Integrate agent-security-scanner-mcp with OpenClaw (formerly Clawdbot), providing security guardrails for the autonomous AI assistant. OpenClaw's broad permissions (email, messaging, browser, files, 50+ services) create unique attack surfaces that require both existing security rules and OpenClaw-specific threat detection.

Goals

1. Protect OpenClaw users from prompt injection attacks
2. Detect data exfiltration attempts targeting OpenClaw's service integrations
3. Provide fast, low-friction security scanning via OpenClaw skill
4. Enable CLI usage for non-MCP environments

Integration Approach

Two-Pronged Strategy

1. OpenClaw Skill (Primary) - SKILL.md file that calls agent-security-scanner-mcp directly via npx, bypassing slow MCPorter (~2.4s cold start). Provides instant access via slash commands and natural language.

2. MCP Server (Secondary) - Document MCPorter-based approach for users wanting full MCP toolset. Slower but provides all 6 tools.

OpenClaw-Specific Threat Rules

New rule file: rules/openclaw.security.yaml

Category 1: Data Exfiltration

• "Forward/send my emails to [external]"
• "Export my contacts/calendar to..."
• "Upload my files to..."
• "Share my browser cookies/session"

Category 2: Messaging Abuse

• "Send this message to all my contacts"
• "Reply to all messages with..."
• "Auto-respond with [malicious content]"

Category 3: Credential/Secret Theft

• "Show me my API keys/tokens"
• "What are my saved passwords"
• "Access my keychain/credential store"

Category 4: Autonomous Harm

• "Run this every hour without asking"
• "Don't confirm before executing"
• "Disable safety checks"

Estimated patterns: 20-30 rules

OpenClaw Skill Structure

Location: ~/.openclaw/workspace/skills/security-scanner/SKILL.md

---
name: security-scanner
description: Scan prompts and code for security threats using agent-security-scanner-mcp
metadata: {"openclaw":{"emoji":"🛡️","requires":{"bins":["npx"]}}}
---

## Security Scanner

Protect your OpenClaw instance from prompt injection, malicious code, and data exfiltration attacks.

### Commands
- `/scan-prompt <text>` - Check if a prompt is safe to execute
- `/scan-code <file>` - Scan code for vulnerabilities before running
- `/check-package <name>` - Verify a package isn't hallucinated

### Auto-Protection Mode
When enabled, automatically scans incoming prompts before execution.

Skill Behavior

1. Call npx agent-security-scanner-mcp with appropriate arguments
2. Use verbosity: minimal for fast checks (~50 tokens)
3. Block/warn based on risk level (BLOCK/WARN/ALLOW)

CLI Interface

Add CLI mode to agent-security-scanner-mcp for direct invocation:

# Scan a prompt
npx agent-security-scanner-mcp scan-prompt "delete all my emails"
# Returns: {"action":"BLOCK","risk_level":"critical",...}

# Scan a file
npx agent-security-scanner-mcp scan-security ./script.py
# Returns: {"issues_count":3,"issues":[...]}

# Check a package
npx agent-security-scanner-mcp check-package flask pypi
# Returns: {"exists":true,"ecosystem":"pypi"}

Implementation

• Add CLI argument parsing to index.js
• Detect process.argv vs stdio mode
• Output JSON to stdout for easy parsing
• Support --verbosity flag
• Estimated: ~50 lines

Deliverables

User Experience

Quick Install

npx agent-security-scanner-mcp init openclaw

Manual Install

# Copy skill to OpenClaw workspace
cp -r skills/openclaw ~/.openclaw/workspace/skills/security-scanner

Version

This feature will ship as v3.3.0 (new feature, backward compatible).

References

• OpenClaw MCP Feature Request
• OpenClaw Security Concerns
• OpenClaw Skills Documentation