A comprehensive security auditing tool designed to scan AI agent skills, plugins, and MCP servers for malicious patterns and vulnerabilities.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agentshield-scanner
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 agentshield-scanner 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 AgentShield?
AgentShield serves as a critical defense layer for the modern AI ecosystem, specifically designed to audit Openclaw Skills and MCP servers before they are deployed or executed. It provides developers and security researchers with a powerful engine to detect high-risk behaviors such as data exfiltration, hardcoded credentials, and dynamic code execution backdoors. By analyzing both the source code and the metadata of a skill, it ensures that the AI agents you interact with are safe, transparent, and trustworthy.
In an era where third-party integrations are common, using AgentShield to verify Openclaw Skills protects your sensitive data from being leaked via hidden HTTP requests or unauthorized file access. The tool bridges the gap between functionality and security, offering a clear risk scoring system that translates complex code analysis into actionable security insights.
AgentShield Use Cases
- Performing a pre-install safety check on a remote repository containing new Openclaw Skills.
- Auditing local MCP server implementations for potential server-side request forgery (SSRF) or credential leaks.
- Integrating automated security scans into CI/CD pipelines to prevent low-scoring plugins from being merged.
- Verifying the integrity of .difypkg plugin archives before importing them into production environments.
- Detecting sophisticated prompt injection patterns across multiple languages within an agent's configuration.
How AgentShield Works
- The scanner is triggered via a CLI command pointing to a local directory, a GitHub URL, or a specific package.
- AgentShield parses the directory structure, identifying key files such as SKILL.md, Python scripts, and JavaScript/TypeScript code.
- The engine applies 30+ specialized security rules, utilizing AST (Abstract Syntax Tree) taint tracking for Python to follow dangerous data flows.
- It compares the declared permissions in the skill metadata against the actual operations performed in the code to find privilege mismatches.
- A final security score (0-100) is generated, categorizing the scan results into High, Medium, and Low risk findings for easy interpretation.
AgentShield Setup
You can use AgentShield immediately via npx to audit your Openclaw Skills without needing a global installation.
# Basic scan of a local directory
npx @elliotllliu/agent-shield scan ./path/to/skill/
# Audit a remote repository before installation
npx @elliotllliu/agent-shield install-check https://github.com/user/repo
# Scan a specific plugin archive
npx @elliotllliu/agent-shield scan ./plugin.difypkg
AgentShield Data Schema & Taxonomy
AgentShield evaluates Openclaw Skills using a structured risk taxonomy. The findings are organized as follows:
| Risk Level | Detection Categories | Security Impact |
|---|---|---|
| High Risk | data-exfil, backdoor, crypto-mining | Potential total system compromise or data theft |
| Medium Risk | env-leak, network-ssrf, supply-chain | Exposure of environment variables or dependency vulnerabilities |
| Low Risk | excessive-perms, hidden-files | Minor configuration issues or overly broad access rights |
Users can also use the --json flag to receive a machine-readable schema of all detected vulnerabilities, including file paths and line numbers.
AgentShield Advanced Features
- CI/CD Gatekeeping: Use the
--fail-underflag to break builds if the security score of your Openclaw Skills does not meet a specific threshold. - Tool Shadowing Detection: Identifies cases where a skill's code behavior semantically mismatches its description to prevent social engineering attacks.
- Multi-Step Attack Chain Analysis: Tracks complex sequences of events that might indicate a coordinated exploit attempt.
- Cross-File Taint Tracking: Monitors data flow across multiple files to detect injected code that bypasses simple single-file scanners.
- MCP Runtime Security: Specialized rules for monitoring and securing Model Context Protocol server environments.
SKILL.md
Loading
Related Openclaw Skills
Huisuiyun Invoice Discern
An AI-powered document intelligence skill that automatically recognizes and extracts detailed information from various types of invoices and receipts.
Variant Annotation
A powerful bioinformatics tool for automated gene variant clinical significance annotation and ACMG classification.
Academic Research
A powerful research automation skill that searches millions of academic papers and generates synthesized literature reviews without requiring an API key.
Halo Manager
An Openclaw Skills powered tool for managing Halo blog posts with Markdown support and official API integration.
TikTok Claim Rewriter
A specialized compliance skill designed to transform high-risk marketing claims into policy-safe, high-converting TikTok scripts.
Baidu ERNIE 5.0 Integration
A comprehensive integration guide for connecting Baidu ERNIE 5.0 thinking models to Clawdbot using the Qianfan platform.
Featured*
