guard-scanner_skill

Overview

This skill is a security scanner for AI agent skills designed to be run before installing or running any new skill from ClawHub or external sources. It detects prompt injection, credential theft, exfiltration, identity hijacking, and 14 additional threat categories. It also provides a Runtime Guard hook that can block dangerous tool calls in real time to prevent malicious behavior during execution.

How this skill works

The scanner performs static analysis of skill code to find malicious patterns such as concealed network calls, shell execution, and instructions that attempt to override agent identity or memory. It applies behavioral heuristics and signatures to flag prompt-injection patterns and data-exfiltration attempts. Additionally, the Runtime Guard intercepts tool invocations and enforces policies to block or sandbox suspicious actions during runtime.

When to use it

• Before installing any skill from ClawHub or third-party sources.
• When auditing archived or legacy skills for safe reuse.
• Prior to granting a skill access to credentials, memory, or external tools.
• When deploying skills in production or multi-tenant environments.
• As part of a CI pipeline that vets skills before merge or release.

Best practices

• Run the scanner with default and strict policy profiles to compare findings.
• Review flagged items manually and prioritize remediation of credential and execution risks.
• Use the Runtime Guard in live environments where skills may call external tools.
• Combine scanner output with dedicated secrets scanning and runtime monitoring.
• Keep scanner rules and signatures updated to catch evolving injection techniques.

Example use cases

• Scanning a newly downloaded skill to detect hidden curl/bash download-and-execute commands.
• Auditing archived skill collections for instructions that persist secrets to disk or memory.
• Blocking a skill at runtime from making outbound network requests or writing tokens to files.
• Integrating scans into CI to prevent unsafe skills from reaching production.
• Reviewing a skill that attempts to change agent identity or replace core instruction files.