Skip to content

SKILL.md

Latest commit

 

History

History
76 lines (57 loc) · 3.55 KB

SKILL.md

File metadata and controls

76 lines (57 loc) · 3.55 KB
name skillguard
description Security audit agent skills for vulnerabilities and malicious patterns. Use when asked to "scan a skill", "check if a skill is safe", "audit a skill for security", "review skill security", "is this skill malicious", or when installing/reviewing any third-party agent skill. Accepts local SKILL.md files, skill directories, .skill packages, or GitHub URLs. Maps findings to OWASP Agentic Top 10 (2026) and MITRE ATLAS.

SkillGuard — Agent Skill Security Auditor

Audit agent skills against industry-standard frameworks before installation.

How to Use

  1. Read the target skill's SKILL.md and all files in the skill directory
  2. Read references/owasp-agentic-top10.md for the risk categories
  3. Read references/mitre-atlas-mapping.md for technique-level mapping
  4. Read references/audit-checklist.md for the structured audit procedure
  5. Perform the audit following the checklist
  6. Output the report in the format below

Audit Procedure

For each skill under review:

Step 1 — Inventory

List all files in the skill. Note languages, dependencies, external URLs, and any scripts.

Step 2 — Frontmatter Validation

Check SKILL.md has valid YAML frontmatter with name and description per the Agent Skills spec. Flag missing or malformed fields.

Step 3 — Intent Verification

Read the description field. Then read ALL code and instructions in the skill. Answer: does the skill's actual behavior match its stated purpose? Mismatches are the strongest signal of malicious intent.

Step 4 — Risk Assessment

Walk through each OWASP Agentic Top 10 category (see references/owasp-agentic-top10.md). For each category, check whether the skill exhibits relevant patterns. Map any findings to MITRE ATLAS techniques where applicable (see references/mitre-atlas-mapping.md).

Step 5 — Verdict

Assign a verdict based on findings:

Verdict Criteria
✅ SAFE No findings, or only informational quality notes
⚠️ SUSPICIOUS Medium-severity findings that need human review
🚨 MALICIOUS High/critical findings indicating intentional harm or dangerous behavior

Report Format

══════════════════════════════════════════════
  SkillGuard Report: <skill-name>
══════════════════════════════════════════════
  Verdict:    ✅ SAFE / ⚠️ SUSPICIOUS / 🚨 MALICIOUS
  Files:      <count>
  Findings:   <count>
══════════════════════════════════════════════

Findings:

1. [SEVERITY] <title>
   OWASP: <Agentic Top 10 category>
   ATLAS: <MITRE ATLAS technique ID, if applicable>
   File: <path>:<line>
   Evidence: <snippet>
   Description: <explanation>

──────────────────────────────────────────────
Summary: <one-paragraph assessment>
Recommendation: <install / review manually / do not install>
══════════════════════════════════════════════

Severity Levels

  • 🔴 CRITICAL — Likely malicious; do not install
  • 🟠 HIGH — Dangerous pattern requiring investigation
  • 🟡 MEDIUM — Risky practice; review before installing
  • 🔵 LOW — Quality/hygiene issue
  • ⚪ INFO — Informational note