GuardSkills — Security Scanner for Skills

Static analysis security gate. Scans skill files for malicious patterns and produces a risk decision before allowing installation.

Rule — ALWAYS scan before installing

Before installing any skill (ClawHub, GitHub, local, or skills.sh), you MUST run a guardskills scan first. Never run clawhub install, skills add, or copy skill files without scanning.

Scan by source

1. ClawHub skills

When the user asks to install a ClawHub skill, scan it first:

guardskills scan-clawhub <owner/skill-slug>

Also accepts full URLs:

guardskills scan-clawhub https://clawhub.ai/owner/skill-slug

2. Local skills

When loading or reviewing a skill folder on disk:

guardskills scan-local <path-to-skill-folder>

If the path contains multiple skills:

guardskills scan-local <path-to-skills-root> --skill <skill-folder-name>

3. GitHub skills

When installing from a GitHub repository:

guardskills add owner/repo --skill <skill-name> --dry-run

Also accepts full URLs:

guardskills add https://github.com/owner/repo --skill <skill-name> --dry-run

4. skills.sh skills

Use the same GitHub flow:

guardskills add owner/repo --skill <skill-name> --dry-run

If scan passes:

guardskills add owner/repo --skill <skill-name>

How to interpret results

Decision	Action
SAFE	Proceed with installation.
WARNING	Ask the user for confirmation before proceeding.
UNSAFE	Block. Tell the user the skill is unsafe and do not install.
CRITICAL	Block. Do not install under any circumstances.
UNVERIFIABLE	Block. Inform the user the skill could not be verified.

Useful flags

• --json — machine-readable JSON output
• --dry-run — scan only, do not install
• --ci — deterministic gate mode (no prompts, no install)
• --strict — use stricter risk thresholds
• --yes — auto-accept WARNING level (only if user explicitly requests)
• --force — override UNSAFE block (only if user explicitly requests)
• --allow-unverifiable — override UNVERIFIABLE block (only if user explicitly requests)

Exit codes

• 0 — allowed / safe
• 10 — warning not confirmed
• 20 — blocked (UNSAFE, CRITICAL, or UNVERIFIABLE)
• 30 — runtime / internal error

Recommended workflow

1. User asks to install a skill.
2. Determine the source (ClawHub, GitHub, local, skills.sh).
3. Run the appropriate guardskills scan command.
4. Read the decision from the output.
5. If SAFE, proceed with clawhub install or equivalent.
6. If WARNING, inform the user and ask for confirmation.
7. If UNSAFE, CRITICAL, or UNVERIFIABLE, block and explain why.
8. Never skip the scan step.

Notes

• guardskills is an additional security layer, not a replacement for manual review.
• A SAFE result means no known high-risk patterns were detected, not a guarantee of safety.
• The scanner checks for: credential exfiltration, remote code execution chains, destructive filesystem operations, privilege escalation, obfuscated payloads, and suspicious network activity.