Skills

OpenClaw Skills Security: a security-audit module to run before you install a skill

For a six-step check of a third-party skill with skill-auditor before you install it into an agent, returning SAFE/SUSPICIOUS/DANGEROUS/BLOCK — provided the modules do not run themselves and must be pasted into a host or symlinked for Codex.

Official source · openclaw-skills-security

Install

ln -s "$PWD/skills/skill-auditor" ~/.codex/skills/skill-auditor

Who it's for

Use these instruction modules maintained by UseClawPro (UseAI.pro) when you need a security-oriented check of skills in the OpenClaw ecosystem, especially worries about prompt injection, dependencies, or credential leaks. End users get two audit skills, plus 11 reusable modules; the first time, use only the audit skills, and do not assemble names for those 11 modules yourself.

Hard precondition: these are SKILL.md instruction modules and will not run themselves. They must be loaded into a host agent (Codex CLI / Claude Code / OpenClaw) or pasted into any LLM conversation. Without a host, and with no plan to paste, do not expect a double-click to work.

The notes give a global symlink command only for Codex CLI. Claude Code and OpenClaw are named as loadable hosts, but their install paths are not written; do not invent ~/.claude or other directories.

First try in 3 minutes

Official page:

https://github.com/useai-pro/openclaw-skills-security

The first time, paste; do not change this machine's skill directories. Open skills/skill-auditor/SKILL.md, paste it into the current agent or any LLM; paste the target skill's SKILL.md as well; then ask only: Audit this skill. Return a SKILL AUDIT REPORT.

Do not let it install the skill under review, and do not let it change the repository. Success signal: it returns a SKILL AUDIT REPORT and finishes the six-step protocol — metadata & typosquat, permission, dependency, prompt injection, network & exfiltration, content red flags — and the verdict is one of SAFE / SUSPICIOUS / DANGEROUS / BLOCK.

If the report tells you to “install into the production agent at once”, stop first and read the verdict yourself.

Use cases

Case 1: Paste-audit one target skill. Follow the three steps above to get a SKILL AUDIT REPORT. Check that all six steps and all four verdict grades are present. The notes give no other prompt sentence; do not change the passphrase.

Case 2: Global symlinks in Codex CLI. This changes ~/.codex/skills/. Confirm the directory may hold links, then run the two lines in the notes:

ln -s "$PWD/skills/skill-auditor" ~/.codex/skills/skill-auditor

ln -s "$PWD/skills/setup-auditor" ~/.codex/skills/setup-auditor

After they run, open those two target paths and confirm the links point at the matching directories in this repository. setup-auditor has only the link command, no prompt sentence; do not invent its audit passphrase. Do not apply ~/.codex/skills/ to other hosts.

Boundaries

Modules do not run themselves. Install is evidenced only as paste, plus the two Codex ln -s lines.

The audit protocol is evidenced only as those six steps and four verdict grades. Do not add another scoring system.

The 11 reusable modules are only counted; there is no name roster. Claude Code / OpenClaw have no install commands.

The notes give no language coverage or pricing. Keep undocumented items unknown.