Every skill you install is code that runs with your agent's permissions — its file system, its API keys, its shell. That's the whole point, and also the whole risk. Public skill registries have already seen malicious-skill campaigns, where a useful-looking skill quietly exfiltrates secrets or runs commands you never approved. The fix isn't to avoid skills; it's to vet them first.
That's why a security skill belongs at the bottom of your install list chronologically but the top of your priorities — you want it in place before you start adding the fun stuff.
Vetting on OpenClaw
OpenClaw's community has built a solid security layer on ClawHub. The standouts:
- skill-vetter — security-first vetting of a skill before you trust it.
- SkillScan — scans skills for risky patterns.
- arc-security-audit — a comprehensive audit of your agent's full skill stack.
- agent-skills-audit — a two-pass, multidisciplinary code audit.
openclaw skills install skill-vetterRun it against anything before granting real permissions. Because OpenClaw skills are downloaded from an open registry, this scan is your main line of defense.
Vetting on Hermes
Hermes pushes some of this up into the distribution layer. Skill hubs such as HermesHub run automated security scanning against a large threat-detection ruleset before a skill is listed, and add an agent-to-agent trust-scoring system so skills carry a reputation signal. You still review what runs, but the hub does a first pass for you.
OpenClaw vs Hermes at a glance
- OpenClaw — install a vetting skill (skill-vetter / SkillScan) and scan before trusting.
- Hermes — hub-level automated scanning plus trust scoring, then your review.
This matters most when you also run a self-improving agent — an agent writing its own skills should never run them unreviewed.
Running it from your phone
The strongest control is a human approval step, and it doesn't have to keep you at a desk. Onepilot puts that step on your phone: review a new or updated skill over SSH and greenlight it before it runs with real access, on either OpenClaw or Hermes.