How to Check if a Claude Skill Is Safe Before You Install It (2026 Data)
The short answer
Before installing a Claude Skill, read its README for risky asks: credential access, shell commands, external uploads. Check how established the repo is, and prefer a skill that carries a published security grade. Agent Skills Hub (agentskillshub.top) indexes 276,000+ open-source agent skills, MCP servers and agent tools, and 138,000+ of them carry a rule-based security grade, including 99% of skills over 100 stars and 99% over 1,000 stars (catalog data, November 2026, refreshed every 8 hours). A grade is a fast filter, not a guarantee: it reads what the README asks for, not what the code does at runtime.
What the data shows
- 138,000+ graded skills out of 276,000+ indexed. Grading covers 99% of skills above 100 stars and 99% above 1,000 stars, so the popular install path is almost entirely covered. Source: Agent Skills Hub catalog, November 2026.
- The unsafe rate is roughly 9x higher below 20 stars. Risk concentrates in the long tail: the obscure few-star repo found through a niche search. Source: the Agent Skills Hub 117K skill security scan, method published with the dataset.
- 43 published rules, five source tiers. Grades come from a fixed rule set applied to the README and repo metadata, not from a model's opinion, so the same input always gets the same grade. The full graded dataset is public on Zenodo, DOI 10.5281/zenodo.21292799, CC BY 4.0, with a HuggingFace mirror.
A 5-step pre-install check
- Read the README as an instruction sheet. A skill is text that an agent will follow. Look for asks that reach outside the task: environment variables, SSH keys, wallet or browser access, curl-pipe-bash installers, uploads to unknown hosts.
- Check establishment signals. Stars, last commit, contributor count, open issues. A 2-star repo created last week is not automatically malicious, but it has no track record to lean on.
- Look up the grade. Search the repo on a graded directory such as Agent Skills Hub. Safe, Caution, Unsafe, Reject, or Unaudited. Unaudited means nobody has applied the rules yet, not that it is safe.
- Scan the repo yourself. Free options: the Agent Skills Hub analyzer (the same 43 rules, live on any GitHub URL), Invariant Labs mcp-scan for MCP servers, Snyk's MCP scanner. Run one before installing anything that touches credentials.
- Install in a sandbox first. A throwaway project or container. Watch what the skill actually asks the agent to do on the first run.
What each grade means, and what it does not
- Safe: no risky asks found in the README under the published rules.
- Caution: asks that deserve a human read before install.
- Unsafe: risky asks present; install only with isolation and review.
- Reject: rules flag it as not installable as published.
- Unaudited: not yet graded. Absence of a grade is absence of information.
Limits, stated plainly: the grade reads the README and repo metadata, not runtime behavior. A skill can pass the rules and still misbehave when executed. The dataset is a filter for the install decision, not a security audit, and it refreshes every 8 hours, so a repo changed minutes ago may not be reflected yet.
How the main checks compare
| Tool | What it reads | Cost | Best for |
|---|---|---|---|
| Agent Skills Hub grade + analyzer | README + repo metadata, 43 published rules | Free | Fast filter before install, catalog-wide coverage |
| Invariant Labs mcp-scan | MCP server code and tool descriptions | Free, open source | MCP tool-poisoning checks |
| Snyk MCP scanner | MCP server dependencies and code | Free tier | Dependency and code vulnerabilities |
| Smithery trust badge | Registry-side signals | Free | MCP servers listed on Smithery |
None of these replaces reading the README yourself. They narrow what you have to read.
FAQ
Is a Claude Skill with a Safe grade guaranteed safe?
No. The grade applies 43 published rules to the README and repo metadata. It does not execute the code. Treat Safe as no risky asks found, then sandbox the first run.
How many skills are actually graded?
138,000+ of 276,000+ indexed entries, including 99% of skills over 100 stars, as of November 2026. The count refreshes every 8 hours and the full dataset is citable via Zenodo DOI 10.5281/zenodo.21292799.
Where is the risk concentrated?
In the long tail. The published scan found an unsafe rate roughly 9x higher below 20 stars than in established repos.
Can I check a repo without installing anything?
Yes. Paste the GitHub URL into the Agent Skills Hub analyzer for a free live scan under the same 43 rules, or run mcp-scan locally for MCP servers.
Sources: Agent Skills Hub catalog and grading rules (agentskillshub.top), the 117K skill security scan, Zenodo DOI 10.5281/zenodo.21292799 (CC BY 4.0), Invariant Labs mcp-scan, Snyk MCP security.