Are agent skills safe? What to check before installing one
Not automatically. A skill is instructions plus optional scripts, and your agent follows and runs them with your permissions, so treat an unfamiliar skill like an unfamiliar dependency: read it, look at what it executes and calls over the network, and pin it to a source you can check. A scanner helps but does not replace reading the scripts/ folder.
What to check before you install
- Read
SKILL.mdfor instructions that hide actions from you or tell the agent to send data somewhere. - Open
scripts/and anypackage.json. Look forcurl | sh, postinstall scripts, credential-file reads, and network calls. - Look for invisible characters. Hidden Unicode in rules files has been used to smuggle instructions past reviewers.
- Check where it came from: the source repository, the commit, the licence. Install counts and names are not trust signals.
- Check what it asks for: tool permissions, MCP servers, hooks.
- Re-check after updates. An approval is a point in time; a later version can differ.
What we found in popular skills
On 2026-10-29 we scanned 511 unique skills from 34 repositories linked from the skills.sh leaderboard: 96 ship executable scripts, 85 make network calls from those scripts, 33 fail the Agent Skills format, and we confirmed 0 malicious ones by hand-reading every critical flag. Read that as "0 found by this method", not "safe": pattern matching is not a sandbox. Full method and numbers are on the proof page linked below.
Outside research is less comfortable. Published work reports malicious skills in large public collections and scanner bypasses, and a hidden-Unicode attack on rules files; see the sources.
A worked example
Skillhawk's secure command audits the config around skills too. This is trimmed output from the repository's own deliberately bad test fixture:
$ skillhawk secure test/fixtures/secure-bad
HIGH .claude/settings.json:2 permission "Bash(*)" lets the agent run any shell command unprompted
HIGH .claude/settings.json:6 hook: downloads and runs remote code in one step
HIGH .mcp.json:5 MCP server "db" has a secret inlined in DB_PASSWORD; it lands in git and backups
MEDIUM .mcp.json:3 MCP server "files" runs @example/server-filesystem unpinned
The Skillhawk command that helps
Run these on a skill folder before you install it, and in CI for your own:
npx skillhawk scan <dir> # security-scan every skill found under a folder
npx skillhawk secure <dir> --gate high # also MCP configs, hooks, dependencies, remote scripts; non-zero exit fails CI
npx skillhawk clean scripts <skill-dir> # review the scripts a skill ships
These are static checks. They can miss a careful attacker, and they do not watch a skill after you install it.
Data behind this: the 511-skill scan.
Questions
Can an agent skill be malicious?
Yes. Skills are instructions and scripts the agent acts on with your permissions. Research reports malicious skills in public collections; our own scan of 511 popular skills found none by hand-review, with the limits of pattern matching stated.
What should I check before installing a skill?
Read SKILL.md and everything in scripts/, look for network calls, curl-pipe-shell and hidden Unicode, verify the source repository and commit, and run a scanner such as skillhawk secure.
Sources
- ReversingLabs: Weaponizing AI coding (Rules File Backdoor)
"The attack leverages configuration files that are usually trusted and overlooked during code reviews." - arXiv 2602.06547: "Do Not Mention This to the User" (large-scale scan of public skills)
- OWASP Agentic Skills Top 10
- Cloud Security Alliance: MCP security, tool poisoning
"rug pulls exploit the gap between initial approval and ongoing verification"