find-skills
vercel-labs/skills/skills/find-skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- b716f9dc2ba3
- Files read
- 1
This skill guides your AI coding agent to search the skills.sh catalog and use the npx skills tool to discover and install third-party skills when you ask how to do tasks. Its biggest risk is downloading and running unpinned code from npm at runtime and installing third-party skills globally across all your projects without pinning versions or checking their files for harmful content.
What it can reach
- Shell commands
- Runs npx skills commands in the terminal to search for, install, update, and initialize agent skills.
- Network
- Connects to skills.sh, the npm registry, and GitHub to look up leaderboards, run CLI searches, and download skill packages.
- Files
- Writes downloaded third-party skill files into your global user-level or project-level agent configuration folders.
Findings
-
MediumGlobal, unprompted installation of unpinned third-party skills without content review
skills/find-skills/SKILL.md:100
Step 6 instructs the agent to install third-party skills globally across your user account (-g), skip confirmation prompts (-y), and pull from a mutable GitHub reference (<owner/repo@skill>) rather than a fixed commit. Step 4 also tells the agent to judge safety only by install counts and GitHub stars, which attackers can fake, without reading the skill files first. If an attacker tricks the agent via prompt injection or publishes a repository that turns malicious, installing it globally plants those instructions across all of your projects and future coding sessions.
Fix: Remove -g so skills install only into the current project, remove -y so confirmation prompts remain active, pin repositories to a specific commit SHA, and instruct the agent to read and review all files in a skill before installing it.
-
MediumDownloading and running an unpinned npm package at runtime
skills/find-skills/SKILL.md:27
Every CLI command in this skill runs npx skills without pinning an exact package version. Whenever the agent searches for, installs, or updates skills, npx can download and execute the latest published version of the skills package from npm on your computer. Because Step 3 runs npx skills find before you ever agree to install a skill, a supply-chain compromise of the skills npm package or its dependencies could execute malicious code on your machine during a routine search.
Fix: Pin the CLI to a vetted exact version (for example, npx skills@x.y.z) or require installing a pinned version locally beforehand instead of fetching latest code at runtime.
-
LowBroad trigger on ordinary how-to questions sends external queries and loads untrusted listings
skills/find-skills/SKILL.md:14
Because this skill activates whenever you ask a general question like how do I do X, routine coding questions can cause the agent to fetch https://skills.sh/ and run npx skills find without you asking to browse external packages. Aside from sending keywords from your request to external services, this pulls skill descriptions written by strangers in the open catalog into the conversation, exposing ordinary coding sessions to indirect prompt injection.
Fix: Narrow the trigger conditions so the skill activates only when the user explicitly asks to search for or install an agent skill, and treat third-party skill descriptions as untrusted data.
