← MCP server and skill grades

skills

shadcn-ui/ui/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
e62c9435fff7
Files read
29

This project provides two AI agent skills for building, styling, and managing shadcn/ui components and migrating React projects from Radix UI to Base UI. The biggest risk is that the shadcn skill pre-approves all shadcn@latest CLI commands without confirmation and automatically runs an unpinned npx package on load, which could let prompt injection or a malicious project npm config overwrite files and run untrusted code.

What it can reach

Shell commands
Automatically executes npx shadcn@latest info --json when the shadcn skill loads, pre-approves wildcard shadcn@latest CLI commands via npx, pnpm dlx, and bunx, and instructs the agent to run package managers, git, TypeScript checks, and a Node coverage script.
Network
Downloads the unpinned shadcn@latest CLI and fetches component code, presets, and documentation from ui.shadcn.com, npm, GitHub, and configured registries.
Files
Reads workspace configuration and source files (including via scripts/showcase-coverage.mjs) and creates, modifies, or overwrites UI components, CSS files, and .migration/ reports in the project.

Findings

1 high1 medium
  1. HighOver-broad allowed-tools wildcard lets the agent install remote code and overwrite files without confirmation

    skills/shadcn/SKILL.md:5

    The shadcn skill activates automatically on matching projects (user-invocable: false) and pre-approves every command matching npx shadcn@latest *, pnpm dlx shadcn@latest *, or bunx --bun shadcn@latest *. Because the shadcn CLI's add, init, and apply subcommands accept arbitrary HTTPS URLs and GitHub repositories along with -y and --overwrite, a prompt injection inside fetched documentation, a third-party DESIGN.md, or an untrusted repository file can make the agent write arbitrary files into the project and install attacker-chosen npm dependencies without a user approval prompt.

    Fix: Narrow allowed-tools to read-only subcommands such as info, docs, search, and view so file-writing and package-installing commands still require explicit user confirmation.

  2. MediumSkill automatically downloads and runs unpinned npx shadcn@latest on load

    skills/shadcn/SKILL.md:17

    When the shadcn skill loads—which happens automatically whenever a conversation involves shadcn/ui or a project with components.json—the !npx shadcn@latest info --json preprocessing directive immediately runs in the current directory before the AI responds. Using the unpinned @latest tag (also repeated across CLI instructions in both skills) downloads and executes whatever version is currently published as latest on npm at runtime, and if opened inside a cloned untrusted repository containing a malicious .npmrc (such as a custom registry URL), npx can fetch and execute an attacker-controlled package without the user running npm install or approving a command.

    Fix: Avoid automatic shell execution during skill loading, pin a specific verified version of the shadcn package instead of @latest, and do not trust workspace npm registry overrides when running npx.