← MCP server and skill grades

skills

addyosmani/agent-skills/skills

A
Type
Skill
Scanned
Nov 13, 2026
Version
301b02b12bff
Files read
33

This project is a collection of 23 engineering workflow skills and two small helper scripts that guide AI coding agents through planning, writing, testing, reviewing, and shipping software. The biggest risk is in its cross-model review instructions, which suggest passing code through echo or a shell heredoc in ways that can run embedded commands when reviewing untrusted code, alongside minor risks like an unpinned npx package setup and in-place file edits with test execution during code reviews.

What it can reach

Files
Guides the agent to read and edit project source files, configuration, git state, and markdown planning or constraint documents, and includes helper scripts that read JSON scan reports and create a docs/ideas folder.
Shell commands
Instructs the agent to run project build, lint, test, git, and security-scanner commands, install optional quality-gate packages after asking, and run external AI CLIs (codex or gemini) only with explicit user confirmation.
Network
Instructs the agent to fetch official framework documentation pages, push git branches or tags when releasing, and optionally send reviewed code snippets to external AI CLIs if the user opts in.
Browser
Guides connecting the agent to Chrome through the chrome-devtools MCP server (defaulting to an isolated temporary browser profile) to inspect pages, console logs, network requests, and screenshots.

Findings

4 low
  1. LowUnsafe shell piping and heredoc advice can execute code during cross-model review

    skills/doubt-driven-development/SKILL.md:134

    When an artifact being reviewed contains $(...) or backticks, Step 3 advises the agent to pass it via echo or a heredoc instead of inline arguments. In standard shells, both double-quoted echo and unquoted heredocs (<<EOF) still expand $(...) and backticks as shell commands before piping to the review tool, so reviewing unfamiliar or untrusted code could run embedded commands on your machine if the user approves the shell invocation.

    Fix: Remove the echo suggestion and explicitly require writing the prompt file with the agent's native file-write tool (without invoking a shell) or using a single-quoted heredoc delimiter (<<'EOF') so shell substitutions are never expanded.

  2. LowUnpinned npx @latest package in recommended Chrome DevTools MCP setup

    skills/browser-testing-with-devtools/SKILL.md:35

    The setup snippet configures the MCP server to run npx -y chrome-devtools-mcp@latest, which skips confirmation and automatically downloads and runs the newest version from npm at startup. If that upstream npm package were ever compromised, malicious updates would execute automatically whenever the browser MCP server starts.

    Fix: Pin chrome-devtools-mcp to a specific audited version number instead of using the floating @latest tag.

  3. LowHardcoded /tmp file path exposes reviewed code on multi-user systems

    skills/doubt-driven-development/SKILL.md:146

    The cross-model review examples tell the agent to write the prompt and reviewed code (often security-sensitive logic) to a fixed /tmp/doubt-prompt.md path in world-writable /tmp. Under default permissions, other local users on a shared machine can read the sensitive code left in /tmp, or pre-create the predictable file to cause collisions or tamper with the review input.

    Fix: Create a private temporary file with mktemp under restrictive permissions (umask 077) or inside a private user directory, and delete it right after the review command finishes.

  4. LowCode review skill edits live source files and runs the test suite

    skills/code-review-and-quality/SKILL.md:164

    Although users normally expect a code review to be read-only, Step 2 instructs the agent to edit the changed source file on disk to invert a condition, run the project's test suite, and restore the file afterward. Checking out an untrusted pull request to ask for a review will execute the contributor's test code, and an interrupted session can leave broken logic behind in your working tree.

    Fix: Document in the skill description that it mutates files and executes tests, ask before running tests on untrusted branches, and perform mutations in an isolated temporary worktree rather than editing the working tree in place.