← MCP server and skill grades

skills

antfu/skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
d524ed287eef
Files read
157

This project is a collection of skills for AI coding agents covering Vue, Vite, Nuxt, and pnpm documentation, along with workflows for opening pull requests, triaging GitHub issues, simplifying code changes, and scaffolding TypeScript projects. Its biggest risks are a fallback upload script that can send GitHub Enterprise tokens and screenshots to public GitHub servers, an issue-triage workflow that trusts unverified digest issues any stranger can open on a public repository, and project setup templates that rely on unpinned third-party release workflows and auto-linked npm package skills.

What it can reach

Files
Reads repository files, git history, and UI screenshots, creates temporary files and git worktrees, and edits project code and configuration when simplifying diffs or scaffolding projects.
Network
Communicates with GitHub APIs and uploads.github.com via gh and curl to read or update issues, labels, and pull requests and to upload PR screenshot attachments.
Shell commands
Instructs the agent to run git, gh, package manager commands, test/lint/build scripts, and the bundled upload-github-attachment.sh bash script.
Secrets
The bundled upload-github-attachment.sh script reads GH_TOKEN, GITHUB_TOKEN, or gh auth token to authenticate file uploads to GitHub.
Browser
The antfu-create-pr skill instructs the agent to launch browser tools such as Playwright, Storybook, or a local dev server to capture before/after UI screenshots.

Findings

4 medium3 low1 info
  1. MediumAttachment upload script hardcodes uploads.github.com and can leak GitHub Enterprise tokens and files

    skills/antfu-create-pr/scripts/upload-github-attachment.sh:78

    The fallback attachment upload script resolves the repository and authentication token through gh (which honors GH_HOST and GitHub Enterprise remotes), but unconditionally sends the bearer token and file bytes via curl to https://uploads.github.com/user-attachments/assets. If a user runs the skill in a self-hosted GitHub Enterprise repository on an older gh CLI, their Enterprise token and private screenshot are sent to public GitHub servers.

    Fix: Verify before uploading that GH_HOST is unset or github.com and that gh repo view --json url resolves to https://github.com/, refusing to run against other hosts.

  2. MediumIssue triage resume step trusts unverified digest issues that anyone can open

    skills/antfu-issue-triage/SKILL.md:24

    At the start of triage, the skill searches all open and closed issues matching "Triage digest in:title" without checking that the author is a repository maintainer or collaborator. On any public repository, a stranger can open an issue with "Triage digest" in the title to mark items as already checked, advance the recorded scan range so the agent skips newer issues and PRs, get their issue edited as the official digest, or plant prompt-injection text in content the skill frames as trusted prior state.

    Fix: Filter digest searches to trusted authors (such as --author @me or verifying authorAssociation is OWNER, MEMBER, or COLLABORATOR) and treat all issue bodies as untrusted text.

  3. MediumScaffolded GitHub Actions workflows use unpinned @main reusable workflows with write and OIDC publish permissions

    skills/antfu/references/setting-up.md:91

    When setting up a new project, the antfu skill instructs the agent to add CI, autofix, and release workflows that reference sxzz/workflows reusable workflows on the mutable @main branch rather than pinning a full commit SHA. In particular, release.yml grants contents: write and id-token: write (npm OIDC publishing) to an unpinned external workflow, exposing downstream projects if that upstream branch is ever compromised.

    Fix: Pin external reusable workflows in .github/workflows/*.yml to an immutable 40-character commit SHA instead of @main.

  4. MediumProject setup wires skills-npm into prepare to auto-link agent skills from installed npm packages

    skills/antfu/SKILL.md:160

    The antfu skill instructs the agent to run skills-npm inside the package.json prepare script so that every install symlinks agent skills shipped inside npm dependencies (and fetches skills declared in package manifests) into the coding agent's skills directory. Even when pnpm blocks untrusted dependency build scripts via allowBuilds, the root prepare script still runs, allowing a newly added or updated dependency to inject unreviewed skill instructions into the user's coding agent.

    Fix: Avoid running skills-npm automatically on every install in prepare, or use pnpm's explicit per-package permissions approval before linking third-party skills.

  5. LowGitHub token passed in curl command-line arguments is visible to local processes

    skills/antfu-create-pr/scripts/upload-github-attachment.sh:72

    The script passes the user's GitHub bearer token directly in argv via --header "Authorization: Bearer $token". While curl uploads the attachment, any other local user or process on a shared machine can read the token from the process table (ps or /proc/<pid>/cmdline).

    Fix: Pass the Authorization header to curl via standard input using --header @- or --config - so the token does not appear in process arguments.

  6. Lowpnpm CI guide recommends executing an unpinned remote install script

    skills/pnpm/references/best-practices-ci.md:19

    For CI environments outside GitHub Actions and official containers, the pnpm skill advises downloading the latest install.sh script from get.pnpm.io or GitHub releases and executing it directly with sh without pinning a release version or verifying a SHA-256 checksum.

    Fix: Pin a specific release artifact and verify its SHA-256 checksum before executing the installer script in CI.

  7. LowPredictable /tmp/pr-base worktree path enables local race conditions on shared machines

    skills/antfu-create-pr/references/visual-evidence.md:18

    When capturing before/after UI screenshots, the visual-evidence guide instructs the agent to check out the base branch into the fixed path /tmp/pr-base and then run the project's dev server or browser tests there. On a multi-user machine, another local user can pre-create an empty world-writable /tmp/pr-base directory to read private repository files or tamper with scripts before they run.

    Fix: Create the temporary worktree directory with mktemp -d instead of using a fixed /tmp/pr-base path.

  8. InfoPR creation skill publishes a summary of the user's session prompts in every pull request

    skills/antfu-create-pr/SKILL.md:67

    The antfu-create-pr skill requires every pull request body to include a "User prompts" section naming the AI model and harness and summarizing everything the user asked for during the coding session. Although it instructs the agent to omit secrets, neither the skill's frontmatter description nor README.md mentions this behavior, so users who discuss internal business context or private constraints in chat may not expect a summary of their prompts to be posted publicly on GitHub.

    Fix: Make the User prompts section optional or note it prominently in the skill description and README so users know session prompts will be summarized in PR descriptions.