skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- a72290d14024
- Files read
- 66
This project is a collection of thirteen Markdown skills that guide AI coding agents in reviewing, building, stress-testing, and polishing user interfaces. The biggest risk is that interface-review can check out and run untrusted pull request code in a local /tmp worktree to render a preview, while afk and explain-interface can expose the agent to untrusted web or issue content and files in sibling project folders.
What it can reach
- Files
- Reads and edits project files, creates git branches and worktrees (including under /tmp), updates .git/info/exclude, and in the afk skill reads and copies code from sibling project directories on disk.
- Shell commands
- Instructs the agent to run git, GitHub CLI (gh), curl, and the project's local build, test, lint, and dev/preview server commands.
- Network
- Fetches web pages and stylesheets with curl, fetches pull request refs and metadata from git remotes, and reads connected Figma, analytics, error-tracking, and AI endpoints.
- Browser
- Opens local dev URLs and external websites, executes read-only DOM and CSS inspection functions, resizes pages, and captures screenshots through browser automation tools.
Findings
-
HighPull request review can check out and run untrusted fork code in /tmp
skills/interface-review/SKILL.md:103
Although the skill states that a change review is read-only and fetches pull requests from forks into refs/remotes/pr/<n>, it allows creating a local worktree at /tmp/review-<n> and rendering the app whenever the project exposes a cheap preview or the user asks for a rendered review. Rendering a pull request from an untrusted fork runs the project's preview or build toolchain against attacker-controlled config files and source code on the reviewer's computer, and uses a predictable path in shared /tmp.
Fix: Never create a worktree or start a local preview server for untrusted or fork pull requests without explicit confirmation from the user that untrusted code will run, and create temporary worktree directories with mktemp -d instead of a predictable /tmp/review-<n> path.
-
MediumStylesheet fetch instructions allow local/internal fetches and shell expansion
skills/explain-interface/no-browser.md:14
When inspecting a live URL without a browser (or when read-the-system.md falls back to this method for cross-origin stylesheets), the skill extracts .css links using href="[^"]*\.css[^"]*" and tells the agent to fetch each one the same way as curl -sL --max-time 25 "$URL". Because curl -sL permits non-HTTP schemes such as file:// and follows redirects to localhost, private LAN hosts, or cloud metadata services, a malicious page can trick the agent into fetching local files or internal URLs into style.css. In addition, the regex allows $(...) and backticks inside href, so interpolating an extracted URL inside double quotes can run shell commands.
Fix: Restrict curl to HTTP and HTTPS with --proto '=http,https' and --proto-redir '=http,https', reject localhost and private IP ranges, and pass validated URLs without double-quoted shell expansion.
-
MediumUnattended AFK mode combines pre-approved tools with untrusted external inputs
The user-invoked afk skill is meant to run unattended for hours and instructs testing dev servers, browsers, and connected services beforehand so nothing pauses for approval while the user is away. It then instructs the agent to read open issues, user feedback, 404 paths, HTTP referrers, and error logs—all writable by strangers on the internet—inside the same session that edits code, adds dependencies, and runs shell commands and curl. While the skill tells the model to ignore instructions in external text, prompt instructions alone are not a hard boundary against prompt injection once tool approvals are bypassed.
Fix: Avoid reading untrusted public inputs (issues, user feedback, raw referrers/404s, error strings) in the same unattended session that has pre-approved file-write, shell, and network tools, or require explicit opt-in per external source.
-
MediumAFK skill reads and copies code from sibling directories outside the current project
The afk skill description states that it works through one project while you are away, yet its instructions tell the agent to inspect sibling project directories on the user's machine and port code or dependencies from them into commits on the current repository. If a user keeps private client or employer repositories (or untrusted cloned repositories) in the same parent folder as the current project, running /afk can copy confidential code across project boundaries or read untrusted files without being asked.
Fix: Keep all file reads within the current repository root by default, and only inspect or copy from another project directory when the user explicitly names it.
