← MCP server and skill grades

skills

liarjsdev/liarjs-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
4068c7920975
Files read
7

This collection of four skills teaches an AI coding agent how to audit browser fingerprints for bot-detection inconsistencies, explain failing checks, and add fingerprint checks to Playwright suites or CI pipelines using the liarjs tool. The biggest risk is that the skills grant broad shell access and download a floating version range (liarjs@0.3) from npm at runtime, so a compromised package update or a prompt injection could run unwanted commands on your computer.

What it can reach

Shell commands
Runs shell commands via Bash, primarily npx liarjs@0.3 and npm install --save-dev liarjs.
Browser
Launches a local Chrome, Chromium, or Edge browser with a temporary profile, or attaches over Chrome DevTools Protocol (--cdp) to an already-running browser session when asked.
Network
Downloads the liarjs package from npm and has the tested browser fetch https://liarjs.dev/api/net.json by default (unless --offline or a custom --endpoint is used) to inspect IP, ASN, TLS, and HTTP headers.
Files
Reads project and report files, and writes or edits JSON scan baselines (scan.json, baseline.json), CI configuration files, and browser test files.

Findings

1 medium2 low
  1. MediumDownloads and executes a floating npm package version at runtime

    skills/browser-fingerprint-audit/SKILL.md:20

    None of the executable liarjs code is included in this repository; instead, all four skills instruct the agent to run npx liarjs@0.3 (or an unpinned npm install --save-dev liarjs). Although the CI skill refers to liarjs@0.3 as pinned, 0.3 in npm matches any new 0.3.x release, so npx can silently pull and execute newly published code on the user's machine if the upstream npm package is compromised.

    Fix: Pin liarjs to an exact version (such as 0.3.0) in package.json with a committed lockfile and integrity hash, and run the locally installed binary instead of fetching a floating range via npx.

  2. LowUnrestricted Bash permission in skill frontmatter

    skills/browser-fingerprint-audit/SKILL.md:5

    All four SKILL.md files request bare Bash in allowed-tools (and two also request bare Edit and Write), which pre-approves any shell command in agents that enforce skill tool allowlists. Because the agent may inspect untrusted scan.json artifacts or attach to pages during troubleshooting, a prompt injection could execute arbitrary shell commands without a confirmation prompt even though the skills only need to invoke liarjs.

    Fix: Restrict allowed-tools to the specific command prefixes each skill requires, such as Bash(npx liarjs@0.3 *), instead of granting unrestricted Bash.

  3. LowBaseline and artifact instructions can publish host IP and hardware metadata

    skills/fingerprint-ci-gate/SKILL.md:70

    Unless --offline is passed, liarjs writes both raw client and server objects into baseline.json and scan.json, including the scanning machine's public IP address, ASN, Cloudflare region, WebRTC candidates, GPU model, and installed OS fonts. The CI skill instructs committing baseline.json to git history and uploading scan.json as a build artifact without warning that this exposes the developer's or private runner's IP address and device fingerprint.

    Fix: Warn that baseline.json and scan.json contain the host's public IP address and raw device fingerprint, recommend --offline when network checks are unnecessary, or strip raw server/client network fields before committing baselines to shared repositories.