← MCP server and skill grades

agent-browser

vercel-labs/agent-browser/skills/agent-browser

B
Type
Skill
Scanned
Nov 13, 2026
Version
2f533ec7d74c
Files read
1

This skill connects an AI agent to the agent-browser CLI so it can automate Chrome/Chromium browsers, cloud browsers, and Electron desktop apps like Slack and VS Code. Its biggest risk is that it pre-approves all agent-browser commands without user confirmation, meaning prompt injection from a visited webpage could trick the agent into taking unauthorized actions in logged-in sites or connected desktop apps.

What it can reach

Shell commands
Pre-approves running any agent-browser CLI subcommand without confirmation and instructs installing the CLI globally via npm.
Browser
Controls local Chrome or Chromium via the Chrome DevTools Protocol to visit URLs, click elements, fill forms, record videos or HAR traffic, and read page snapshots.
System
Automates local Electron desktop applications including VS Code, Slack, Discord, Figma, Notion, and Spotify.
Network
Connects to arbitrary websites as well as Vercel Sandbox microVMs, protected Vercel deployments, and AWS Bedrock AgentCore cloud browsers.
Secrets
Uses persisted login sessions, browser state, and an authentication vault to interact with authenticated websites and workspaces.

Findings

1 medium2 low
  1. MediumWildcard auto-approval for all agent-browser CLI commands

    skills/agent-browser/SKILL.md:4

    The skill frontmatter pre-approves every agent-browser subcommand so the agent can execute them without asking the user each time. Because agent-browser reads untrusted web pages into the agent's context, hidden text on a malicious page (prompt injection) could trick the agent into silently navigating to attacker URLs, submitting forms in logged-in sessions, sending Slack messages, or controlling connected Electron apps like VS Code.

    Fix: Restrict allowed-tools to read-only documentation lookups such as Bash(agent-browser skills:*) so browser navigation, form submission, binary installation, and desktop app automation still prompt the user for approval.

  2. LowOverrides built-in web tools that have tighter safety boundaries

    skills/agent-browser/SKILL.md:3

    The skill description instructs the agent to prefer agent-browser over any built-in web tools while hiding the skill from slash-command menus via hidden: true. Coding agents typically isolate built-in web fetch tools with per-domain prompts, read-only HTTP fetches, or separate summarization passes, whereas routing simple page reads through a local Chrome instance executes page JavaScript locally and feeds raw accessibility trees directly into the main agent context.

    Fix: Remove the blanket instruction to override built-in web tools so routine read-only URL fetches continue to use the host agent's safer built-in web fetch tool, reserving agent-browser for interactive browser tasks.

  3. LowUnpinned global npm package and binary install command

    skills/agent-browser/SKILL.md:12

    The setup instruction installs the latest version of agent-browser globally from npm and downloads browser components without pinning a specific version. If the npm package were ever compromised, a user or agent running this setup command would execute untrusted package install code and load altered runtime skill prompts via agent-browser skills get.

    Fix: Pin the installation command to a specific audited version (for example, npm i -g agent-browser@x.y.z) before running agent-browser install.