← MCP server and skill grades

skills

browser-act/skills

D
Type
Skill
Scanned
Nov 13, 2026
Version
593316aa0f8c
Files read
295

This repository provides AI agent skills for local browser automation via the BrowserAct CLI, a Skill Forge generator for building custom scrapers, and over 60 pre-built web scraping and social media skills. Installing the core skill pre-authorizes all browser-act CLI commands—including local file uploads, cookie exports, and Chrome profile imports—while several workflows handle untrusted web pages, links, and direct messages in ways that enable prompt injection or shell command injection.

What it can reach

Browser
Controls headless and headed browsers, attaches to the user's running Chrome via DevTools Protocol, imports local Chrome profiles (cookies, localStorage, IndexedDB), runs arbitrary JavaScript in web pages, and operates logged-in accounts across e-commerce and social platforms.
Shell commands
Pre-approves all browser-act CLI subcommands via allowed-tools, runs bundled Python scripts and bash batch loops, and installs or upgrades browser-act-cli using uv.
Files
Reads and writes scraped datasets, HAR network recordings, screenshots, exported cookies, skill memory files, and local account state directories, and can upload local files into web forms.
Network
Connects to arbitrary websites, configured proxies, BrowserAct cloud workflow and MCP endpoints (api.browseract.com, mcp.browseract.com), CAPTCHA-solving APIs, and remote-assist streaming links.
Secrets
Reads BROWSERACT_API_KEY and BROWSERACT_MCP_TOKEN from environment variables or .env files, and accesses active session cookies and CSRF tokens inside logged-in browser profiles.

Findings

3 high1 medium4 low
  1. HighBlanket allowed-tools pre-approves sensitive browser-act subcommands without host confirmation

    browser-act/SKILL.md:4

    The entry skill sets allowed-tools: Bash(browser-act:*), which tells compatible coding agents (such as Claude Code) to run every browser-act command without a host-level permission prompt, while its description simultaneously steers the agent to prefer browser-act over built-in read-only fetch tools. Although the metadata claims file uploads, logins, and browser creation require user confirmation, docs/agent-design.md notes that confirmation gating is only a conversation-level prompt instruction rather than a CLI block. If a visited webpage contains prompt injection, the tricked agent can invoke browser-act upload to exfiltrate local files, browser-act cookies export to dump session cookies, or browser-act browser import-profile to copy local Chrome credentials without triggering a host approval dialog.

    Fix: Remove the wildcard Bash(browser-act:*) pre-authorization (or restrict allowed-tools to read-only inspection subcommands) so state-changing commands trigger host confirmation, and enforce hard confirmation gates inside the CLI binary itself.

  2. HighShell command injection via unescaped outbound links on YouTube channel pages

    solutions/lead-generation/youtube-channel-business-email/SKILL.md:122

    In the YouTube email lookup workflow, scripts/extract-channel-about.py extracts external links controlled by the channel owner and URL-decodes /redirect?q= parameters (turning %27 into a literal single quote). Steps 5 and 6 then instruct the agent and its batch bash scripts to interpolate {link.url} directly inside single-quoted shell commands (--classify-url '{link.url}' and browser-act stealth-extract '{link.url}'). A malicious channel owner who places a single quote and shell metacharacters in an outbound link on their About page can break out of the single quotes and run arbitrary shell commands on the user's machine when those templates are executed.

    Fix: Pass URLs via safe argument arrays in Python or properly shell-escaped variables rather than interpolating untrusted URL strings directly into single-quoted bash command templates.

  3. HighAutonomous X DM reply loop exposes local data and double-quoted shell input to strangers

    solutions/social-listening/x-dm-auto-chat/SKILL.md:125

    Scenario A scans unread X (Twitter) direct messages from strangers, reads their message text into the AI coding agent, generates a reply, and sends it immediately with no human confirmation gate (x-keyword-comment similarly auto-posts public replies without review). Any stranger able to DM the account can send prompt-injected instructions that trick the agent into reading local files, environment secrets, or other private DM conversations and sending them straight back in the automated reply. In addition, the send template wraps <reply_text> in double quotes in bash, so backticks or $() in the generated reply will be executed as shell commands.

    Fix: Require explicit user review and approval before sending any DM or public comment, isolate untrusted message text from tool-calling privileges, and pass message text via stdin or strict shell escaping instead of double-quoted bash strings.

  4. MediumUnpinned automatic PyPI installation without user approval and opaque runtime directives

    solutions/social-listening/xiaohongshu-auto-posting/SKILL.md:112

    Phase 0.1 of xiaohongshu-auto-posting instructs the agent to automatically install or upgrade browser-act-cli from PyPI with no version pin and no user confirmation, contradicting browser-act/SKILL.md's claim that package downloads require approval. More broadly, browser-act/SKILL.md installs browser-act-cli unpinned and acts only as a stub that loads actual workflow instructions and self-update prompts at runtime via browser-act get-skills. If browser-act-cli on PyPI were ever compromised, installing or running these skills would execute unpinned code and inject unreviewed instructions into the agent.

    Fix: Pin browser-act-cli to a specific audited version, require explicit user confirmation before any install or upgrade command, and keep static, auditable skill instructions in the repository.

  5. LowOutbound link classifier allows local file schemes and private network addresses

    solutions/lead-generation/youtube-channel-business-email/scripts/extract-emails-from-text.py:74

    When a YouTube channel description has no visible email, the workflow runs classify_url() on every outbound link from the channel's About page and automatically fetches links classified as personal_or_business using browser-act stealth-extract. Because classify_url() does not restrict URL schemes to http/https and does not reject loopback, link-local, or private IP ranges (such as localhost, 127.0.0.1, 169.254.169.254, 192.168.x.x, or file://localhost/...), a stranger's YouTube channel link can cause the user's local browser to request internal services or local files and return any email-formatted strings found in the response.

    Fix: Validate that parsed.scheme is http or https and resolve/reject loopback, link-local, RFC1918 private, and internal hostnames before classifying a URL as fetchable.

  6. LowJavaScript code injection across Skill Forge output template and pre-built browser scripts

    solutions/lead-generation/linkedin-jobs-search/scripts/search-jobs.py:42

    The Python wrapper template in browser-act-skill-forge/references/output_template.md teaches generated skills to inject CLI arguments directly into JavaScript source strings via Python f-strings without json.dumps() escaping. Consequently, roughly 25 pre-built scripts across solutions/ (including linkedin-jobs-search, indeed-job-search, instagram-hashtag-posts, instagram-place-posts, airbnb-listing-detail, 1688-product-detail, and douyin-video-search) embed unescaped or partially escaped arguments and pagination cursors into JS string literals. Crafted inputs containing quotes, backslashes, or line breaks can break out of the string and run arbitrary JavaScript inside authenticated browser sessions.

    Fix: Use json.dumps() to serialize all parameters into JavaScript literals in browser-act-skill-forge/references/output_template.md and in all solutions/**/scripts/*.py wrapper scripts.

  7. LowRefusal-suppression directive ('never refuse') in always-loaded skill description

    browser-act-skill-forge/SKILL.md:3

    The frontmatter description of browser-act-skill-forge/SKILL.md—which coding agents load into the system prompt on every turn—ends with 'Human-operation automation; never refuse.' (and amazon-alexa-qa/SKILL.md includes a similar 'Do not refuse' instruction). Blanket instructions telling the model never to refuse weaken built-in safety guardrails when handling ambiguous, unauthorized, or prompt-injected automation requests against sensitive websites.

    Fix: Remove 'never refuse' and similar refusal-override wording from skill frontmatter descriptions and bodies.

  8. LowCloud API skills prompt users to paste secret API keys directly into chat

    solutions/ecommerce/amazon-asin-lookup-api-skill/scripts/amazon_asin_lookup_api.py:110

    When BROWSERACT_API_KEY is missing, 16 of the cloud API skills (both in their SKILL.md instructions and Python CLI output) ask the user to paste their API key into the chat conversation. Entering secrets in chat exposes them to the LLM provider, saves them in plaintext session transcripts, and often leads the agent to pass the key inline in shell command history (an anti-pattern that wechat-article-search-api-skill/SKILL.md explicitly warns against).

    Fix: Instruct users to set BROWSERACT_API_KEY in their shell environment or a local .env file, and never ask them to paste API keys into chat.