← MCP server and skill grades

anti-detect-browser-skills

antibrow/anti-detect-browser-skills

B
Type
Plugin
Scanned
Nov 13, 2026
Version
b800e3a54d8d
Files read
11

This Claude Code plugin provides four Markdown skills that teach AI coding agents how to automate web browsing and multi-account scraping with the AntiBrow SDK and MCP server. Installing the plugin runs no code by itself, but following its instructions downloads and runs a closed-source Chromium binary that handles your live logins and can expose broad browser and screen-streaming tools to the agent.

What it can reach

Files
Installing the plugin adds Markdown skill files, while the AntiBrow SDK and CLI it instructs the agent to run read and write browser profiles, cookies, captured passkeys, cached scraping recipes, and the browser binary under ~/.anti-detect-browser/ and ~/.antibrow/.
Network
Connects to antibrow.com and its CDN to download the browser kernel, verify licenses daily, optionally sync profiles, and stream Live View sessions, as well as connecting to GitHub for recipes and any websites or proxies opened in the browser.
Shell commands
Instructs the AI coding agent to run npm, pip, npx, python, docker, and anti-detect-browser CLI commands to install packages, manage profiles, run isolation checks, and execute scraping recipes.
Secrets
Reads ANTIBROW_API_KEY or ANTI_DETECT_BROWSER_KEY and proxy credentials from environment variables or ~/.antibrow/license.key, and stores website cookies, session tokens, and captured passkeys inside local profile directories.
Browser
Launches and controls persistent fingerprinted Chromium sessions via Playwright, CDP, or MCP tools capable of navigating pages, clicking, filling forms, running arbitrary in-page JavaScript, taking screenshots, and streaming live browser screens.

Findings

2 medium3 low
  1. MediumClosed-source browser binary is downloaded at runtime without pinned build hashes

    anti-detect-browser/SKILL.md:366

    Following the skills installs an MIT-licensed wrapper SDK that downloads and executes a 190 to 320 MB closed-source Chromium binary from AntiBrow's CDN on first launch and phones home daily for license checks. Although the skill claims both artifacts are pinnable and verifiable, new profiles default to resolving the newest kernel build from a remote manifest by major version number rather than an immutable hash pinned inside the SDK, meaning a compromised vendor CDN or manifest could deliver untrusted native code even when the npm or PyPI package version is pinned.

    Fix: Run the closed-source browser inside an isolated VM or container, pre-warm a vetted kernel directory, and verify downloaded kernel binaries against immutable cryptographic hashes before execution.

  2. MediumDomain allowlist cannot stop third-party scraping recipes from modifying or leaking data on the same site

    multi-account-scraping/SKILL.md:132

    The scraping skill pulls community JavaScript recipes from antibrow/recipes and encourages running reviewed recipes inside logged-in profiles, stating that recipes only read data and are safe because meta.domains blocks requests to undeclared hosts. However, because recipe code executes directly inside the authenticated page context with full fetch and DOM access, host-level filtering cannot block state-changing POST or DELETE actions on that site or stop a compromised recipe from exfiltrating private account data by posting or messaging it to another account on the same allowed domain.

    Fix: Run third-party recipes only in throwaway --temporary profiles unless you have personally audited the exact recipe source code and verified it performs no writes on the allowed domain.

  3. LowRecommended MCP configuration exposes live screen-streaming and profile tools by default

    browser-mcp-agent/SKILL.md:53

    The .mcp.json setup snippet launches anti-detect-browser --mcp with all server tools enabled, including start_live_view (which streams the browser screen to a shareable URL), list_profiles, evaluate, and claim_proxy. Although the skill prose warns users to manually restrict tools in their MCP client, an unfiltered setup lets a prompt-injected agent open an existing logged-in profile, start a Live View screen stream, and leak the shareable stream URL or session data by navigating to an attacker-controlled page.

    Fix: Configure an explicit MCP client tool allowlist limited to basic browsing tools (launch_browser, navigate, get_content, screenshot, close_browser) and keep untrusted browsing in temporary profiles.

  4. LowUnpinned npx and package installation commands in skill instructions

    multi-account-isolation/SKILL.md:85

    Several skill files instruct the agent or CI pipeline to download and run unpinned packages from public registries, including npx liarjs for unattended CI checks in multi-account-isolation/SKILL.md, npm install -g anti-detect-browser and pip install antibrow in multi-account-scraping/SKILL.md, and pip install antibrow in anti-detect-browser/SKILL.md. Because AI agents load individual SKILL.md files on demand without necessarily reading README.md, following these commands resolves and executes whatever latest package version is currently published.

    Fix: Pin exact package versions and verify package integrity hashes for all npm, npx, and pip commands inside every SKILL.md file.

  5. LowPasskey credentials are captured into portable, syncable storage by default

    anti-detect-browser/SKILL.md:372

    In the launch options taught by the skill, webauthn_capture defaults to True, storing newly created WebAuthn passkeys in a portable profile store rather than an OS- or hardware-bound enclave so they travel with profile exports or cloud sync. Unlike session cookies, captured passkeys remain persistent cryptographic authenticators even after logging out, and the credential-audit documentation in multi-account-isolation/SKILL.md omits passkeys when listing sensitive artifacts stored on disk or synced to the cloud.

    Fix: Pass webauthn_capture=False when launching profiles for real accounts unless portable passkey storage is explicitly required, and avoid enabling sync=True on profiles holding sensitive passkeys.