The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (HIGH): The installation instructions require downloading a ZIP file from an untrusted GitHub repository (femto/mcp-chrome) and loading it as an 'unpacked' extension in Developer Mode. This method bypasses the automated and manual security reviews conducted by the Google Chrome Web Store, making it a primary vector for delivering malicious code.
[DATA_EXFILTRATION] (HIGH): The skill is designed to operate within the user's existing Chrome profile. This grants the agent access to sensitive information including:
Active Sessions: The chrome_network_request tool can perform HTTP requests using the user's live cookies and authentication headers.
Personal Data: The chrome_history and chrome_bookmark_search tools allow searching through private browsing data.
Sensitive Content: chrome_get_web_content and chrome_screenshot can capture any information visible on the screen, including banking details or private messages if those tabs are open.
[COMMAND_EXECUTION] (MEDIUM): The skill requires a global NPM installation (npm install -g mcp-chrome-bridger). Global packages can execute arbitrary code on the host machine during installation via post-install scripts and have persistent access to the command line.
[PROMPT_INJECTION] (LOW): This skill is highly susceptible to Indirect Prompt Injection. Because it reads live web content (chrome_get_web_content), an attacker could place hidden instructions on a website that command the AI to use its interaction tools (chrome_click_element, chrome_fill_or_select) to perform unauthorized actions on other open tabs using the user's active login sessions.

browser-automation