The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill enables an agent to browse the open web and ingest content via get_content, screenshot, and evaluate tools. This introduces a surface where malicious third-party content (e.g., hidden instructions in a webpage) could influence the agent's behavior.
Ingestion points: get_content, screenshot, evaluate (in SKILL.md)
Boundary markers: None detected in the skill logic; the documentation advises users to treat page text as data, not instruction.
Capability inventory: navigate, click, fill, evaluate, launch_browser (in SKILL.md)
Sanitization: None; the skill relies on the model's ability to distinguish external content from system instructions.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the anti-detect-browser (NPM) or antibrow (PyPI) packages. Upon first execution, these packages download a closed-source browser kernel binary (approximately 190MB to 320MB) from the vendor's CDN at antibrow.com.
[COMMAND_EXECUTION]: The evaluate tool provides the capability to execute arbitrary JavaScript code within the context of the active browser page. While intended for automation, this is a high-privilege function that could be abused if the agent is misled by malicious page content.
[DATA_EXFILTRATION]: The start_live_view tool generates a shareable URL that streams the browser's screen. If used during a logged-in session, this URL could expose sensitive user data or session cookies if the link is accessed by unauthorized parties.
[DYNAMIC_EXECUTION]: The skill facilitates the execution of JavaScript in the browser (evaluate) and relies on a dynamically downloaded binary kernel to perform its core spoofing functions.

browser-mcp-agent