The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions include a one-time download and installation of a ~200MB browser binary (cloakbrowser) from an external source using the npx command.
[REMOTE_CODE_EXECUTION]: The installation process involves executing code from an unverified source via npx js-reverse-mcp and cloning a repository from a public GitHub account (zhizhuodemao).
[DYNAMIC_EXECUTION]: The evaluate_script tool provides the capability to execute arbitrary JavaScript code within the execution context of a web page, which can be used to manipulate page state or access sensitive data.
[DATA_EXFILTRATION]: The skill provides tools like save_script_source and evaluate_script that can write captured remote content (scripts or execution results) to the local file system at user-defined paths, creating a potential path for data exfiltration.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests untrusted data from web pages through tools such as get_script_source, list_console_messages, and get_websocket_messages (defined in SKILL.md).
Boundary markers: Absent; no delimiters or "ignore instructions" markers are mentioned to prevent the agent from being influenced by data found on analyzed websites.
Capability inventory: High-impact capabilities include arbitrary JavaScript execution via evaluate_script, local file writing via save_script_source, and browser navigation via navigate_page (defined in SKILL.md).
Sanitization: Absent; no sanitization, filtering, or validation of ingested web content is documented.

js-reverse-mcp-debugging