The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses node:child_process to manage the browser lifecycle and host environment. Specifically, start.js spawns Chrome/Chromium processes, stop.js sends termination signals to PIDs, and browser-processes.js executes the ps command to audit running browser instances. Additionally, start.js uses rsync to synchronize profile data.\n- [DATA_EXFILTRATION]: The skill can access sensitive user data by copying the local Chrome profile directory to an isolated cache (~/.cache/agent-web/browser/profile-copy) when the --profile flag is used. Furthermore, watch.js maintains detailed logs of all console messages, exceptions, and network requests (including URLs and metadata) in the local cache, which could contain sensitive information from browsing sessions.\n- [DYNAMIC_EXECUTION]: The eval.js script and the underlying cdp.js library facilitate the execution of arbitrary JavaScript code within the context of active browser tabs using the Runtime.evaluate CDP method. While this is a core feature, it allows for the execution of complex logic within the browser environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with external web pages, which are untrusted sources of information. \n
Ingestion points: Browser content extracted via eval.js, pick.js, or captured in screenshots.\n
Boundary markers: None identified in the skill's instructions or logic to separate untrusted web content from agent instructions.\n
Capability inventory: JavaScript execution, network logging, screenshot capture, and local file writing (logs).\n
Sanitization: The skill does not implement specific sanitization for data retrieved from web pages before it is returned to the agent.\n- [PERSISTENCE]: The start.js script automatically launches a detached background process (watch.js) to log browser activity. This process remains running independently of the initial command execution to provide continuous monitoring and logging.

web-browser