The Agent Skills Directory

[COMMAND_EXECUTION]: The eval and evalraw commands in scripts/cdp.mjs allow the execution of arbitrary JavaScript code and raw Chrome DevTools Protocol (CDP) commands within the browser context, providing full programmatic control over the active session.\n- [DATA_EXFILTRATION]: The skill operates within the user's headful Chrome profile, granting access to active cookies, session tokens, and personal account data across all open tabs. Commands such as shot (screenshot), snap (accessibility tree), and html (page source) provide a direct path for the agent to retrieve sensitive information.\n- [PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection because it processes untrusted data from web pages that could contain instructions designed to manipulate the agent.\n
Ingestion points: Data enters the agent's context through snapshotStr (accessibility tree) and htmlStr (page HTML) in scripts/cdp.mjs.\n
Boundary markers: Absent. The skill does not implement delimiters or specific warnings to help the model distinguish between system instructions and data ingested from the browser.\n
Capability inventory: The skill exposes interactive actions including element interaction (click, clickxy), text entry (type), arbitrary code execution (eval), and navigation (nav).\n
Sanitization: Absent. Data from web pages is retrieved and passed to the agent without escaping or validation, allowing embedded instructions to be processed directly.

chrome-cdp