chrome-cdp-access

Fail

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [DATA_EXPOSURE]: The skill instructs the agent to copy the user's active Chrome profile directory, including sensitive files like 'Local State' and session data (cookies, history), to a hidden directory in the home folder ('~/.hermes/chrome-cdp') to bypass SingletonLock and access the user's authenticated sessions.
  • [DATA_EXFILTRATION]: The script 'scripts/cdp_probe_403.py' captures and prints sensitive request headers, specifically including session cookies, while navigating the browser to a hardcoded external Cloudflare tunnel URL ('https://tions-becomes-affecting-zus.trycloudflare.com/'). This exposes the user's authentication context to the agent and potentially the remote service endpoint.
  • [REMOTE_CODE_EXECUTION]: The skill enables the Chrome DevTools Protocol on port 9222 and explicitly instructs the agent to disable security origin checks using the '--remote-allow-origins="*"' flag. This allows any local process to interact with and execute code within the browser session, bypassing standard browser security boundaries.
  • [DYNAMIC_EXECUTION]: Multiple scripts ('cdp_eval.js', 'cdp_do.js', etc.) allow for the dynamic evaluation of arbitrary JavaScript within the browser context via the CDP 'Runtime.evaluate' method. This allows the agent to generate and run code that can interact with any web page the user has open.
  • [INDIRECT_PROMPT_INJECTION]: By reading DOM content, titles, and console logs from arbitrary web pages via CDP, the skill creates an attack surface where malicious content on a website could influence the agent's behavior or decision-making process.
  • [COMMAND_EXECUTION]: Uses AppleScript ('osascript') to forcefully quit the user's browser and shell commands to manage files and launch processes with dangerous flags.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 3, 2026, 10:26 AM