computer-use-desktop

Warn

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions in references/lightpanda.md to download a binary from a third-party GitHub repository (lightpanda-io/browser). While the tool is a known headless browser, the source is not included in the defined list of trusted organizations.
  • [CREDENTIALS_UNSAFE]: The script scripts/cdpx_cookies.js uses the Chrome DevTools Protocol (CDP) to extract all session cookies from a running browser for specified domains. These cookies include sensitive authentication tokens (such as session IDs for LinkedIn and X). The extracted credentials are saved to a local file for use by other components of the skill.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to read and process data from external platforms like Discord, Twitter, and LinkedIn. This creates an attack surface where malicious content embedded in these platforms could attempt to influence the agent's behavior. The documentation includes explicit warnings to treat this information as data rather than instructions, which is a recommended mitigation.
  • [COMMAND_EXECUTION]: Multiple Python scripts (discord_deep.py, discord_read.py, tw100.py) utilize subprocess.run to execute the agent-desktop CLI tool. Other scripts use system utilities like ioreg and lsappinfo to monitor the environment.
  • [DYNAMIC_EXECUTION]: The skill uses ctypes in several Python scripts (ax-force-a11y.py, ax-probe.py, session-lock-check.py) to dynamically load and call functions from macOS system frameworks such as CoreGraphics and ApplicationServices. Furthermore, JavaScript components use Runtime.evaluate via CDP to run arbitrary code within browser sessions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 3, 2026, 10:26 AM