computer-use

Warn

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: MEDIUMPRIVILEGE_ESCALATIONPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PRIVILEGE_ESCALATION]: The scripts/setup-vnc.sh script utilizes sudo to perform high-privilege system modifications, including installing packages via apt, copying service configuration to /etc/systemd/system/, and replacing the system binary /usr/bin/xfdesktop with a shell script wrapper.
  • [PERSISTENCE]: The skill establishes persistence by installing four systemd services (xvfb.service, xfce-minimal.service, x11vnc.service, novnc.service) that are configured to automatically start at boot and restart upon failure.
  • [COMMAND_EXECUTION]: Through the use of xdotool in various scripts (e.g., scripts/type_text.sh, scripts/key.sh), the agent is granted full interactive control over the desktop environment, allowing it to execute arbitrary commands within terminals or GUI applications.
  • [INDIRECT_PROMPT_INJECTION]: The skill's architecture is inherently vulnerable to indirect prompt injection due to its reliance on interpreting visual output.
  • Ingestion points: Screen captures are generated by scripts/screenshot.sh and scripts/zoom.sh and returned to the agent as base64-encoded images.
  • Boundary markers: None identified; there are no clear delimiters to separate trusted system output from untrusted content displayed in browsers or text editors.
  • Capability inventory: Full GUI control including terminal access, file management, and browser interaction via xdotool scripts.
  • Sanitization: No filtering or sanitization is performed on the visual data before processing.
  • [EXTERNAL_DOWNLOADS]: The skill's documentation in SKILL.md provides instructions to fetch and install the Google Chrome browser package directly from dl.google.com using wget and dpkg.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 4, 2026, 03:03 AM