computer-use-agents

Pass

Audited by Gen Agent Trust Hub on Oct 31, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides patterns for direct computer control using tools like pyautogui and xdotool in references/patterns.md. These tools allow an AI to perform clicks, type text, and press keys, which can be misused if the agent is redirected to sensitive applications or interfaces.
  • [DYNAMIC_EXECUTION]: The bash tool implementation in references/patterns.md uses subprocess.run(command, shell=True) to execute arbitrary shell commands provided by the vision-language model. While the code includes a blacklist of dangerous patterns (e.g., rm -rf), this remains a high-privilege execution surface.
  • [REMOTE_CODE_EXECUTION]: The bash tool and browser automation capabilities (playwright) create a surface where an agent could be instructed to download and execute remote scripts or payloads, particularly if browsing untrusted websites.
  • [EXTERNAL_DOWNLOADS]: The Dockerfile pattern in references/patterns.md uses pip3 install -r /tmp/requirements.txt to fetch dependencies, and the SandboxedAgent class uses requests.post to communicate with the agent's API within the sandbox.
  • [INDIRECT_PROMPT_INJECTION]: The skill documents the risk of and patterns for agents processing untrusted external data.
  • Ingestion points: Screenshots of the desktop/web pages (pyautogui.screenshot, scrot) and DOM snapshots (page.accessibility.snapshot) in references/patterns.md.
  • Boundary markers: System prompts in ComputerUseAgent.run and BrowserUseAgent.run_with_llm that define the agent's role and expected JSON action format.
  • Capability inventory: GUI control (mouse/keyboard), shell command execution via bash tool, and file editing via str_replace_editor documented across references/patterns.md.
  • Sanitization: Implementation patterns for injection scanners (scan_for_injection), confirmation gates (ConfirmationGate class), and command blacklists provided in references/sharp_edges.md and references/patterns.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 31, 2026, 05:27 AM