The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides tools for comprehensive GUI automation using pyautogui and xdotool, allowing the agent to click, type, and move the mouse across the operating system.
[COMMAND_EXECUTION]: Implements a bash tool that executes arbitrary shell commands via subprocess.run(command, shell=True). While it includes a basic sanitizer for patterns like 'rm -rf', it does not prevent all forms of malicious command execution.
[DATA_EXFILTRATION]: The skill captures full-system screenshots and accessibility snapshots, which are then base64-encoded and transmitted to an external vision-language model (Anthropic) for analysis.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing untrusted external data (such as web page screenshots or DOM trees) and allowing that data to influence high-privilege actions like shell command execution and filesystem modification.
Ingestion points: Browser content captured via Playwright DOM snapshots and full-screen screenshots in capture_screenshot() and get_page_snapshot().
Boundary markers: The system prompt instructs the agent to respond with a JSON object but lacks explicit instructions to ignore commands embedded within the visual or text data of the processed computer screen.
Capability inventory: Full GUI control via pyautogui, arbitrary shell execution via subprocess.run, network requests via requests, and file modification via _handle_editor.
Sanitization: Includes a rudimentary blacklist for bash commands but lacks validation for the logic or intent of GUI-based actions.
[EXTERNAL_DOWNLOADS]: The Docker environment fetches the ubuntu:22.04 base image and installs various desktop utilities and Python packages from official repositories.

computer-use-agents