The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements desktop automation using the pyautogui library, which allows the AI agent to perform mouse movements, clicks, and keyboard inputs on the host system.
[REMOTE_CODE_EXECUTION]: The AnthropicComputerUse implementation includes a bash tool (BetaToolBash20241022), enabling the agent to execute arbitrary shell commands within its environment.
[DATA_EXFILTRATION]: The skill captures full-screen screenshots using pyautogui.screenshot() and scrot, which are then sent to the vision model for processing. This presents a risk of exfiltrating sensitive or private information visible on the user's desktop.
[COMMAND_EXECUTION]: The agent uses subprocess.run to call system binaries like scrot for taking screenshots.
[REMOTE_CODE_EXECUTION]: The skill includes the str_replace_editor tool, which grants the agent the ability to view and modify files on the filesystem.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its ingestion of untrusted data.
Ingestion points: Screen captures and outputs from shell commands.
Boundary markers: Absent; there are no clear delimiters used to separate user instructions from potentially malicious content found on-screen.
Capability inventory: Full desktop control via pyautogui, shell access via bash, and file system access via the text editor tool.
Sanitization: Absent; the skill does not provide mechanisms to filter or validate text or instructions encountered during screen analysis.

computer-use-agents