The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various system utilities to perform screen captures and manage application focus across different operating systems.
Evidence: scripts/take_screenshot.py invokes screencapture on macOS, and scrot, gnome-screenshot, or import (ImageMagick) on Linux.
Evidence: scripts/take_screenshot.py uses osascript to activate macOS applications before capture.
Evidence: The Python and shell scripts execute swift to run local Swift scripts for permission and window metadata retrieval.
[PRIVILEGE_ESCALATION]: Documentation within the skill encourages the agent to prompt for escalated permissions (e.g., sudo) to bypass system or sandbox-level restrictions.
Evidence: SKILL.md instructs: "If you see 'screen capture checks are blocked in the sandbox' ... rerun the command with escalated permissions."
Evidence: SKILL.md instructs: "If saving to the OS default location fails with permission errors in a sandbox, rerun the command with escalated permissions."
[INDIRECT_PROMPT_INJECTION]: The skill exposes a functional surface area that could be abused via indirect prompt injection to capture sensitive screen data without explicit user intent for that specific content.
Ingestion points: Parameters such as --app, --window-name, and --path are passed to scripts, potentially using values derived from untrusted external content (e.g., a website the agent is summarizing).
Boundary markers: The instructions lack specific guidance or delimiters to prevent the agent from being influenced by malicious instructions embedded in the data it is currently processing.
Capability inventory: The skill possesses file-writing capabilities and targeted window capture across multiple operating systems.
Sanitization: While the Python script uses subprocess.run with argument lists to mitigate shell injection, there is no application-level logic to validate the sensitivity of the target window or the output path against attacker-influenced inputs.
[DYNAMIC_EXECUTION]: The skill dynamically executes code from local source files rather than pre-compiled binaries.
Evidence: scripts/take_screenshot.ps1 uses the PowerShell Add-Type cmdlet to compile and execute C# code for interfacing with the Windows API (user32.dll).
Evidence: macOS functionality relies on interpreting .swift files at runtime via the swift command-line tool.

screenshot