The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/look.py utilizes subprocess.run() to execute macOS binaries screencapture and sips for capturing and resizing images. The commands are safely constructed using argument lists, which prevents shell injection vulnerabilities.- [DATA_EXPOSURE_AND_EXFILTRATION]: The primary function of the skill is to capture and store screenshots. Visual data from the user's screen is saved to local storage at /tmp/claude-screenshots/ or a specified path. No unauthorized data transmission or network exfiltration logic was detected in the scripts.- [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The script identifies a dependency on the pyobjc-framework-Quartz package. This is a legitimate and standard library used for interfacing with macOS system APIs.- [INDIRECT_PROMPT_INJECTION]: The skill presents an indirect prompt injection surface when listing or capturing windows. An attacker could potentially influence agent behavior by naming a window or website with malicious instructions.
Ingestion points: Window titles are retrieved and printed in scripts/look.py via the get_windows function.
Boundary markers: The output does not wrap external data in delimiters or include warnings to ignore instructions within titles.
Capability inventory: The agent is authorized to use Bash and Read tools.
Sanitization: No validation or filtering is applied to the window titles before they are displayed to the agent.

screenshot