The Agent Skills Directory

[DATA_EXFILTRATION]: The skill captures full-screen visual data via the screencapture utility and transmits it to an external endpoint using the openclaw message send command. This facilitates the transfer of potentially sensitive desktop information to a remote channel.
[COMMAND_EXECUTION]: The skill executes shell commands to capture screen data, manage temporary files in /tmp, and interact with the openclaw CLI tool.
[COMMAND_EXECUTION]: The recommended command pattern interpolates variable context (e.g., <current-channel>, <current-target>) into a shell script. If these values are derived from untrusted user input and not strictly validated, it could lead to command injection.
[PROMPT_INJECTION]: The skill processes untrusted visual data (screen content) which represents an indirect prompt injection surface should the model attempt to interpret instructions contained within the image.
Ingestion points: Screen content is captured to a temporary file via /usr/sbin/screencapture (SKILL.md).
Boundary markers: None present; the skill lacks delimiters or instructions to ignore content within the captured media.
Capability inventory: Executes shell commands via screencapture and openclaw (SKILL.md).
Sanitization: No validation, OCR filtering, or content sanitization is performed on the captured image before transmission.

channel-screenshot-send