The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled shell script scripts/grab-clipboard-image.sh to interact with a local Unix socket. The script utilizes standard system utilities including nc (Netcat) for socket communication and file for MIME-type detection.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes data from the user's clipboard, which is an untrusted external source. An attacker could potentially place an image on the clipboard containing visual instructions designed to manipulate the agent's behavior.
Ingestion points: Data is ingested through the Unix socket defined by the CLIPBOARD_SOCKET environment variable (defaulting to ~/.clipboard.sock) via the grab-clipboard-image.sh script.
Boundary markers: The instructions mandate that the agent describe the image content to the user before performing any actions, which serves as a human-in-the-loop verification step to mitigate accidental obedience to injected instructions.
Capability inventory: The skill has the capability to execute shell commands, read and write files to the local disk, and render images into the conversation context.
Sanitization: The script performs basic validation by checking the MIME type of the retrieved data using the file utility to ensure it is a valid image format (PNG, JPEG, etc.) before saving it with an appropriate extension.

clipboard-screenshot