The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/describe_image.py executes the macOS system utility sips using subprocess.run. This is used to downsample images larger than 5MB to ensure compatibility with the vision API. This utilizes a standard system tool for local processing.
[DATA_EXFILTRATION]: Local image files are read, base64 encoded, and transmitted to the external API endpoint https://open.bigmodel.cn/api/paas/v4/chat/completions. While this is the core functionality for providing vision capabilities, it involves sending local data to a third-party service provider.
[INDIRECT_PROMPT_INJECTION]: This skill serves as a significant ingestion point for untrusted external data. It processes local image files (screenshots, charts, photos) and accepts user-provided questions to describe them. There are no boundary markers or sanitization processes to prevent malicious instructions embedded within the image text from being transcribed and subsequently influencing the agent's logic.
Ingestion points: scripts/describe_image.py reads file paths from command-line arguments and file content from the local disk.
Boundary markers: Absent. The transcribed text is returned directly to the agent without delimiters or safety warnings regarding embedded content.
Capability inventory: The skill performs local file reads, executes the sips command, and makes outbound HTTPS requests.
Sanitization: Absent. The raw response from the vision model is printed to standard output for the agent to consume.

lov-describe-image