The Agent Skills Directory

[DATA_EXFILTRATION]: The vision.js script reads local files based on a user-provided path and transmits the base64-encoded content to a remote API. While the tool is intended for image recognition, the script's extension validation is permissive, defaulting to image/jpeg for unknown extensions. This creates a potential vector for exfiltrating sensitive non-image files if an agent is directed to process a sensitive file path.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting external data (images and text prompts) and returning the resulting analysis to the agent's context.
Ingestion points: imageSource and prompt CLI arguments in scripts/vision.js (file: scripts/vision.js).
Boundary markers: Absent. The user-provided prompt is interpolated directly into the API request payload.
Capability inventory: File system read access via fs.readFileSync and network communication via fetch (file: scripts/vision.js).
Sanitization: Absent. The script encodes image data in base64 but does not sanitize the text prompt or validate the contents of the image files.
[EXTERNAL_DOWNLOADS]: The script makes network requests to https://dashscope.aliyuncs.com. This is the official endpoint for Alibaba Cloud's DashScope vision service, a well-known service used here for the skill's primary function.

vision