The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions and code samples for processing external images via the Vision API, creating a surface for indirect instructions embedded within visual content.
Ingestion points: Samples in samples/vision-image-understanding.md and scripts/vision-input.md demonstrate ingesting images from public URLs (e.g., api.nga.gov) and user-provided base64 strings.
Boundary markers: The example prompts do not implement explicit boundary markers or instructions for the model to ignore potential directives embedded in the visual media.
Capability inventory: The skill demonstrates file system write operations (fs.writeFileSync, f.write) and network communication with the OpenAI API across multiple sample scripts.
Sanitization: No specific validation or sanitization of image content is performed before processing, which is standard for the described API usage but constitutes a vulnerability surface.
[COMMAND_EXECUTION]: The skill provides numerous shell command examples for interacting with external APIs and processing responses.
Evidence: scripts/image-generation.md, scripts/video-generation.md, and scripts/vision-input.md include curl commands that pipe sensitive API output to jq and base64 for local file creation.
[EXTERNAL_DOWNLOADS]: The skill references and demonstrates fetching resources from external network locations.
Evidence: samples/vision-image-understanding.md and scripts/vision-input.md fetch sample images from the National Gallery of Art's public API (api.nga.gov).
Evidence: The skill makes standard calls to the official OpenAI API domain (api.openai.com) for generation and analysis tasks.

openai-images-video