The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including images and PDFs fetched from URLs or local files, creating a surface for indirect prompt injection attacks where instructions could be embedded within the visual or document content.
Ingestion points: Untrusted data enters the agent context through httpx.get(image_url), open("image.jpg", "rb"), and open("document.pdf", "rb") as demonstrated in the Python code patterns.
Boundary markers: The provided instructions and code snippets do not include explicit boundary markers or warnings for the agent to ignore instructions embedded within the visual content.
Capability inventory: The skill allows access to high-privilege tools including Bash, Read, Write, Edit, and WebFetch, which could be exploited if an indirect injection successfully overrides agent behavior.
Sanitization: No specific sanitization, filtering, or validation of the content within the images or PDFs is described.
[EXTERNAL_DOWNLOADS]: The skill utilizes the WebFetch tool and the httpx Python library to download images from remote URLs for analysis. This behavior is documented as a core feature for multimodal processing.

vision-multimodal