The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an untrusted data ingestion surface in scripts/image_search.py where it fetches external image metadata (titles, source descriptions, and URLs) from DuckDuckGo search results. There are no explicit boundary markers or instructions in the output JSON to declare the content untrusted to the downstream agent. The capability inventory includes network operations via requests.get and local file writing via open. However, proper sanitization is present; file paths are generated safely using structured incremental names (image_{i:03d}) and strict extension whitelisting, which effectively neutralizes any path traversal risks from external inputs.

image-search