The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest whole-slide images (WSIs) from potentially untrusted sources for analysis. It also possesses capabilities to write output files (HDF5, CSV, and PNG) to the local file system. This combination creates a surface for indirect injection if an attacker-controlled image file could somehow influence the agent's behavior during processing, though binary image formats carry negligible risk for traditional prompt injection.
Ingestion points: Slide files (SVS, NDPI, MRXS, TIFF) loaded via SlideData in SKILL.md.
Boundary markers: None identified; image data is processed via standard vision libraries.
Capability inventory: File system write operations including slide.write(), df.to_csv(), and img.save() documented in SKILL.md.
Sanitization: Standard image processing pipeline; no specific text-based sanitization for binary slide data.
[EXTERNAL_DOWNLOADS]: The skill instructions include downloading software and models from established sources.
Evidence: Installation commands for packages from the conda-forge repository and pytorch.org.

pathml