The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted medical imaging files (DICOM), creating a vulnerability surface for indirect prompt injection via metadata tags. \n
Ingestion points: Data enters the agent's context through pydicom.dcmread() calls found in SKILL.md and all helper scripts in the scripts/ directory. \n
Boundary markers: There are no explicit delimiters or instructions defined to treat extracted DICOM metadata as untrusted or to ignore embedded instructions. \n
Capability inventory: The skill provides capabilities for local file system writes via ds.save_as() and processing of extracted metadata. \n
Sanitization: The anonymize_dicom.py script removes specific Protected Health Information (PHI) but does not validate or sanitize metadata for potential instruction injection. \n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of well-known and established Python libraries for medical imaging and data science. \n
Dependencies include pydicom, pillow, numpy, matplotlib, pylibjpeg, pylibjpeg-libjpeg, pylibjpeg-openjpeg, and python-gdcm.

pydicom