The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process DICOM files, which are untrusted external binary data. This creates a potential surface for indirect prompt injection or structural exploits if the underlying parser or processing scripts are vulnerable to maliciously crafted files.
Ingestion points: The pydicom.dcmread() function is used in SKILL.md, scripts/anonymize_dicom.py, scripts/dicom_to_image.py, and scripts/extract_metadata.py to parse external files.
Boundary markers: No explicit boundary markers or delimiters are used to separate untrusted data from processing logic in the examples provided.
Capability inventory: The skill has capabilities to write to the file system (ds.save_as, image.save) and modify the environment via package installation.
Sanitization: While an anonymization script (scripts/anonymize_dicom.py) is provided to remove Protected Health Information (PHI) from metadata, there is no evidence of sanitization against structural or binary-level exploits within the DICOM format itself.
[EXTERNAL_DOWNLOADS]: The instructions include commands to download and install several external Python libraries from public registries.
Evidence: Installation instructions for pydicom, pillow, numpy, matplotlib, pylibjpeg, pylibjpeg-libjpeg, pylibjpeg-openjpeg, and python-gdcm using uv pip install.
[COMMAND_EXECUTION]: The skill instructs the user and agent to execute shell commands for library installation and for running utility scripts.
Evidence: Shell commands for uv pip install and script execution (e.g., python scripts/anonymize_dicom.py) are documented throughout SKILL.md and script headers.

pydicom