The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash(python:*) tool to execute Python scripts for medical image processing. This is a restricted, purpose-built usage for tasks such as anonymization (anonymize_dicom.py) and format conversion (dicom_to_image.py).
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to ingest and process external medical data files (.dcm), it has an inherent attack surface for indirect prompt injection if an attacker embeds instructions in DICOM tags.
Ingestion points: Data enters the system via pydicom.dcmread in SKILL.md and all provided scripts.
Boundary markers: The skill provides instructions to handle exceptions when reading files from untrusted sources and to verify anonymization thoroughly.
Capability inventory: The skill has file-write capabilities (ds.save_as) and local script execution.
Sanitization: The skill explicitly includes anonymize_dicom.py to remove or replace sensitive patient data, which acts as a primary sanitization layer for clinical workflows.
[DYNAMIC_EXECUTION]: The skill uses dynamic imports (e.g., from pydicom.pixels import apply_voi_lut inside try-except blocks) to maintain compatibility between different versions of the pydicom library (2.x vs 3.x). This is a standard software engineering practice for library migration and does not involve executing untrusted or remote code.

alterlab-pydicom