The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted medical imaging data which could contain malicious instructions embedded within DICOM metadata tags.
Ingestion points: The scripts extract_metadata.py, anonymize_dicom.py, dicom_inventory.py, and dicom_to_image.py all parse local DICOM files which may contain external content.
Boundary markers: SKILL.md includes a dedicated 'Mandatory safety boundary' section that explicitly instructs the agent on how to handle PHI, restricting output to technical allowlists and aggregate summaries.
Capability inventory: The skill provides comprehensive local file system capabilities, including reading sensitive medical files and writing rendered images or pseudonymized derivatives.
Sanitization: Robust path validation is implemented in scripts/_common.py, including checks for parent traversal, symlink resolution, and root directory escaping. Metadata extraction tools use restricted allowlists to prevent the exposure of sensitive fields.
[EXTERNAL_DOWNLOADS]: The skill specifies installation of several well-known scientific and medical imaging libraries from the public PyPI registry.
Dependencies include pydicom, numpy, Pillow, and various pylibjpeg plugins. All packages are pinned to specific versions to ensure environment stability and mitigate supply chain risks.

pydicom