The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted DICOM data which could contain malicious content within metadata tags or private elements.
Ingestion points: DICOM datasets are ingested via pydicom.dcmread in multiple scripts including anonymize_dicom.py, extract_metadata.py, and dicom_inventory.py.
Boundary markers: The skill includes a 'Mandatory safety boundary' and explicitly warns users against printing full datasets or logs containing clinical data.
Capability inventory: The skill allows writing modified DICOM files, rendering image frames, and generating technical audit reports.
Sanitization: Implements strict tag allowlisting, recursive private tag removal, and technical inventory filters that omit quasi-identifiers and free-text fields.
[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions for installing dependencies from the official Python Package Index (PyPI).
Evidence: References standard packages such as pydicom, numpy, Pillow, and transfer-syntax plugins like pylibjpeg and python-gdcm.
[SAFE]: The skill scripts incorporate multiple best-practice security controls for handling sensitive data.
Path Validation: The _common.py module contains a checked_input and checked_output framework that enforces local filesystem paths, prohibits parent directory traversal (..), and rejects symlinks and network URIs.
Resource Bounding: Scripts enforce configurable limits on file sizes, frame counts, sequence nesting depth, and total data elements to prevent denial-of-service through resource exhaustion.
Secure Cryptography: Pseudonymization and UID remapping use hmac with SHA-256 and site-specific keys, ensuring deterministic and secure linkage without exposing original identifiers.

pydicom