The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The scripts process untrusted external data (DICOM files) and display metadata to the agent's console, creating a surface for indirect prompt injection.
Ingestion points: scripts/extract_metadata.py, scripts/anonymize_dicom.py, and scripts/dicom_to_image.py ingest external files using pydicom.dcmread().
Boundary markers: Absent. There are no delimiters or warnings to the agent that output from DICOM tags should be treated as untrusted content.
Capability inventory: The scripts perform file system operations (read/write) and console logging. They do not have network access or advanced system execution capabilities.
Sanitization: While anonymize_dicom.py removes PHI, extract_metadata.py decodes and prints raw tag values (e.g., PatientName), which could contain malicious instructions designed to influence the agent when it reads the output.
[Data Exposure & Exfiltration] (SAFE): The skill is designed to improve data privacy by providing a dedicated anonymization script that removes or replaces sensitive PHI tags (PatientName, PatientID, etc.). No hardcoded credentials or unauthorized network operations were detected.
[Unverifiable Dependencies & Remote Code Execution] (SAFE): The skill references standard, reputable Python libraries (pydicom, numpy, pillow). No automated downloads or suspicious remote code patterns were found in the scripts.

pydicom