pydicom

Pass

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: LOWPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (LOW): The skill processes untrusted external binary data (DICOM files) and extracts textual metadata, creating a surface for injection. * Ingestion points: The skill uses pydicom.dcmread() to ingest binary data in scripts/anonymize_dicom.py, scripts/dicom_to_image.py, and scripts/extract_metadata.py. * Boundary markers: Absent. Metadata extracted in extract_metadata.py (such as PatientName or StudyDescription) is returned as raw text strings without delimiters or instructions to ignore embedded commands. * Capability inventory: The skill has filesystem write access via ds.save_as() and image.save(). No network access or arbitrary command execution was found. * Sanitization: None. Extracted metadata is formatted for display but not sanitized or escaped for safe consumption by an LLM.
  • Data Privacy (INFO): The script scripts/anonymize_dicom.py is a privacy-enhancing utility designed to remove PHI. While beneficial, users should ensure the PHI_TAGS list meets their specific regulatory requirements.
Audit Metadata
Risk Level
LOW
Analyzed
Feb 16, 2026, 09:39 AM