dicom-metadata-extract

Pass

Audited by Gen Agent Trust Hub on Aug 31, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external DICOM files via the pydicom library in scripts/extract_metadata.py.
  • Ingestion points: The extract function in scripts/extract_metadata.py reads file content provided by the user.
  • Boundary markers: The skill uses structured JSON output but lacks explicit instructions within the data stream to ignore instructions embedded in DICOM tags.
  • Capability inventory: The skill possesses Bash and file-write capabilities (via the --output flag in scripts/extract_metadata.py).
  • Sanitization: The script uses json.dumps() for output formatting, providing basic structural isolation, but does not filter content for prompt injection patterns.
  • [COMMAND_EXECUTION]: The skill's test suite (tests/test_basic.py) uses subprocess.run to execute the extraction script. While this is standard for testing, it involves shell-level invocation of the Python interpreter.
  • [SAFE]: The skill is authored by a known vendor (NVIDIA) and focuses on a specific, documented medical research task. It includes multiple layers of evaluation data (BENCHMARK.md, evals/) and explicit safety disclaimers warning against clinical use or reliance on its PHI detection for regulatory compliance.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 31, 2026, 07:00 PM
Security Audit — agent-trust-hub — dicom-metadata-extract