The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external DICOM files provided by the user, which constitutes a potential attack surface for indirect prompt injection. If a malicious DICOM file contains strings in header fields (such as StudyDescription or PatientName) that are designed as instructions, an agent processing the tool's output might be influenced.
Ingestion points: The scripts/extract_metadata.py script reads the file specified by the dicom_path argument.
Boundary markers: No explicit boundary markers or instructions to ignore embedded content are provided in the tool output or agent instructions.
Capability inventory: The skill requires file_read permissions for the input DICOM, file_write for the optional JSON output, and shell access to execute the Python script entrypoint.
Sanitization: The script uses a _safe_str() helper to cast values to strings or representations, providing basic type safety, but it does not perform semantic filtering or sanitization of the header content itself.
[COMMAND_EXECUTION]: The skill uses subprocess.run() within its test suite (tests/test_basic.py) to verify script functionality. This is a standard practice for development testing and does not pose a risk in the production execution context of the agent, as the command only invokes the skill's own local script with controlled fixture paths.

dicom-metadata-extract