dicom-web-query

Pass

Audited by Gen Agent Trust Hub on Oct 5, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes JSON data fetched from remote DICOMweb servers (via QIDO-RS and WADO-RS) and uses this data to inform agent actions, creating an attack surface where a malicious server could return instructions to influence the agent.
  • Ingestion points: JSON responses from requests.get calls in qido_studies, retrieve_study, and get_study_metadata in SKILL.md.
  • Boundary markers: The skill does not define specific delimiters or instructions to prevent the agent from obeying strings embedded in the medical metadata.
  • Capability inventory: The skill includes capabilities for file system writes (download_study_dicom, get_thumbnail) and network storage (store_dicom).
  • Sanitization: There is no evidence of sanitization or validation of the remote data before it is parsed or used by the agent.
  • [EXTERNAL_DOWNLOADS]: The skill provides Python code to download DICOM objects and thumbnails from external servers. For example, download_study_dicom fetches ZIP archives and retrieve_image_pixels fetches JPEG data from a remote BASE_URL.
  • [COMMAND_EXECUTION]: The instructions provide Python implementations for network communication and local file I/O. Specifically, download_study_dicom and get_thumbnail write binary data to the local disk, while store_dicom performs network POST requests with local file content.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 5, 2026, 05:08 PM