imaging-data-commons
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill facilitates the download of medical imaging datasets from official National Cancer Institute (NCI) buckets on AWS S3 and Google Cloud Storage (idc-open-data). These are well-known research services.
- [DYNAMIC_EXECUTION]: Includes a configuration script that uses
subprocess.runto callpip3for upgrading theidc-indexpackage. This is a standard setup procedure for ensuring the agent uses the library version matching the current data release. - [INDIRECT_PROMPT_INJECTION]: The skill processes metadata and clinical records from external sources, which creates a potential surface for indirect prompt injection.
- Ingestion points: Data enters the context from SQL query results (client.sql_query), clinical tables (client.get_clinical_table), and DICOM metadata read from downloaded files.
- Boundary markers: No delimiters or safety warnings are defined for the ingested research data.
- Capability inventory: The skill has the capability to execute subprocesses (for package upgrades), write files to the local system, and open URLs in a browser.
- Sanitization: There are no documented sanitization or validation routines for the external metadata strings.
Audit Metadata