imaging-data-commons

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The SKILL.md and referenced guides explicitly instruct the agent to query public IDC metadata and BigQuery tables, fetch files from public S3/GCS buckets and DICOMweb proxy endpoints (e.g., client.sql_query(), BigQuery "bigquery-public-data.idc_current.*", series_aws_url, and DICOMweb proxy URL), which are untrusted third‑party data sources whose content (clinical notes, SR measurements, segment annotations, DOIs, manifest URLs) the agent is expected to read and use to decide downloads, viewer actions, and downstream processing — enabling indirect prompt-injection via that content.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.70). The skill contains runtime code that upgrades/installs the idc-index package (via subprocess.run(["pip3","install",..."]) if the version is outdated) and explicitly references the idc-index GitHub repo https://github.com/ImagingDataCommons/idc-index, meaning remote code will be fetched and executed as a required dependency at runtime.