llm-radiology-use
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external clinical data such as radiology reports and patient histories by interpolating them directly into prompt templates.
- Ingestion points: Untrusted data enters the agent context via the
report_textvariable inanalyze_report, thecurrentandpriorvariables incompare_reports, and thefindingsandhistoryvariables inget_differentialwithinSKILL.md. - Boundary markers: No explicit delimiters (like triple backticks or XML tags) or "ignore embedded instructions" warnings are used when interpolating external text into the prompt strings.
- Capability inventory: The skill possesses the capability to write data to external datastores via the
send_results_to_healthlakefunction using theboto3library and to transmit data to various medical APIs viarequests.post. - Sanitization: There is no evidence of input validation, escaping, or filtering of the clinical text before it is sent to the LLM providers.
- [EXTERNAL_DOWNLOADS]: The skill makes network requests to external API endpoints for medical reasoning and data storage.
- Evidence: The
query_medpalmandmedical_insightsfunctions communicate with Google's generative language and health APIs (googleapis.com). Theextract_medical_entitiesfunction communicates with Microsoft's Azure AI Health service (cognitiveservices.azure.com). Theconfigure_healthlakefunction uses theboto3client to interact with Amazon Web Services.
Audit Metadata