tooluniverse-rare-disease-diagnosis
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it processes untrusted clinical data (phenotypes, symptoms, and genetic variants) to automate queries across multiple external medical databases and generate local diagnostic reports.
- Ingestion points: Clinical symptoms and genetic variant data provided by the user are processed in 'Phase 1: Phenotype Standardization' (SKILL.md) and 'Phase 4: Variant Interpretation' (REFERENCE.md).
- Boundary markers: The instructions do not define clear delimiters or data framing (such as XML tags) to isolate patient data from instructions, nor do they include 'ignore embedded instructions' directives.
- Capability inventory: The agent has the capability to write local markdown and CSV files ('[PATIENT_ID]_rare_disease_report.md') and perform extensive network queries via specialized medical research tools.
- Sanitization: No explicit sanitization or validation logic for the input clinical data is described in the workflow.
- [DATA_EXFILTRATION]: The skill performs network operations to external medical and genomic databases to retrieve disease and gene information based on user input.
- Evidence: Network tool calls are made to services including Orphanet, OMIM, DisGeNET, ClinGen, GTEx, and PubMed to standardize phenotypes and interpret genetic variants.
- Context: These operations target well-known scientific and medical services required for the skill's primary diagnostic purpose.
Audit Metadata