biomedical-skills

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The modules for literature review and automated research (e.g., med-researcher-guide, medgeclaw-guide, med-researcher-r1-guide) are designed to ingest untrusted data from external scientific sources such as PubMed and PMC for summarization and synthesis.
  • Ingestion points: ncbi.nlm.nih.gov, ebi.ac.uk, and clinicaltrials.gov API responses containing unstructured medical text.
  • Boundary markers: No explicit instruction-isolation markers or 'ignore embedded instructions' prompts are included in the illustrative code snippets.
  • Capability inventory: Access to networking libraries (requests) and local tool execution (Bioinformatics CLIs).
  • Sanitization: Relies on standard library parsers for structured data (JSON/XML) without specific input validation for NL instructions.
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions and examples for downloading research datasets and command-line utilities from well-known and established scientific infrastructure.
  • Evidence: Fetches genomics tools from NCBI's public FTP server, protein structure data from EMBL-EBI and AlphaFold services, and drug data from the FDA's open portal.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 09:31 AM