bioresearcher-deep-research
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructions provide configuration for the harness to install several Node.js packages via npx, including 'biomcp@1.1', 'webr@0.6', and 'mysql2@3'. Additionally, the report conversion step uses the Python 'markdown' package fetched via 'uv run'.
- [COMMAND_EXECUTION]: The skill uses the 'Bash' tool to execute a bundled Python script ('scripts/markdown-to-html.py') for generating HTML research reports. It also utilizes the 'biomcp' server to interact with local SQL databases and genomic data artifacts when configured.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large amounts of data from external biomedical sources such as PubMed, ClinicalTrials.gov, and patent databases.
- Ingestion points: Data enters the context through 'biomcp' tools like 'article_search', 'trial_get', and 'patent_search' referenced in 'SKILL.md' and various 'references/' documentation files.
- Boundary markers: The skill enforces a strict 6-section report template ('references/report-template.md') and a numbered citation format ('references/citations.md') to structure outputs.
- Capability inventory: The agent has 'Bash' and 'Write' access, allowing it to write reports to the local file system and convert them via Python scripts.
- Sanitization: Instructions mandate verifying identifiers (PMID, NCT ID, Accessions) against tool output and explicitly listing 'Evidence Gaps' when source data is unavailable or conflicting.
- [DYNAMIC_EXECUTION]: The skill supports advanced genomic analysis through environment-gated tools ('analysis_r_' and 'analysis_biowasm_') which execute WebAssembly-based code bundles for differential expression and sequence processing at runtime. These features require explicit opt-in via environment variables.
Audit Metadata