find-cohort-gap

Pass

Audited by Gen Agent Trust Hub on Oct 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources which could contain malicious instructions designed to influence the agent's research gap analysis or output ranking.
  • Ingestion points: The scripts/build_cohort_profile.py script reads local codebooks (.csv, .xlsx, .json, .md) and fetches content from user-provided context URLs. Furthermore, Phases 1, 3, and 4 utilize the /search-lit skill to ingest PubMed abstracts and metadata.
  • Boundary markers: External context extracted by the script is wrapped in # Context: [source] headers in the context_extract.md file. The SKILL.md includes an 'Anti-Hallucination' section requiring verification of all clinical definitions and references.
  • Capability inventory: The skill is granted Bash, Write, Edit, Read, Grep, and Glob tools. The bundled Python script uses subprocess.run to call system utilities.
  • Sanitization: No explicit sanitization or filtering of natural language instructions within the ingested documents or PubMed data is performed before processing.
  • [COMMAND_EXECUTION]: The scripts/build_cohort_profile.py script executes the pdftotext system utility via subprocess.run to extract text from PDF files. While the arguments are passed as a list (reducing shell injection risks), it constitutes a subprocess call on external, potentially untrusted files.
  • [EXTERNAL_DOWNLOADS]: The scripts/build_cohort_profile.py script performs network operations using urllib.request to fetch domain context from arbitrary URLs provided by the user. This is an intended feature for research but allows the agent to interact with external web servers.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 4, 2026, 01:28 PM