sr-screener

Pass

Audited by Gen Agent Trust Hub on Nov 2, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted research data (RIS exports, PDFs), creating a potential surface for indirect prompt injection.\n
  • Ingestion points: Bibliographic exports are parsed in scripts/prepare_records.py, and full-text PDFs are read in scripts/prepare_fulltext.py.\n
  • Boundary markers: The skill defines an instruction-data boundary in SKILL.md and agents/screening_reviewer_agent.md to prevent data from being interpreted as instructions.\n
  • Capability inventory: Reviewer subagents are restricted to Read and Grep tools, limiting exposure.\n
  • Sanitization: srlib.py sanitizes CSV/XLSX exports by escaping formula-triggering characters.\n- [DYNAMIC_EXECUTION]: scripts/build_workflow.py generates JavaScript scripts for the Workflow tool using predefined templates filled with configuration data.\n- [PROMPT_INJECTION]: SKILL.md includes a [direct-mode] token that bypasses intent clarification, overriding routing logic while remaining within safe operational boundaries.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 2, 2026, 04:44 PM