answering-research-questions

Pass

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill identifies an attack surface for indirect prompt injection by ingesting and processing data from external scientific literature.
  • Ingestion points: External research abstracts and full-text articles retrieved via rp_search and rp_fulltext in Phase 3 and Phase 4.
  • Boundary markers: The instructions do not specify the use of delimiters or provide the agent with guidance to ignore potential instructions embedded within the research text.
  • Capability inventory: The agent is granted capabilities to write to the file system (mkdir, file writing) and execute locally generated Python scripts.
  • Sanitization: No explicit sanitization or validation of the retrieved academic content is performed before the agent evaluates it for data extraction.
  • [COMMAND_EXECUTION]: The workflow involves generating and executing Python scripts to handle research data.
  • Evidence: Phase 7 and Phase 8 instruct the agent to create and run scripts such as generate_summary.py and JSON filtering snippets to consolidate findings and manage research sessions.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 6, 2026, 12:19 AM
Security Audit — agent-trust-hub — answering-research-questions