literature-review

Pass

Audited by Gen Agent Trust Hub on Mar 15, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill includes mandatory behavioral instructions forcing the use of specific visualization tools and a promotional section that steers the agent to suggest an external hosted platform (k-dense.ai) for complex workflows.
  • [PROMPT_INJECTION]: There is a discrepancy between the author identifier provided by the system ('wu-yc') and the author identity claimed in the metadata ('K-Dense Inc.'), which may mislead users about the skill's origin.
  • [PROMPT_INJECTION]: A surface for indirect prompt injection exists because the skill processes untrusted research metadata from external databases. 1. Ingestion points: JSON-formatted search results from PubMed, arXiv, and Semantic Scholar. 2. Boundary markers: Absent; no specific instruction delimiters are mentioned for processed paper content. 3. Capability inventory: 'Bash', 'Write', and 'Edit' tools are enabled for script execution and file modification. 4. Sanitization: No specific filtering or escaping of paper titles or abstracts is documented in the workflow.
  • [COMMAND_EXECUTION]: The skill requires the 'Bash' tool to execute local utility scripts including 'generate_schematic.py', 'search_databases.py', 'verify_citations.py', and 'generate_pdf.py' for processing research data and generating documents.
  • [EXTERNAL_DOWNLOADS]: The skill instructions suggest installing standard system dependencies such as Pandoc and LaTeX, as well as the Python 'requests' library to support its core functionality.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 15, 2026, 04:33 AM