scientific-brainstorming
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill establishes a comprehensive methodology for evidence-aware scientific brainstorming, prioritizing human-first ideation and independent review. It includes explicit operating rules to maintain safety, ethical integrity, and regulatory compliance.
- [COMMAND_EXECUTION]: Includes local Python scripts (session_scaffold.py, evaluate_matrix.py, validate_register.py) for data handling and matrix calculations. These utilities are deterministic, restricted to the Python standard library, and implement security measures such as bounded input reading, symlink protection, and private file permissions (0o600).
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch citation metadata from arXiv, a well-known academic repository, to ensure accurate referencing of research papers. This is an informational data retrieval task involving a trusted academic service.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests brainstorming content and scoring data through JSON and CSV files. This attack surface is managed by strict schema validation in the provided scripts and instructions for the agent to use boundary markers and explicit labeling to separate data from processing logic.
Audit Metadata