light-idea-generation
Pass
Audited by Gen Agent Trust Hub on Jun 13, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions call for the execution of local Python scripts, specifically
search_normalize.pyandsnowball.py, located within a companion skill folder (light-literature-search). These scripts are used for fetching and normalizing research paper metadata. As these are part of the vendor's own ecosystem and perform expected bibliographic tasks, they do not pose a security risk. - [PROMPT_INJECTION]: The skill processes user-provided project backgrounds and external data from literature searches, creating a surface for indirect prompt injection where content in search results could theoretically influence agent behavior.
- Ingestion points: Project background variables in
SKILL.mdand paper metadata (titles, abstracts) retrieved from external APIs. - Boundary markers: The skill utilizes a structured Markdown template (
templates/idea_card.md) which enforces a schema for the generated ideas, helping to isolate external content. - Capability inventory: The skill can execute local Python scripts via the CLI and write analysis results to
idea_candidates.md. - Sanitization: No explicit sanitization or filtering of external paper content is described, relying on the model's internal safety guardrails for data processing.
- [SAFE]: All identified external references target well-known academic services or reputable research organizations. The skill demonstrates best practices for secret management by instructing users to configure API keys via environment variables rather than hardcoding credentials.
Audit Metadata