ai-ml-skills
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: Several modules (e.g.,
npcpy-research-guide,ai-agent-papers-guide,autonomous-agents-papers-guide) introduce an attack surface for indirect prompt injection by ingesting and reasoning over external, untrusted data.\n - Ingestion points: The skills demonstrate workflows that ingest content from academic PDFs (
workflow.analyze_paper("paper.pdf")) and web search results (Agenttools likeweb_search).\n - Boundary markers: Implementation examples lack robust delimiters or specific instructions to the model to disregard potential directives embedded within the ingested data.\n
- Capability inventory: The documented agents possess tools for network access and file system interaction, which could be abused if the model follows malicious instructions from an external source.\n
- Sanitization: No explicit sanitization or filtering logic is demonstrated for external data streams.\n- [EXTERNAL_DOWNLOADS]: The documentation includes standard commands for installing machine learning and utility libraries via
pip, includingnpcpy,vmas,pykan,lm-eval, andlightning. These references target established research tools and public packages.\n- [SAFE]: The majority of the skills provide high-quality documentation for well-known AI frameworks. Sensitive operations, such as Hugging Face API authentication, are correctly handled using environment variables to avoid hardcoding secrets.
Audit Metadata