skills/aipoch/open-science/scgpt/Gen Agent Trust Hub

scgpt

Pass

Audited by Gen Agent Trust Hub on Jul 14, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted biological data files (.h5ad) which represent a potential attack surface for indirect prompt injection via malicious metadata.
  • Ingestion points: External data is ingested using ad.read_h5ad("dataset.h5ad") in SKILL.md.
  • Boundary markers: None identified; data is processed directly by the model tasks.
  • Capability inventory: The skill uses subprocess-like orchestration for remote compute jobs, file writing for output artifacts, and specialized biological data processing via the scgpt and scanpy libraries.
  • Sanitization: No specific sanitization or validation of the AnnData object structure or metadata is mentioned before processing.
  • [EXTERNAL_DOWNLOADS]: The skill references weights and checkpoints from the Wang Lab at the University of Toronto (bowang-lab/scGPT). These are sourced from a recognized academic research laboratory and used for their intended purpose in biological modeling.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 14, 2026, 10:23 PM
Security Audit — agent-trust-hub — scgpt