single-cell-foundation-model-scgpt
Pass
Audited by Gen Agent Trust Hub on Mar 25, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references the installation of the 'scgpt' and 'flash-attn' Python packages via pip. These are well-known libraries in genomics and deep learning research.
- [COMMAND_EXECUTION]: The instructions involve running local Python scripts, such as 'examples/finetune_integration.py', within the repository at '/DATA/disk0/zhaosy/home/scGPT' to perform model training and evaluation.
- [DATA_EXPOSURE]: The skill accesses a specific absolute path on the local filesystem ('/DATA/disk0/zhaosy/home/scGPT') to retrieve model checkpoints and repository code. This access is limited to the defined project scope and does not target sensitive system directories.
- [PROMPT_INJECTION]: A potential surface for indirect prompt injection exists via the ingestion of untrusted data from AnnData (.h5ad) and configuration files (args.json, vocab.json) as mentioned in SKILL.md and references/local-usage.md. There are no explicit boundary markers or sanitization procedures described for these external inputs. The skill's capabilities include executing local scripts and library functions that process this data, though the risk is mitigated by the specialized nature of the tools.
Audit Metadata