single-popv-annotation
Pass
Audited by Gen Agent Trust Hub on Aug 6, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Downloads pretrained cell type classification models from the HuggingFace Hub (e.g.,
popv/immune_all). This is a standard and well-recognized method for sharing machine learning models within the scientific community. - [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external datasets (H5AD files). Maliciously crafted metadata or labels in these files could theoretically attempt to influence an agent's logic if displayed in the final output.
- Ingestion points: Data is loaded via
ov.readand theProcess_Queryclass in bothSKILL.mdandreference.md. - Boundary markers: None; the data is interpreted as categorical and numerical biological features.
- Capability inventory: Includes file system writing for results and model checkpoints, as well as network access for fetching pretrained models.
- Sanitization: Not present for the internal cell label content, though the skill performs some structural validation on the dataframes (e.g., checking for NaN values).
Audit Metadata