weights-and-biases
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches and installs the wandb package from the official Python Package Index (PyPI) and provides numerous links to documentation and examples on GitHub.
- [INDIRECT_PROMPT_INJECTION]: The skill demonstrates workflows for ingesting external data (datasets, models) via W&B Artifacts, which introduces an attack surface where an agent could potentially be influenced by instructions hidden within those files.
- Ingestion points: Identified in
references/artifacts.md(e.g.,artifact.download()) andSKILL.md(e.g.,run.use_artifact). - Boundary markers: No explicit instructions or delimiters are provided in the code snippets to warn the agent about ignoring potential instructions within ingested data.
- Capability inventory: The skill uses tools for file writing (
torch.save,wandb.save), logging to external cloud services (wandb.log), and package installation via pip. - Sanitization: No specific data validation, schema enforcement, or content sanitization is demonstrated for the downloaded artifacts before processing.
Audit Metadata