weights-and-biases
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data by downloading artifacts (datasets, models) from the Weights & Biases platform.
- Ingestion points: Untrusted data enters the agent context via the
artifact.download()method as shown inSKILL.mdandreferences/artifacts.md. - Boundary markers: The instructions do not explicitly implement boundary markers or "ignore instructions" warnings for the content of downloaded artifacts.
- Capability inventory: The skill includes file writing capabilities (
torch.save,model.save), network operations (wandb.log,wandb.save), and environment variable configuration (export WANDB_API_KEY). - Sanitization: There are no explicit sanitization or validation steps shown for the content of downloaded artifacts before they are used in training or evaluation scripts.
Audit Metadata