weights-and-biases
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill facilitates downloading artifacts such as datasets and models from the Weights & Biases platform (wandb.ai), which is a well-known service for MLOps.
- [INDIRECT_PROMPT_INJECTION]: The skill instructions include an attack surface for indirect prompt injection via the processing of external data. Ingestion points include
artifact.download()andrun.use_artifact()methods as shown inreferences/artifacts.md. No explicit boundary markers or 'ignore instructions' warnings are provided for the downloaded data. The capability inventory includes the use oftorch.load()on these downloaded files. There is no mention of sanitization or integrity verification for the downloaded artifacts before loading. - [DYNAMIC_EXECUTION]: The skill demonstrates the use of
torch.load()for model checkpoints inreferences/artifacts.md. Although this is standard practice in the machine learning community, it involves the Python unpickling process which can execute arbitrary code embedded in a malicious model file.
Audit Metadata