ray-train
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill demonstrates how to ingest external datasets into training pipelines, specifically through the
train_datasetparameter in the HuggingFace Transformers workflow. This establishes a vulnerability surface where adversarial instructions within the training data could be executed or influence the training process.\n - Ingestion points: The
train_datasetvariable in theTransformersTrainerexample inSKILL.md.\n - Boundary markers: No delimiters or 'ignore embedded instructions' warnings are included in the code examples to separate data from execution logic.\n
- Capability inventory: The training process involves distributed task execution on Ray nodes, file system persistence for checkpoints (
torch.save), and inter-node network communication.\n - Sanitization: The provided workflows do not demonstrate validation, filtering, or sanitization of input data before processing.
Audit Metadata