unsloth
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, specifically training datasets in JSONL format or from the Hugging Face Hub. This content is interpolated into model training and inference workflows.
- Ingestion points: Dataset loading occurs via
load_dataset("json", ...)in Python scripts and the--local-datasetargument in theunsloth trainCLI command. - Boundary markers: The skill utilizes structured chat templates (e.g., ChatML, Alpaca) to delimit conversation turns, which provides some structural separation but does not prevent instructions embedded within the data from being processed.
- Capability inventory: The skill executes shell commands for training, exporting models, and serving (via
unsloth,ollama, andvllm), performs file system writes for saving checkpoints and exported GGUF/merged weights, and conducts network operations to download/upload models from/to the Hugging Face Hub. - Sanitization: No explicit sanitization, validation, or filtering of the input dataset content is mentioned.
Audit Metadata