nvalchemi-training-api
Pass
Audited by Gen Agent Trust Hub on Jul 3, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill documents the
TrainingStrategy.load_checkpointmethod for resuming training runs. In Python machine learning frameworks like PyTorch, loading checkpoints commonly uses thepicklemodule for deserialization, which can be exploited for arbitrary code execution if a checkpoint is loaded from an untrusted source. - [INDIRECT_PROMPT_INJECTION]: The training pipeline described in the skill ingests external data batches through training and validation loaders.
- Ingestion points: External data enters via
strategy.run(train_loader)andValidationConfig(validation_data=val_loader)(SKILL.md). - Boundary markers: The documentation does not specify the use of delimiters or specific instructions to ensure the model or agent ignores instructions that could be embedded within the training data.
- Capability inventory: The training strategy has capabilities for file reading and model reconstruction via
load_checkpointand code execution through the training run logic (SKILL.md). - Sanitization: There are no mentioned methods for validating or sanitizing the data batches before they are processed by the training strategy.
Audit Metadata