ml-model-training
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from 'data.csv' using the pandas library. If this data source is attacker-controlled, it could contain malicious instructions designed to influence the agent's behavior during training or evaluation phases.
- Ingestion points: 'SKILL.md' (pd.read_csv('data.csv')).
- Boundary markers: No delimiters or explicit instructions are provided to the agent to distinguish between data content and instructions.
- Capability inventory: The skill includes scripts that write files to the local system, such as 'torch.save()' and 'model.save()', which could be misused if the agent is compromised.
- Sanitization: No input validation or data sanitization is performed on the contents of the CSV file before processing.
- [DYNAMIC_EXECUTION]: The PyTorch implementation in the reference files utilizes 'torch.load()' to restore saved model weights. This function uses the pickle module internally, which is susceptible to unsafe deserialization and can lead to arbitrary code execution if a malicious model file is loaded.
- Evidence: 'references/pytorch-training.md' (the 'Trainer.fit' method calls 'torch.load('best_model.pt')').
Audit Metadata