torchforge-rl-training
Warn
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes workflows for training models using external datasets, which creates a surface for indirect prompt injection.
- Ingestion points: External datasets such as GSM8K referenced in config/grpo_math.yaml.
- Boundary markers: There are no explicit markers or delimiters used to separate user instructions from dataset content.
- Capability inventory: The training scripts compute rewards and update model parameters based on the dataset content.
- Sanitization: No sanitization or filtering of the dataset content is specified in the provided instructions.
- [EXTERNAL_DOWNLOADS]: The skill directs the user to install software from sources and execute scripts that are not verified as trusted vendors. Evidence: Instructions to run ./scripts/install.sh and clone github.com/meta-pytorch/monarch.
- [COMMAND_EXECUTION]: The skill requires the execution of multiple system commands to provision environments and run training jobs. Evidence: Usage of conda, python -m apps.grpo.main, and sbatch.
Audit Metadata