ml-training-recipes

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches specialized CUDA kernels from the kernels-community repository on GitHub for optimized transformer attention calculations (references/architecture.md).
  • [EXTERNAL_DOWNLOADS]: Downloads pretrained protein language models from Hugging Face's official repository using the transformers library (references/biomedical.md).
  • [COMMAND_EXECUTION]: Recommends using standard git commands (commit, reset) to track and revert model training experiments in a systematic loop (references/experiment-loop.md).
  • [COMMAND_EXECUTION]: References standard command-line tools for the nnUNetv2 framework to automate biomedical image segmentation workflows (references/biomedical.md).
  • [INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection due to its handling of external data and libraries.
  • Ingestion points: Processes SMILES strings, protein sequences, EHR data, and raw text chunks during model training and tokenizer creation (references/biomedical.md, references/experiment-loop.md).
  • Boundary markers: The provided examples do not use explicit delimiters to isolate external data inputs within prompts.
  • Capability inventory: Employs dynamic kernel loading, CLI tool execution, and JIT compilation using torch.compile.
  • Sanitization: No specific sanitization or input validation logic is included in the reference implementations.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:07 PM
Security Audit — agent-trust-hub — ml-training-recipes