torchcode-pytorch-interview-practice
Warn
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill requires installing the
torch-judgepackage from PyPI and cloning theTorchCoderepository from an unverified GitHub account (duoan/TorchCode). - [DYNAMIC_EXECUTION]: The grading system relies on the
torch_judge.check()function to dynamically execute and test user-provided PyTorch code against internal benchmarks. - [COMMAND_EXECUTION]: The setup process involves running shell commands including
pip install,git clone,make run, anddocker runto pull images fromghcr.io/duoan/torchcode. - [INDIRECT_PROMPT_INJECTION]: The skill ingests user-defined PyTorch functions for evaluation, creating a surface for potential exploitation if malicious code is processed.
- Ingestion points: Implementation code in Jupyter notebook templates (e.g.,
templates/01_relu.ipynb). - Boundary markers: None documented.
- Capability inventory: The
torch_judge.check()utility executes user code, performs gradient verification, and calculates timing. - Sanitization: None documented.
Audit Metadata