nanogpt
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [DYNAMIC_EXECUTION]: The training and data preparation scripts in
references/data.mdandreferences/training.mdusepickle.load()andtorch.load()to deserialize model metadata and checkpoints. Deserializing untrusted data with these methods can lead to arbitrary code execution. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large external datasets for training and inference, which creates a vulnerability to indirect instructions.
- Ingestion points: Data is fetched from GitHub and HuggingFace as described in
references/data.md. - Boundary markers: The instructions do not define boundary markers or safety guidelines to differentiate between data and instructions within the processed text.
- Capability inventory: The skill includes file system writes, network downloads, and text generation capabilities.
- Sanitization: No sanitization or validation of the input text data is performed before it is processed by the model.
- [EXTERNAL_DOWNLOADS]: The skill fetches the TinyShakespeare dataset from a public GitHub repository and uses HuggingFace to download the OpenWebText dataset.
Audit Metadata