transformers
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill documents the trust_remote_code=True parameter for loading models, which allows for the execution of repository-provided Python code from the Hugging Face Hub.
- [EXTERNAL_DOWNLOADS]: Fetches pre-trained models and datasets from the well-known Hugging Face Hub service (huggingface.co).
- [DYNAMIC_EXECUTION]: Includes examples of torch.compile for runtime JIT compilation of models for performance optimization.
- [INDIRECT_PROMPT_INJECTION]: Documents the ingestion of untrusted data via model pipelines and datasets, establishing an attack surface. (1) Ingestion points: Input prompts to the pipeline API in SKILL.md and datasets loaded via load_dataset in reference/fine-tuning.md. (2) Boundary markers: Absent from code examples. (3) Capability inventory: Network access for fetching models, local file system writes for saving results, and local model inference. (4) Sanitization: Not addressed in the provided implementation examples.
Audit Metadata