The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process model identifiers and revision hashes from external sources such as environment variables (HF_MODEL_ID, HF_REVISION) to load machine learning assets.
Ingestion points: Environment variables accessed in gpt.md and configuration strings in SKILL.md.
Capability inventory: Uses the transformers library to download and load weights/code, which involves network access and file system operations.
Boundary markers: The instructions explicitly mandate pinning revisions and using safetensors to ensure the integrity of the loaded assets.
Sanitization: No explicit validation or filtering of the model identifiers is performed before they are passed to the loading functions.
[EXTERNAL_DOWNLOADS]: The skill fetches model configurations, tokenizers, and weights from Hugging Face's official repository. Hugging Face is a well-known service for machine learning models. The skill mitigates risks by recommending the use of pinned commit SHAs for revisions and the safetensors format to avoid the security risks associated with unpickling untrusted model weights.

huggingface-transformers