The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is designed to fetch scientific metadata and documentation from an external domain, https://huggingscience.co, using a bundled Python script. This network activity is the primary mechanism for the skill's catalog discovery feature.
[PROMPT_INJECTION]: The process of ingesting and parsing external markdown data from the Hugging Science catalog introduces a surface for indirect prompt injection. Maliciously formatted entries in the remote catalog could potentially influence the agent's reasoning or the instructions it follows when selecting resources.
Ingestion points: scripts/fetch_catalog.py (fetches from huggingscience.co endpoints including llms-full.txt and topic files).
Boundary markers: Content is parsed into structured objects, but titles and descriptions are processed as free-form strings.
Capability inventory: The skill facilitates subprocess execution (local scripts), file system access (via the datasets library), and network operations (via huggingface_hub and gradio_client).
Sanitization: The skill parses the data for structure but does not implement validation or filtering to detect embedded instructions in the metadata.
[REMOTE_CODE_EXECUTION]: The skill documentation (using-models.md) provides guidance on using the trust_remote_code=True parameter when loading models from the Hugging Face Hub. This is a standard requirement for many scientific foundation models that use custom architectures, as it allows the execution of arbitrary Python code included in the model repository. While the skill warns users and targets well-known services, this remains an inherent risk of the research ML ecosystem.

hugging-science