The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches catalog data from the external domain huggingscience.co, which could potentially contain malicious instructions embedded in dataset or model descriptions.
Ingestion points: scripts/fetch_catalog.py fetches markdown files from the remote catalog.
Boundary markers: The script prepends an UNTRUSTED_BANNER to fetched content to warn the agent that the data is third-party and not authoritative instructions.
Capability inventory: The skill is capable of network requests, file system access, and model execution via libraries like transformers and gradio_client.
Sanitization: The _defang function in scripts/fetch_catalog.py strips markdown code fences (```) and frontmatter delimiters (---) from the fetched text to prevent the agent from misinterpreting data as executable blocks or system configuration.
[REMOTE_CODE_EXECUTION]: The skill facilitates the use of models from the Hugging Face Hub that may require the execution of custom modeling code.
Evidence: references/using-models.md explicitly discusses the use of the trust_remote_code=True flag for architectures like Evo-2 and various genomics models.
Mitigation: The skill mandates that the agent must inform the user about the implications of this flag and request explicit consent before proceeding, treating the repository as an untrusted source until confirmed.
[EXTERNAL_DOWNLOADS]: The scripts/fetch_catalog.py utility programmatically retrieves resource lists from https://huggingscience.co, which is the dedicated domain for this catalog.

hugging-science