The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large volumes of untrusted external data, such as web scrapes from Common Crawler, which represents a potential attack surface for indirect prompt injection.
Ingestion points: The skill reads data from various formats including Parquet, JSONL, and cloud storage via DocumentDataset.read_parquet in SKILL.md.
Boundary markers: The provided instructions do not define specific delimiters or isolation protocols for separating external data from agent instructions.
Capability inventory: The skill can write processed datasets to the file system using dataset.to_parquet and perform distributed execution via Ray and Dask backends.
Sanitization: While the skill includes PII redaction and quality filtering stages, these are primarily data curation tools and do not specifically sanitize the input to prevent instruction injection targeting the processing agent.
[SAFE]: All package installations (uv pip install), model downloads (nvidia/quality-classifier-deberta, sentence-transformers/*), and documentation references point to trusted organizations (NVIDIA, Hugging Face) and well-known services. The operations described are consistent with standard high-performance data engineering workflows.

nemo-curator