The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides an example of model loading using joblib.load(). Because joblib uses pickle-based serialization, loading files from untrusted sources is unsafe and can result in arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process datasets from external locations like S3, GCS, and SQL databases. This creates a surface for indirect prompt injection where malicious instructions within the data could influence the agent.
Ingestion points: Data loading methods such as read_parquet, read_csv, and read_datasource in SKILL.md.
Boundary markers: No delimiters or warnings are used to isolate untrusted data in the examples.
Capability inventory: The skill can write files, connect to external databases, and execute distributed tasks.
Sanitization: No input validation or sanitization logic is shown in the transformation steps.
[EXTERNAL_DOWNLOADS]: The skill references standard packages and downloads models from Hugging Face, which are trusted and well-known services.

ray-data