The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the ingestion of data from various untrusted external sources, creating a surface for indirect prompt injection where data could contain instructions meant to influence the agent.
Ingestion points: The documentation in references/io-data.md and references/quickstart.md covers loading data from CSV, Parquet, JSON, S3, SQL databases, and remote HTTP URLs.
Boundary markers: The skill lacks explicit instructions or delimiters to isolate data from instructions or to warn the agent about untrusted content in data files.
Capability inventory: SKILL.md mentions a research workflow involving the execution of Python scripts in the scripts/ directory to process data, as well as a reference to a mandatory script execution protocol in agent_reference/SCRIPT_EXECUTION_REFERENCE.md.
Sanitization: There are no mentions of data validation or sanitization steps before processing external content.

polars