The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data from multiple sources, creating a surface for indirect prompt injection. If an agent processes untrusted data using the documented Polars tools, it may be vulnerable to malicious instructions embedded within the data content.\n
Ingestion points: The skill provides instructions for reading data via pl.read_csv, pl.read_parquet, pl.read_database_uri, pl.read_ndjson, pl.read_excel, and pl.scan_parquet in SKILL.md and references/io_best_practices.md.\n
Boundary markers: The skill documentation does not mention or provide examples of using boundary markers or specific instructions to the agent to ignore instructions embedded in the processed dataframes.\n
Capability inventory: The skill enables extensive data operations, including reading and writing to the local file system, performing network operations to connect to various databases (PostgreSQL, MySQL, SQLite, BigQuery), and interacting with cloud storage services like AWS S3, Azure Blob Storage, and Google Cloud Storage (detailed in SKILL.md and references/io_best_practices.md).\n
Sanitization: There are no documented methods or instructions for sanitizing, escaping, or validating the content of external data to prevent the execution of embedded prompts.

polars-dataframes