The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface where the agent could potentially be influenced by malicious instructions embedded within ingested datasets or their associated metadata.
Ingestion points: Data and metadata are loaded from external sources such as local files, ZIP archives, and cloud storage (S3 and Google Cloud Storage) using functions like zarr.open, zarr.open_array, xr.open_zarr, and da.from_zarr as documented in SKILL.md and references/api_reference.md.
Boundary markers: The provided instructions lack safety markers, delimiters, or explicit warnings to the agent to disregard potential instructional content found within numeric datasets or JSON-serializable array attributes.
Capability inventory: The skill provides the agent with capabilities to write to the local filesystem (LocalStore, ZipStore) and interact with cloud storage via network operations (s3fs, gcsfs) throughout the documentation in SKILL.md and references/api_reference.md.
Sanitization: There is no evidence of input validation or sanitization protocols to filter out non-numeric or instructional strings from array attributes or data contents before they are processed by the agent.

zarr-python