The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill provides the agent with a file-write capability via the output_file parameter in scripts/download_era5.py. An attacker could use indirect prompt injection (e.g., via a malicious data request) to trick the agent into specifying a sensitive path like ~/.bashrc or ~/.ssh/authorized_keys as the output destination, potentially leading to system compromise or denial of service through file overwriting.
Ingestion points: User-provided parameters for output_file in the Python script and output_file in the programmatic download_era5 function.
Boundary markers: Absent. There is no validation or sanitization of the output path to ensure it remains within a dedicated data directory.
Capability inventory: File system write (result.download(str(output_path))) and directory creation (output_path.parent.mkdir).
Sanitization: Absent. The script uses the provided path directly without checking for path traversal or restricted directories.
[Data Exposure & Exfiltration] (LOW): The skill requires and handles a COPERNICUS_API_KEY. While it does not hardcode credentials, it explicitly instructs the agent to use and manage these secrets.
[External Downloads] (LOW): The skill relies on the cdsapi library and makes network requests to https://cds.climate.copernicus.eu/api. This is a legitimate service for climate data, but involves external network communication.

era5-download