The Agent Skills Directory

[DATA_EXFILTRATION]: The skill documents and enables interaction with cloud storage providers (AWS S3, Google Cloud Storage) using the fsspec library and standard provider environment variables (e.g., AWS_ACCESS_KEY_ID, GOOGLE_APPLICATION_CREDENTIALS). While the skill instructions emphasize local processing and require user confirmation for remote write paths, the capability to read from and write to arbitrary cloud buckets using sensitive credentials represents a potential data exfiltration surface.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process molecular data from external sources, including SDF, CSV, and Excel files, as well as remote URLs. These data formats can contain metadata or text fields that, if maliciously crafted, could influence agent behavior during processing. The analysis confirms that while chemical standardization is performed, there is no explicit sanitization for potential prompt injection patterns within the data metadata.
Ingestion points: dm.read_sdf, dm.read_csv, dm.read_excel, and dm.open_df functions in references/io_module.md allow loading data from local and remote files.
Boundary markers: The skill does not define specific boundary markers for text metadata extracted from chemical files.
Capability inventory: The skill environment allows Python code execution, file system writes, and network access to cloud storage providers.
Sanitization: The instructions focus on chemical structure sanitization and standardization, which does not mitigate natural language prompt injection risks.

datamol