The Agent Skills Directory

DATA_EXFILTRATION (MEDIUM): The references/io_module.md and references/descriptors_viz.md files describe functions like dm.save_df, dm.to_sdf, and dm.viz.to_image which support an outfile parameter. The documentation explicitly states that these functions support remote file paths through fsspec integration, covering protocols such as S3, GCS, Azure, and HTTP/HTTPS. This capability allows an agent to exfiltrate data from the local environment to an external endpoint.
EXTERNAL_DOWNLOADS (LOW): The library's I/O module (references/io_module.md) provides functions like dm.read_sdf, dm.read_csv, and dm.open_df that can fetch and process molecular data from arbitrary remote URLs. This creates a vector for ingesting untrusted content into the agent's context.
PROMPT_INJECTION (LOW): The skill is vulnerable to Indirect Prompt Injection due to its extensive data ingestion surface.
Ingestion points: dm.read_sdf, dm.read_smi, dm.read_csv, dm.read_excel, dm.read_mol2file, dm.read_pdbfile, and dm.open_df (documented in references/io_module.md).
Boundary markers: Absent. The documentation does not suggest using delimiters or instructions to ignore embedded natural language commands within molecular data files.
Capability inventory: The skill can perform network writes (via fsspec), local file writes (dm.save_df), and complex visualizations (dm.viz.to_image).
Sanitization: While the library performs chemical sanitization (fixing valency, aromaticity), it does not provide sanitization for natural language instructions that may be embedded in metadata fields (e.g., SDF tags or CSV columns).

datamol