The Agent Skills Directory

[SAFE]: The skill demonstrates a strong security posture by explicitly warning against the use of Python's pickle module for deserializing untrusted data in references/workflows_and_best_practices.md. It correctly identifies the risk of arbitrary code execution and recommends safer alternatives like SMILES, SDF, or RDKit's native binary format for local caches.
[SAFE]: All provided scripts (scripts/molecular_properties.py, scripts/similarity_search.py, scripts/substructure_filter.py) perform local chemical analysis and data processing. There are no indications of unauthorized network activity, credential harvesting, or access to sensitive system directories.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it is designed to ingest and process untrusted molecular data files.
Ingestion points: The function read_molecule_records in scripts/_common.py reads data from SMILES, MOL, and SDF files.
Boundary markers: The skill does not implement explicit delimiters or warnings to the agent regarding instructions potentially embedded within molecular names or properties.
Capability inventory: The skill's primary capabilities are scientific calculations and local file writes (CSV, SDF, SMILES). It does not contain capabilities for subprocess execution, dynamic code evaluation (eval/exec), or outbound network operations that could be abused via data injection.
Sanitization: The skill relies on RDKit's robust internal parsers and Chem.SanitizeMol to validate the chemical integrity of the input data, which effectively prevents non-chemical data from being processed as molecules.
[SAFE]: The installation instructions target the legitimate rdkit package on PyPI and use well-known package managers like uv and conda, following standard developer workflows for scientific software.

rdkit