The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation in SKILL.md (within the 'Performance Optimization' section) recommends using the pickle module to save and load molecules. Using pickle.load() on data from an untrusted source is a significant security risk, as it allows for the execution of arbitrary code during deserialization.
Evidence: The following code snippet is provided in SKILL.md: with open('molecules.pkl', 'rb') as f: mols = pickle.load(f)
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process molecular data from external files (e.g., SDF, SMILES, PDB). These formats can contain metadata fields or molecular labels that could be manipulated to include malicious instructions targeting the agent's logic during property calculation or filtering tasks.
Ingestion points: The scripts scripts/molecular_properties.py, scripts/similarity_search.py, and scripts/substructure_filter.py all utilize Chem.SDMolSupplier and Chem.SmilesMolSupplier to read data from external files.
Boundary markers: The skill does not provide instructions to the agent to treat molecular metadata or property fields as untrusted data or to ignore embedded natural language instructions.
Capability inventory: The skill scripts have the capability to write output files to the filesystem (SDF, SMILES, and CSV formats) and perform extensive analysis that influences the agent's response.
Sanitization: While RDKit performs 'sanitization' of chemical structures (validating valence and aromaticity), this is a scientific validation and does not provide security sanitization against malicious prompt content.
[EXTERNAL_DOWNLOADS]: The skill provides documentation for installing RDKit via standard package managers. RDKit is a well-known and established scientific library, and these references are considered safe distribution channels for the toolkit.
Evidence: references/api_reference.md includes commands for conda install -c conda-forge rdkit and pip install rdkit-pypi.

rdkit