The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation and examples recommend using the Python pickle module for caching molecular embeddings. Using pickle.load() is insecure as it can be exploited to execute arbitrary code if the cache file is supplied by an untrusted source or tampered with.
[REMOTE_CODE_EXECUTION]: The skill provides methods to load transformer configurations from YAML and JSON files. This functionality can be a vector for code execution if the library utilizes unsafe deserialization methods (such as yaml.load() without a safe loader) when reading state files.
[EXTERNAL_DOWNLOADS]: The skill facilitates the automated download of pre-trained machine learning models and embeddings from external providers like HuggingFace Hub, Microsoft Research, and DGL-LifeSci. While these are established services, the skill dynamically fetches and loads binary model data at runtime.
[PROMPT_INJECTION]: The skill processes molecular SMILES strings from external datasets, creating a surface for indirect prompt injection. Maliciously crafted molecular data could influence downstream agent behavior if the resulting features are interpolated into prompts without sanitization.
Ingestion points: SMILES strings processed by MoleculeTransformer and PretrainedMolTransformer.
Boundary markers: Default examples do not specify the use of delimiters or instructions to ignore embedded content.
Capability inventory: The skill uses parallel processing (multiprocessing) and file system writes for persisting featurizer states.
Sanitization: Molecular standardization and salt removal are listed as optional preprocessing steps rather than default security controls.

molfeat