The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's documentation and examples encourage the use of the Python pickle module for caching molecular embeddings.
Evidence: Code snippets in SKILL.md and references/examples.md demonstrate loading data using pickle.load(f) for performance optimization.
Risk: The pickle module is insecure and can be used to execute arbitrary code during deserialization. If an attacker provides a malicious pickle file, it could lead to full system compromise.
[REMOTE_CODE_EXECUTION]: The skill uses state management methods that involve deserializing configuration files from disk.
Evidence: SKILL.md describes loading state via MoleculeTransformer.from_state_yaml_file("featurizer_config.yml").
Risk: This functionality depends on the security of the underlying library's YAML and JSON parsing. Unsafe deserialization of configuration files is a known vector for code execution.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of pre-trained models from external repositories.
Evidence: The ModelStore class and PretrainedMolTransformer are used to fetch models such as ChemBERTa-77M-MLM and ChemGPT-1.2B.
Context: These operations typically target well-known services like the HuggingFace Model Hub. While these are trusted sources, they represent a dependency on external code and weights.

molfeat