The Agent Skills Directory

[SAFE]: The skill implements standard molecular machine learning workflows using the legitimate DeepChem library. No malicious intent, obfuscation, or unauthorized data access was detected. Findings related to data ingestion and external downloads are consistent with the skill's primary purpose and use trusted sources.
[EXTERNAL_DOWNLOADS]: The skill downloads benchmark datasets and pretrained models from well-known and reputable sources.
Evidence: dc.molnet.load_* functions in scripts/graph_neural_network.py and scripts/predict_solubility.py fetch datasets from the official DeepChem MoleculeNet repository.
Evidence: dc.models.HuggingFaceModel in scripts/transfer_learning.py and references/workflows.md downloads models from Hugging Face, such as seyonec/ChemBERTa-zinc-base-v1 and Rostlab/prot_bert.
[INDIRECT_PROMPT_INJECTION]: The skill provides the ability to ingest and process chemical data from external files, which is a common pattern in scientific AI agents.
Ingestion points: CSVLoader, SDFLoader, and FASTALoader (located in SKILL.md and scripts/) allow reading SMILES, structural, and sequence data.
Boundary markers: Not present, but the data is processed into numerical tensors rather than interpreted as natural language instructions.
Capability inventory: The skill can read local files and perform intensive mathematical computations (training/inference), but lacks direct shell command injection capabilities.
Sanitization: Data is validated and featurized using DeepChem's built-in chemical validators and featurizers.
[DYNAMIC_EXECUTION]: The skill uses dynamic function calls to handle different datasets.
Evidence: getattr(dc.molnet, f'load_{dataset_name}') in scripts/graph_neural_network.py dynamically loads the requested MoleculeNet benchmark. This is safe as the input is validated against a whitelist of supported datasets defined in the script's argument parser.

deepchem