The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches pretrained model weights and configurations from Hugging Face (e.g., ChemBERTa and MolFormer) as part of its transfer learning functionality. These are sourced from well-known repositories and are standard for molecular ML tasks.
[COMMAND_EXECUTION]: The provided Python scripts perform model training and evaluation using the DeepChem API. All operations, including dataset loading and model fitting, are constrained to the intended scientific workflows.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill reads external data via CSVLoader, SDFLoader, FASTALoader, and JsonLoader (SKILL.md, api_reference.md).
Boundary markers: Not explicitly defined for raw molecular data.
Capability inventory: The skill performs model training and saves results but lacks dangerous capabilities like arbitrary shell execution or network exfiltration.
Sanitization: Data is processed through established scientific parsers (RDKit/DeepChem) which are not susceptible to traditional instruction-based injection.
[SAFE]: Extensive review found no evidence of obfuscated code, persistence mechanisms, or metadata poisoning. The skill's behavior aligns perfectly with its stated purpose of chemical property prediction.

deepchem