The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external chemical and biological data files, which creates a surface for indirect prompt injection. Maliciously crafted data could potentially influence the agent's logic when it processes or reports on the file contents. \n
Ingestion points: The skill uses dc.data.CSVLoader, dc.data.SDFLoader, and dc.data.FASTALoader in SKILL.md and across scripts in the scripts/ directory to load external molecular datasets. \n
Boundary markers: No specific delimiters or instructions to ignore embedded instructions are implemented in the data loading workflows. \n
Capability inventory: The skill is configured with Bash(python:*), Bash(uv:*), and file system Read, Write, and Edit tools in the SKILL.md frontmatter. \n
Sanitization: Input data is processed for featurization without explicit sanitization of text fields that might be reflected in agent outputs or logs. \n- [EXTERNAL_DOWNLOADS]: The skill downloads standardized benchmark datasets and pre-trained models from established repositories. \n
Evidence: scripts/transfer_learning.py and references/workflows.md document fetching models from Hugging Face (e.g., seyonec/PubChem10M_SMILES_BPE_60k and Rostlab/prot_bert) and datasets from MoleculeNet. These are recognized as well-known services and trusted academic resources.

alterlab-deepchem