The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes clinical text and transcriptions which serves as an ingestion point for untrusted data that could contain malicious instructions.
Ingestion points: references/datasets.md describes the MedicalTranscriptionsDataset which ingests clinical notes and transcriptions.
Boundary markers: No explicit boundary markers or instructions to ignore embedded content are specified during the ingestion of text data.
Capability inventory: The skill allows training and evaluation of machine learning models and writing results/models to the local file system using trainer.save() in SKILL.md.
Sanitization: Basic text preprocessing (lowercasing, punctuation removal) is described in references/preprocessing.md, but this does not prevent instruction-based attacks.
[DYNAMIC_EXECUTION]: Several healthcare-specific models utilize serialized data formats for structural components, which presents a surface for unsafe deserialization if files originate from untrusted sources.
Evidence: references/models.md indicates that models such as GAMENet and SafeDrug load adjacency matrices and molecular graphs from .pkl (Pickle) files provided via user-specified paths (e.g., ddi_adj_path).

pyhealth