The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its data ingestion surfaces.
Ingestion points: The script scripts/analyze_dataset.py reads user-provided CSV files for analysis. Reference patterns in references/classical-ml.md and references/transformers-patterns.md perform data loading from CSV and JSON formats.
Boundary markers: No boundary markers or delimiters are implemented to separate data from instructions within the processed datasets.
Capability inventory: The skill can execute local scripts (scripts/analyze_dataset.py), write files to disk (model exports in references/deployment.md), and perform network operations (fetching models/datasets from HuggingFace).
Sanitization: No validation or filtering is performed on data content to prevent instruction hijacking.
[REMOTE_CODE_EXECUTION]: Reference patterns include the use of joblib.load for model persistence, which poses a risk of arbitrary code execution if an agent is induced to load an untrusted model file.
Evidence: The file references/classical-ml.md provides a load_model utility that uses joblib.load to deserialize model artifacts.
[EXTERNAL_DOWNLOADS]: The skill references and downloads models and datasets from well-known and trusted platforms.
Evidence: Patterns in references/fine-tuning.md and references/transformers-patterns.md utilize from_pretrained and load_dataset to fetch assets from HuggingFace.

engineering-ml-engineer