The Agent Skills Directory

[SAFE]: The skill instructions and associated scripts do not exhibit any malicious behavior. The primary logic involves using a model evaluation plugin and visualizing results via a Python script.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (CSV files) which constitutes a theoretical attack surface for indirect prompt injection. However, the provided visualization script treats these files as numerical data for plotting, and no exploitable instructions were found in the skill's logic.
Ingestion points: Data files such as actual.csv and predicted.csv are read by the visualization_script.py and potentially by the model-evaluation-suite plugin.
Boundary markers: Not explicitly defined in the prompt instructions for isolating external data.
Capability inventory: The skill has access to Bash, Read, and Write tools.
Sanitization: No specific sanitization of CSV content is performed before plotting.

evaluating-machine-learning-models