engineering-features-for-machine-learning

Warn

Audited by Gen Agent Trust Hub on Jul 21, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The script scripts/feature_importance_analyzer.py contains deceptive metadata. While its internal docstring, header, and help description claim it performs feature importance analysis (SHAP and permutation importance), the actual code is a generic directory scanner that recursively traverses the filesystem. This discrepancy is a form of metadata poisoning that could mislead the AI agent into executing filesystem operations under false pretenses.
  • [COMMAND_EXECUTION]: The skill documentation in scripts/README.md lists multiple core scripts (e.g., feature_engineering_pipeline.py, data_visualizer.py) as completed and existing, but they are not present in the provided file set. This lack of package integrity, combined with the deceptive script, presents a risk of unintended behavior when the agent attempts to invoke missing or mislabeled tools.
  • [PROMPT_INJECTION]: The skill processes external data (e.g., assets/example_dataset.csv), creating an indirect prompt injection surface.
  • Ingestion points: Data is loaded via pd.read_csv() in assets/feature_engineering_template.py.
  • Boundary markers: Absent; there are no instructions to prevent the agent from interpreting content within the data as directives.
  • Capability inventory: The skill possesses extensive execution and filesystem traversal capabilities through the Bash(cmd:*) tool and the feature_importance_analyzer.py script.
  • Sanitization: No validation or sanitization of input data content is implemented in the provided scripts.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 21, 2026, 11:31 AM
Security Audit — agent-trust-hub — engineering-features-for-machine-learning