The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill demonstrates model persistence using joblib.load and mlflow.sklearn.load_model. These functions rely on Python's pickle serialization mechanism, which is fundamentally insecure. If an attacker can replace the model artifact files in the artifacts/ or mlruns/ directories, the agent or user would execute arbitrary code upon loading the model.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface for indirect prompt injection by processing external data and possessing file-writing capabilities.\n
Ingestion points: The skill processes dataframes clean_df and X_new which represent external data input (SKILL.md).\n
Boundary markers: No delimiters or explicit instructions to ignore embedded commands are used when processing data.\n
Capability inventory: The skill has file system write access via joblib.dump, plt.savefig, and Path.write_text, as well as network capability via mlflow tracking (SKILL.md).\n
Sanitization: No data sanitization, validation, or structural constraints are mentioned for the ingested data.

scikit-learn