The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The FeaturePipeline class in references/feature-engineering.md (and similar logic in other reference files) utilizes pickle.load() to deserialize saved pipeline objects and models from the file system. Deserializing untrusted data via pickle is a known vector for arbitrary code execution if the source file is compromised.
[COMMAND_EXECUTION]: The skill generates implementation code for orchestration frameworks including Apache Airflow, Kubeflow Pipelines, and Prefect. These frameworks are designed to execute complex Directed Acyclic Graphs (DAGs) of tasks, which may involve running shell commands or arbitrary Python code in distributed environments.
[DATA_EXFILTRATION]: The skill includes components that interact with external services and shared file systems. For example, deploy_model in references/pipeline-orchestration.md uploads model artifacts to Google Cloud Storage (GCS) buckets, and ShadowDeployment in references/model-validation.md writes prediction data and features to /var/log/shadow_predictions.jsonl. While these are standard MLOps practices, they involve moving potentially sensitive model data and features to external or shared locations.

ml-pipeline