ml-pipeline-automation

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests external data for model training and validation, creating a potential surface for indirect injection if data sources are compromised.
  • Ingestion points: Python scripts in SKILL.md and references/airflow-patterns.md load CSV data from paths like /data/raw/ using pandas.
  • Boundary markers: There are no explicit delimiters or specific markers used to separate data from instructions during the file ingestion phase.
  • Capability inventory: The skill utilizes PythonOperator and BashOperator for script execution, performs network requests via Slack webhooks and MLflow tracking servers, and performs file system writes.
  • Sanitization: Extensive data quality validation logic, including schema checks and statistical validation, is provided in references/pipeline-monitoring.md as a primary mitigation.
  • [DYNAMIC_EXECUTION]: The skill uses templates to generate executable scripts and dynamically registers tasks within the orchestration environment.
  • Pattern evidence: SKILL.md contains instructions to write Python DAG definitions directly to the filesystem. references/airflow-patterns.md demonstrates dynamic DAG generation using the globals() registry.
  • [COMMAND_EXECUTION]: The skill requires the execution of shell commands to initialize Airflow databases, start background services, and trigger pipeline runs. These are standard operations for the described workflow but represent active system interaction capabilities.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 12:00 AM
Security Audit — agent-trust-hub — ml-pipeline-automation