building-automl-pipelines

Pass

Audited by Gen Agent Trust Hub on Apr 20, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes external training datasets without explicit security boundaries.
  • Ingestion points: Training data loaded via the Read tool from external CSV, Parquet, or database sources as instructed in SKILL.md and references/implementation.md.
  • Boundary markers: Absent; there are no instructions for the agent to ignore or delimit potentially malicious commands embedded within training datasets.
  • Capability inventory: Capabilities include Bash(python:*), Write, and Edit, as well as file system operations (shutil.copy2) within scripts/pipeline_deployment.py.
  • Sanitization: No explicit sanitization or validation logic is provided to filter instructions from ingested data.
  • [COMMAND_EXECUTION]: The skill includes a Python script that performs file system tasks.
  • Evidence: The scripts/pipeline_deployment.py utility uses the shutil and os libraries to create directories and copy files from a source to a target.
  • [DATA_EXFILTRATION]: The deployment utility could be misused to transfer sensitive data if the agent is manipulated.
  • Evidence: The scripts/pipeline_deployment.py script accepts arbitrary source and target paths as command-line arguments. Without path restrictions, it could be used to copy sensitive configuration or credential files into user-accessible locations.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 20, 2026, 01:18 PM