ml-experiment-tracker
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill contains Python and shell templates for managing experiments that utilize system commands in a secure manner.
references/reproducibility-checklist.mdincludes Python functions usingsubprocess.check_outputwith static, hardcoded arguments to retrieve Git metadata (e.g.,git rev-parse HEAD).SKILL.mdprovides standard shell command examples for DVC operations such asdvc init,dvc add, anddvc push.- [EXTERNAL_DOWNLOADS]: The skill references standard machine learning packages and official container images from well-known services.
references/mlflow-setup.mdprovides instructions for installingmlflow,psycopg2-binary, andboto3from standard registries.- The same file includes a Docker Compose example referencing official images from GitHub Container Registry (
ghcr.io/mlflow/mlflow) and the official PostgreSQL image. - [CREDENTIALS_UNSAFE]: The documentation contains placeholder and default credentials used solely for educational purposes in local development examples.
references/mlflow-setup.mduses placeholder strings likepostgresql://user:pass@localhost:5432/mlflowand default local credentials (POSTGRES_PASSWORD=mlflow) in a demonstration environment configuration.- [INDIRECT_PROMPT_INJECTION]: The skill describes workflows for processing experiment artifacts and metrics, which represents a potential attack surface if an agent reads malicious or malformed log data.
- Ingestion points: The workflows involve reading metrics, configurations, and artifacts produced during model training (e.g.,
mlflow.log_artifact,wandb.log). - Boundary markers: The provided templates do not specify delimiters for external experiment data.
- Capability inventory: The skill uses file system access and network operations for tracking, alongside static shell command execution.
- Sanitization: No explicit sanitization logic for logged metrics is present in the provided documentation snippets.
Audit Metadata