The Agent Skills Directory

[SAFE]: The skill provides security-conscious guidance by documenting the transition from unsafe pickle-based serialization to safer alternatives like skops and torch.export. It explicitly mentions security warnings associated with pickle and instructs on how to handle legacy models safely.
[SAFE]: The skill documents secure handling of sensitive data, such as using environment variables for webhook secrets (MLFLOW_WEBHOOK_SECRET) rather than hardcoding credentials.
[SAFE]: The skill includes instructions on disabling anonymized usage telemetry in production environments for privacy and compliance.
[DYNAMIC_EXECUTION]: The skill documents the code_paths feature in mlflow.pyfunc.log_model, which allows users to bundle custom Python code with model artifacts. This code is intended to be executed at model load time as part of the model's prediction logic.
[COMMAND_EXECUTION]: The skill describes the use of CLI tools like mlflow models build-docker and mlflow models serve for deploying models into production environments.
[REMOTE_CODE_EXECUTION]: The skill documents mlflow.models.predict, which rebuilds a model's environment and executes its prediction logic in an isolated environment. This is a legitimate feature used for validating model artifacts before deployment.
[INDIRECT_PROMPT_INJECTION]: The skill describes the management of model metadata, registry tags, and evaluation data which are ingested into the MLflow tracking system.
Ingestion points: references/reg-gate-state-in-tags.md (registry tags), references/eval-gate-with-validate-evaluation-results.md (evaluation metrics).
Boundary markers: Not explicitly mentioned.
Capability inventory: references/reg-per-environment-registered-models.md (copy_model_version), references/serve-predict-before-deploy.md (mlflow.models.predict).
Sanitization: Not explicitly mentioned, though the skill focuses on structured MLOps APIs.

mlflow-3