The tracking server is the MLflow service that receives experiment run data, stores metadata about runs, and exposes the main interface for browsing and comparing training results. In practice, it is the control point for experiment tracking: clients send parameters, metrics, tags, and artifact references to it, and users query it to review what happened during a model run.
In cyber security, this matters because the tracking server often sits in the trust path for an entire MLOps workflow. If it is exposed to untrusted networks or misconfigured, attackers may be able to tamper with run history, overwrite artifacts, leak sensitive experiment details, or abuse the service as an entry point into connected stores and systems. Defenders should treat it like any other management interface: restrict network access, require authentication, keep it patched, and monitor logs for unusual API requests or unexpected changes to runs and artifacts.