MLflow model lifecycle and evaluation platform

Record description

MLflow supports experiment tracking and generative-AI evaluation. A 2026 advisory concerns scorer keyword arguments reaching LiteLLM-related network functionality through unauthenticated evaluation endpoints under certain exposure conditions. An MLflow scorer is not itself evidence that autonomous models launched attacks.

The GenAI scorer incident concerns an evaluation endpoint that could promote requester-controlled scorer settings into server-originated outbound requests through LiteLLM. The distinctive dependency is the model-evaluation workflow's network authority, not an LLM selecting malicious targets. This is why hardening a generic scoring API and restricting the evaluator's egress are related but separate defenses. MLflow advisory.

Explicitly related research (4)

Declared connections