OffSec — MLflow LFI via URI fragment (CVE-2024-2928)

AI relevance: MLflow is a common ML lifecycle platform; LFI in its artifact handling can expose model artifacts and AI pipeline credentials.

  • CVE-2024-2928 affects MLflow versions ≤ 2.11.2; patched in 2.11.3.
  • The LFI stems from improper sanitization of the URI fragment (#) during artifact path resolution.
  • MLflow parses the raw HTTP request line, which includes fragments, bypassing the usual path traversal protections.
  • Crafted fragments like #../etc/passwd can escape the artifact directory and read arbitrary files.
  • Attack is unauthenticated and remote if the MLflow server is exposed to the network.
  • Browsers do not send fragments, so exploitation requires raw HTTP tooling (curl, Burp, netcat).
  • CVSS v3.1 score: 7.5 (high).

Why it matters

  • MLflow often co-locates model artifacts, pipeline configs, and service credentials; LFI turns it into a secret-harvesting endpoint.
  • Artifact stores are frequently tied to cloud IAM roles, so leaked credentials can pivot into broader AI infrastructure.

What to do

  • Upgrade MLflow to 2.11.3+ immediately.
  • Restrict network exposure of the tracking server and require authentication/allowlists.
  • Add edge filtering/WAF rules to block #-fragment traversal patterns in artifact routes.
  • Monitor logs for suspicious artifact path requests with fragment traversal.

Sources