Server-Side Request Forgery in the MLflow tracking server via client-supplied artifact locations and model version sources
CVE-2026-118628 · GHSA-mr9f-g8qf-4w4j · published 2026-11-13
- Severity
- High
- CVSS score
- 8.2
- Chance of exploitation (EPSS)
- Not scored
- Fixed in
- 3.18.0
- Exploitation No known exploitation
- Automatable Yes
- Technical impact Partial
- CISA SSVC
Internal reachability and port probing. The server opens a TCP connection to any host and port the client names. File read from internal file services. Against an internal FTP, SFTP or HDFS service that accepts the login supplied in the URI (or anonymous access), the client can list and download files through the artifact endpoints. File write to the same services. /ajax-api/2.0/mlflow/upload-artifact stores client-supplied content (up to 10 MB) through the same repository. Reading internal HTTP responses under --no-serve-artifacts. …
Upgrade mlflow to 3.18.0 or later.
Upgrade to MLflow 3.18.0 or later. 3.17.0 closes the paths described above (https://github.com/mlflow/mlflow/pull/25891). The tracking server refuses to connect to host-addressed artifact locations (ftp, sftp, hdfs, viewfs, http, https, host-bearing mlflow-artifacts, r2, b2, abfss) that name a host other than its configured --default-artifact-root or --artifacts-destination. The check applies when a client submits a location or source, when a stored location is used to serve artifacts, and inside server-side jobs. 3.18.0 closes a bypass of that check through the wasbs:// scheme (https://github.com/mlflow/mlflow/pull/27396). …
No configuration option disables the behavior in affected versions. Until upgrading: Do not expose the tracking server to untrusted networks, and enable authentication so that only trusted users can create experiments and model versions. Restrict the server's outbound network access to its artifact storage, and block internal ranges and the cloud metadata endpoint.
| Package | Registry | Affected | Fixed in |
|---|---|---|---|
| mlflow | PyPI | < 3.18.0 | 3.18.0 |
- Critical@zereight/mcp-gitlab Vulnerable to Server-Side Request Forgery@zereight/mcp-gitlab
- CriticalSSRF + LLM API key exfiltration via user-controlled basepath / baseUrl in 8 chat-model nodesflowise, flowise-components
- CriticalLocalAI POST /models/apply permits unauthenticated server-side request forgery through gallery URLsgithub.com/mudler/LocalAI
- CriticalFastMCP OpenAPI Provider has an SSRF & Path Traversal Vulnerabilityfastmcp
- CriticalBentoML SSRF Vulnerability in File Upload Processingbentoml
- Criticallobe-chat `/api/proxy` endpoint Server-Side Request Forgery vulnerability@lobehub/chat
- CriticalMLflow: Unauthenticated full-read SSRF in webhook delivery: _validate_webhook_url bypassed via unvalidated HTTP redirects (and DNS rebindingmlflowExploited
- CriticalMLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifacmlflow
- CriticalMLflow: Environment variable injection in AI Gateway secrets enables server-side credential exfiltrationmlflow
- CriticalMLflow allows unauthorized access to multipart upload endpoints when the `--serve-artifacts` mode is enabledmlflow
- CriticalMLflow: Improper Origin Validation in MLflow Assistant /ajax-api Endpoints Enables Browser-Mediated Local Command Executionmlflow
- Criticalmlflow: FastAPI job endpoints under `/ajax-api/3.0/jobs/*` are not protected by authentication or authorizationmlflow
The impact, patch and workaround text comes from the advisory, shortened (published by MLflow). Proof-of-concept and reproduction sections and code are left out; the original advisory has them.