Patchfirst
CVE-2026-64849 KEVPublic exploit Critical · 9.3 EPSS 10.0%

MLflow Server-Side Request Forgery Vulnerability

MLflow · MLflow

Summary

The flaw is a server-side request forgery in the MLflow platform's webhook test functionality, where the webhook delivery component follows redirects and re-resolves hostnames after only the original URL has been validated. An attacker with network access can send an unauthenticated request to the webhook test endpoint, without any user interaction or special configuration, to make the server reach internal or cloud metadata services and return the response status and body to the attacker.

Official description · NVD

MLflow is an open source AI engineering platform for agents, large language models, and machine learning models. Starting in 3.3.0 and prior to 3.15.0, the unauthenticated POST /api/2.0/mlflow/webhooks/{id}/test endpoint calls _validate_webhook_url() in mlflow/utils/validation.py only for the original URL while mlflow/webhooks/delivery.py follows redirects and re-resolves the hostname without pinning the validated address, allowing attackers to reach internal or cloud metadata services and receive response_status and response_body. This issue is fixed in version 3.15.0.

Breakdown
78.6 / 100
Exploited
Yes, in CISA KEV
Reports
0
Tactics
2/15

How an attack unfolds

Based on mappings. Select a step to see the evidence.
  1. How attackers get in

  2. What they do

First seen exploited 1 day after disclosure. Public exploit code (PoC-in-GitHub) appeared 2 days after disclosure.

  1. Disclosedday 0
  2. First seen exploitedVulnCheck KEV evidence+1 day
  3. Added to CISA KEVEU database agreesPublic exploit codePoC-in-GitHub+2 days

CISA KEV listing is when exploitation was confirmed, not necessarily the first attack.

7 fixes in mlflow over 7 months. The latest fix listed for mlflow is 3.17.0.

Fixes in mlflowPyPI package, by fixed version. Earlier and later fixes in the same package, not necessarily the same bug.

  1. 3.11.0CVE-2026-413718 May 2026
  2. 3.11.1CVE-2026-338657 Apr 2026
  3. 3.13.0rc0CVE-2026-81472 Jul 2026
  4. 3.15.0CVE-2026-6484917 Aug 2026KEV
  5. 3.17.0CVE-2026-11588623 Oct 2026
  6. 3.17.0CVE-2026-11669027 Oct 2026
  7. 3.17.0CVE-2026-11669127 Oct 2026Latest listed

techniques

2 , grouped by in kill-chain order, 2 cited passages 2

Initial AccessTA0001

Exploit Public-Facing Application T1190 Medium
AI mapping from the description · how it is exploited

unauthenticated public-facing API endpoint exploited

the unauthenticated POST /api/2.0/mlflow/webhooks/{id}/test endpoint
From the official description (NVD) Original (opens in a new tab) ↗

Command and ControlTA0011

Proxy T1090 Medium
AI mapping from the description · primary impact

SSRF proxies requests to internal services

allowing attackers to reach internal or cloud metadata services
From the official description (NVD) Original (opens in a new tab) ↗

Detection

0 rules for this CVE · 171 general rules for its techniques

Rules for this CVE

0 that reference CVE-2026-64849 directly.

CVE-specific rules are not available in this build. Search SigmaHQ ↗

Sigma rules © SigmaHQ, Detection Rule License 1.1; linked, not reproduced.

Am I affected? What do I upgrade to?

From GitHub Security Advisories and OSV

1 package in 1 ecosystem. Main package: mlflow.

Upgrade to 3.17.0 or later

Fixes this CVE and 10 other known vulnerabilities in this package (CVE-2023-1177, CVE-2023-3765, +8).

pippip install "mlflow>=3.17.0"

First fixed in 3.15.0 for CVE-2026-64849.

Affected versions

If you runUpgrade to
All versions before 3.15.03.17.0first fix for this CVE: 3.15.0

“Up to” means the second version is the fix and is not affected.

Check your lockfile against all CVEs

What security researchers reported

Not named in any threat report yet

Also mentioned in 31 community threat pulses (AlienVault OTX)

Weaker evidence

Pulses are indicator collections that anyone can publish. A mention is not confirmation of exploitation.

References

6 links, grouped by NVD tag
↑ ↓ chooseEnter openEsc close