Skip to content
Server Security Center
KEVcritical9.3

CVE-2026-64849: MLflow Server-Side Request Forgery

MLflow · listed in the CISA catalog of actively exploited vulnerabilities since 2026-08-19.

CVSS v3.1 (CVE record)
9.3
critical
EPSS (exploit probability, 30 days)
11%
higher than 95.9% of all CVEs
Added to CISA KEV
2026-08-19
US federal due date 2026-09-02
Automatable (CISA SSVC)
yes
technical impact: partial

CVE-2026-64849 explained

CVE-2026-64849 is a server-side request forgery flaw in MLflow, an open source AI engineering platform. The webhook test endpoint validates only the original URL, then follows redirects and re-resolves hostnames, letting attackers reach internal or cloud metadata services. CISA lists it as known exploited, so affected servers need prompt attention.

Impact

An attacker can make the MLflow server send requests to internal or cloud metadata services and read the returned status and response body. On deployments without server authentication, no credentials are needed. With built-in basic authentication enabled, valid credentials are required. The flaw does not bypass authentication. CISA rates it automatable with partial technical impact. Confidentiality impact is high and integrity impact low, and the impact reaches beyond MLflow itself.

Who is affected

The CVE record lists mlflow from 3.3.0 up to but not including 3.15.0 as affected. The structured fixedIn data in the CVE record names no fixed version. The description, however, states that the issue is fixed in version 3.15.0. Check the vendor advisory to confirm the fixed release before upgrading.

Written with AI from the data on this page and checked against it automatically.

Official description

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 POST /api/2.0/mlflow/webhooks/{id}/test endpoint (accessible without credentials when server authentication is not enabled) 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.

The unauthenticated attack scenario described here applies to deployments without authentication. With MLflow’s built-in basic authentication enabled, requests to this endpoint require valid credentials; this issue does not itself bypass authentication.

This issue is fixed in version 3.15.0.

CISA required action: Apply mitigations in accordance with vendor instructions, ensuring compliance with CISA’s BOD 26-04 Prioritizing Security Updates Based on Risk guidance and CISA’s “Forensics Triage Requirements”. Follow applicable BOD 26-04 guidance for cloud services or discontinue use of the product if mitigations are unavailable. Stakeholders are responsible for evaluating each asset's internet exposure and ensuring adherence to BOD 26-04 patching guidelines.

CISA notes

Affected and fixed versions

From the CVE record published by the vendor or CVE numbering authority.

  • mlflow
    Affected
    • ≥ 3.3.0
    • < 3.15.0
    Fixed in
    not stated

How to fix it

  • The CVE record names no fixed version – check the vendor advisory and CISA's required action for updates and mitigations.
  • Vendor advisory: github.com/mlflow/mlflow/security/advisories/GHSA-7gwp-5p…
  • CISA's deadline for US federal agencies: 2026-09-02 (passed) – a useful benchmark for everyone else, too.

What to do

  1. Upgrade MLflow to 3.15.0, which the description names as the fixed version, after confirming it in the vendor advisory.
  2. Enable MLflow's built-in authentication so the webhook test endpoint cannot be reached without valid credentials.
  3. Do not expose MLflow servers directly to the internet. Restrict access to trusted networks and users.
  4. Limit outbound traffic from MLflow servers to internal services and cloud metadata endpoints where possible.
  5. Review logs and webhook configurations for signs of compromise, following CISA's required action and forensics triage guidance.

The structured CVE data names no fixed version, though the description states the issue is fixed in 3.15.0. Confirm this in the vendor advisory. Until you upgrade, enable authentication, restrict network access to the server, and limit its outbound access. CISA requires applying vendor mitigations, or discontinuing use of the product if mitigations are unavailable.

Written with AI from the data on this page and checked against it automatically.

CISA's assessment (SSVC)

Exploitation
Exploited in the wild
Automatable
Yes – attackers can automate it at scale
Technical impact
Partial – limited control or information

Weakness: CWE-918 Server-Side Request Forgery

Questions about CVE-2026-64849

Is this vulnerability being exploited?
Yes. CISA added CVE-2026-64849 to its Known Exploited Vulnerabilities catalog on 2026-08-19. The due date for US federal agencies is 2026-09-02. Known ransomware use is listed as unknown.
Does enabling authentication protect my server?
It blocks the unauthenticated scenario. With MLflow's built-in basic authentication enabled, the endpoint requires valid credentials, and the issue does not bypass authentication. Authenticated users could still trigger it, so upgrading remains necessary.
Which versions should I run?
The description states the issue is fixed in 3.15.0. Versions from 3.3.0 up to but not including 3.15.0 are affected.

Written with AI from the data on this page and checked against it automatically.

References

Data: CISA Known Exploited Vulnerabilities (public domain) · CVE Program records incl. CISA ADP enrichment · OSV.dev (CC BY 4.0) incl. Ubuntu and Debian security data · FIRST EPSS · endoflife.date. Updated automatically – facts last changed 2026-11-21. Always confirm with the vendor's or distribution's advisory before patching.