CrowdSec detects attacks against MLflow, one hundred fifty-two days after the patch

MLflow Addresses Security Flaw with Patch, But Awareness Takes Time

MLflow, the open-source platform favored by data science teams for tracking experiments and deploying models, recently rolled out a significant patch on February 18. This update addressed a vulnerability that had left part of its interface accessible without authentication. Although the patch was characterized as a bug fix by its maintainers, the CVE identifier (CVE-2026-2652) was not disclosed until 85 days later. Meanwhile, CrowdSec has reported ongoing exploitation attempts since July 20.

Once a vulnerability receives a CVE identifier, it becomes public knowledge, allowing security professionals to gauge its associated risks. Unfortunately, this often centers on communication, as the exposure period effectively begins when the patch is applied. Attackers who see the fix before others gain insight into what they could have exploited just before it was secured.

The vulnerability stemmed from partial access control. An MLflow server running with simple authentication and served by Uvicorn only checked permissions for the routes within its gateway. As a result, routes that trigger tasks or receive execution logs remained open. An attacker able to reach the server could receive responses without credentials, while the administrator believed access had been restricted. For the flaw to be exploited, three conditions needed to align: simple authentication activated, a service provided by Uvicorn, and a server accessible from the outside. In contrast, an MLflow instance without any authentication is simply open to all.

Security Patch Labeled as a Bug Fix

The developer who proposed the patch titled it “adding missing authentication to FastAPI routes,” omitting any mention of the vulnerability or its potential exploitation. They marked the change as “not user-visible” and categorized it under bug fixes. The only reference to the issue was buried within the request body, lacking any explanatory context. Nevertheless, the patch closure included a third access point that neither the U.S. registry nor the security advisory disclosed—specifically, that of the platform’s conversational assistant.

This classification appears in MLflow’s version 3.10.0 release notes, published two days later. The patch was assigned the twelfth spot among roughly fifty bug fixes, without any security section, CVE identifier mention, or indication that authentication was bypassable. Alongside it, four other security patches were listed, including one for server-side request forgery and another for arbitrary file reading.

These release notes did not link to any security advisory, as none had been published in their repository, despite seven being released since 2023. However, their approach has shifted, with three advisories issued between July 27 and August 4. The project appears to be treating security patches distinctly after previously cataloging them among general fixes.

A Long Wait for Vulnerability Disclosure

The MLflow team merged the patch on February 18 and delivered it within version 3.10.0 on February 20. In contrast, the hunting authority huntr.dev did not relay the CVE identifier to the U.S. registry until May 14. By then, MLflow had already released four updates, up to version 3.12.0. Interestingly, the advisory published by the U.S. registry cites two different patched versions—3.10.0 in the body and 3.11.0 in the structured fields—leaving administrators to decide which version is the most prudent to adopt.

This delay allowed researcher DhiyaneshDk to publish a detection template on June 4, 20 days before CrowdSec activated its own rule and 46 days prior to the first observed exploitation attempt. Not all vulnerabilities suffer such a lag; for instance, the Citrix NetScaler appliance vulnerability was exploited within 24 hours. The test for the MLflow flaw involved two requests without needing a payload—simply querying an endpoint and observing the server’s response. Therefore, the challenge of exploitation has never been a protective measure; rather, it was sheer ignorance of the vulnerability that offered any safety.

Moreover, this detection template targeted MLflow routes that were already known. Just six weeks earlier, on April 3, another vulnerability affecting the same task routes had received a CVE identifier and a score of 9.1 out of 10, but no patched version has been associated with it to date. Thus, the unprotected endpoints of MLflow were publicly cataloged since April, while the associated patch remained hidden among bug fixes.

CrowdSec’s Observations and Gaps in Reporting

The French security editor has identified 129 distinct addresses and collected 352 signals over 17 days, averaging 21 signals per day with a peak of 67. Germany and France are the top targets, accounting for 20% and 19% of incidents, respectively. Additionally, CrowdSec categorizes four out of five attempts as infrastructure takeovers, with one in six linked to ransomware preparation and one in twenty searching for data.

CrowdSec is the sole source of these figures and also markets recommended firewalls and block lists at the end of its report. However, CrowdSec does not delimit its observation window with specific dates, has not disclosed the size of its sensor network, and does not clarify how it derives intent distribution. Neither GreyNoise nor Shadowserver have corroborated this activity thus far, and the CISA has not listed any MLflow vulnerabilities in its catalog of exploited flaws.

For information systems directors, there’s more to glean from MLflow than merely applying patches. Monitoring vulnerability trends can provide insights into what developers are publishing versus what they are fixing, with the discrepancy here measured in months. The pressing question for machine learning supply chain vendors remains: how do they communicate security patches without an identifier?

Scroll to Top