CVE-2024-3099 — MEDIUM (CVSS 5.4) AI Security Vulnerability

CISO Take

An authenticated low-privileged user can register a model with a URL-encoded name that collides with an existing production model, silently redirecting pipelines and analysts to a poisoned version. If you run shared MLflow Model Registry instances—common in data science platforms and CI/CD pipelines—restrict model creation to trusted roles and audit existing registries for URL-encoded name variants immediately. Upgrade MLflow past 2.11.1.

Risk Assessment

CVSS 5.4 understates operational risk for ML-heavy organizations. The attack is trivially executable by any authenticated user with low privileges, requires no special AI knowledge, and produces silent model substitution—a high-impact outcome for production inference pipelines. Exposure is highest in multi-tenant or shared MLflow deployments where least-privilege is not enforced on the model registry.

Affected Systems

Package	Ecosystem	Vulnerable Range	Patched
mlflow	pip	—	No patch
25.7K OpenSSF 4.5 624 dependents Pushed 7d ago 24% patched ~64d to patch Full package profile →

Do you use mlflow? You're affected.

Severity & Risk

CVSS 3.1

5.4 / 10

EPSS

0.1%

chance of exploitation in 30 days

Higher than 19% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, CISA SSVC, EPSS, trickest/cve, and Nuclei templates.

Attack Surface

AV Network

AC Low

PR Low

UI None

S Unchanged

C None

I Low

A Low

Recommended Action

5 steps

Upgrade MLflow to a patched version that validates and normalizes model names on registration.
Audit the model registry: query for entries containing URL-encoded characters (%[0-9A-Fa-f]{2} pattern) and remove unauthorized entries.
Restrict model creation to trusted roles—revoke registry write access from general authenticated users.
Implement model artifact hash validation in pipelines consuming models from the registry to detect silent substitution.
Enable audit logging on the MLflow API and alert on model registration events for sensitive model names.

CISA SSVC Assessment

Decision Track*

Exploitation poc

Automatable No

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

Classification

Model Poisoning DoS Framework Model AML.T0010.001 - AI Software AML.T0018.000 - Poison AI Model AML.T0029 - Denial of AI Service AML.T0049 - Exploit Public-Facing Application AML.T0074 - Masquerading

Compliance Impact

This CVE is relevant to:

EU AI Act

Art. 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.3.2 - Organizational roles and access control for AI systems A.7.3 - Data and model integrity controls

NIST AI RMF

MANAGE 2.2 - Mechanisms to sustain value of deployed AI systems

OWASP LLM Top 10

LLM03 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2024-3099?

An authenticated low-privileged user can register a model with a URL-encoded name that collides with an existing production model, silently redirecting pipelines and analysts to a poisoned version. If you run shared MLflow Model Registry instances—common in data science platforms and CI/CD pipelines—restrict model creation to trusted roles and audit existing registries for URL-encoded name variants immediately. Upgrade MLflow past 2.11.1.

Is CVE-2024-3099 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2024-3099, increasing the risk of exploitation.

How to fix CVE-2024-3099?

1. Upgrade MLflow to a patched version that validates and normalizes model names on registration. 2. Audit the model registry: query for entries containing URL-encoded characters (%[0-9A-Fa-f]{2} pattern) and remove unauthorized entries. 3. Restrict model creation to trusted roles—revoke registry write access from general authenticated users. 4. Implement model artifact hash validation in pipelines consuming models from the registry to detect silent substitution. 5. Enable audit logging on the MLflow API and alert on model registration events for sensitive model names.

What systems are affected by CVE-2024-3099?

This vulnerability affects the following AI/ML architecture patterns: MLOps registries, model serving, training pipelines.

What is the CVSS score for CVE-2024-3099?

CVE-2024-3099 has a CVSS v3.1 base score of 5.4 (MEDIUM). The EPSS exploitation probability is 0.06%.

Technical Details

NVD Description

A vulnerability in mlflow/mlflow version 2.11.1 allows attackers to create multiple models with the same name by exploiting URL encoding. This flaw can lead to Denial of Service (DoS) as an authenticated user might not be able to use the intended model, as it will open a different model each time. Additionally, an attacker can exploit this vulnerability to perform data model poisoning by creating a model with the same name, potentially causing an authenticated user to become a victim by using the poisoned model. The issue stems from inadequate validation of model names, allowing for the creation of models with URL-encoded names that are treated as distinct from their URL-decoded counterparts.

Exploitation Scenario

An attacker with a standard MLflow account (e.g., a data analyst or compromised CI service account) identifies a production model named 'fraud-detector/v2' used by the inference pipeline. The attacker registers a model named 'fraud-detector%2Fv2'—URL-encoded to appear identical when decoded. Depending on MLflow's name resolution path, the production pipeline periodically loads the attacker's model variant. The poisoned model is trained to produce incorrect predictions on specific input patterns (e.g., a backdoor trigger), silently degrading fraud detection accuracy. The substitution goes undetected until anomaly metrics surface weeks later.