What is the CVSS score of CVE-2026-24151?

CVE-2026-24151 has a CVSS 3.1 base score of 7.8 (High). CVSS vector: CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H. EPSS exploitation probability: 0.1%.

Which versions of Nvidia are affected by CVE-2026-24151?

A critical remote code execution vulnerability in NVIDIA Megatron-LM could allow attackers to compromise AI infrastructure and execute arbitrary code on systems running this widely-used large…

How to fix or mitigate CVE-2026-24151?

Within 24 hours: Inventory all systems running NVIDIA Megatron-LM and assess exposure in production vs. development environments. Within 7 days: Implement network segmentation to restrict access to Megatron-LM inference services, disable remote model loading where possible, and enforce strict input validation on model files. Within 30 days: Establish continuous monitoring for suspicious file loading activities, evaluate alternative frameworks if business-critical, and prepare remediation playbooks for detected exploitation attempts.

Nvidia CVE-2026-24151

EUVD-2026-15007 HIGH

Deserialization of Untrusted Data (CWE-502)

2026-03-24 nvidia

RCE Information Disclosure Deserialization Nvidia

7.8

CVSS 3.1 · NVD

Severity by source

NVD PRIMARY

7.8 HIGH

AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H

Primary rating from NVD · only source for this CVE.

CVSS VectorNVD

Attack Vector

Local

Attack Complexity

Low

Privileges Required

Low

User Interaction

None

Scope

Unchanged

Confidentiality

High

Integrity

High

Availability

High

Lifecycle Timeline

EUVD ID Assigned

Mar 24, 2026 - 20:31 euvd

EUVD-2026-15007

Analysis Generated

Mar 24, 2026 - 20:31 vuln.today

CVE Published

Mar 24, 2026 - 20:24 nvd

HIGH 7.8

DescriptionCVE.org

NVIDIA Megatron-LM contains a vulnerability in inferencing where an Attacker may cause an RCE by convincing a user to load a maliciously crafted input. A successful exploit of this vulnerability may lead to code execution, escalation of privileges, information disclosure, and data tampering.

AnalysisAI

NVIDIA Megatron-LM contains an insecure deserialization vulnerability (CWE-502) during model inferencing that allows remote code execution when a user loads a maliciously crafted input file. This vulnerability has a CVSS score of 7.8 and requires local access with low privileges but no user interaction, enabling attackers to execute arbitrary code, escalate privileges, disclose sensitive information, and tamper with data. The vulnerability affects NVIDIA's large language model training framework widely used in AI research and production environments.

Technical ContextAI

NVIDIA Megatron-LM is an open-source deep learning framework designed for training large transformer-based language models at scale. The vulnerability stems from CWE-502 (Deserialization of Untrusted Data), which occurs when the inferencing component deserializes model inputs without proper validation. When loading model checkpoints or input data, the framework likely uses Python's pickle or similar serialization mechanisms that can execute arbitrary code embedded in crafted serialized objects. The affected product is identified by CPE cpe:2.3:a:nvidia:megatron_lm:*:*:*:*:*:*:*:*, indicating all versions prior to a patched release are vulnerable. This class of vulnerability is particularly dangerous in machine learning frameworks where model files and data are frequently exchanged between systems and users.

RemediationAI

Consult the NVIDIA security advisory at https://nvidia.custhelp.com/app/answers/detail/a_id/5769 for the patched version of Megatron-LM and upgrade immediately to the remediated release. Until patching is completed, implement strict input validation by only loading model checkpoints and inference inputs from trusted, verified sources with cryptographic signature verification. Restrict filesystem permissions so that the Megatron-LM process can only read from controlled directories, and apply principle of least privilege by running inference workloads under dedicated service accounts with minimal system access. In multi-tenant or shared environments, consider sandboxing inference operations using containers with restricted capabilities or separate virtual machines to limit the blast radius of potential exploitation. Organizations should audit their model supply chain and establish secure model distribution channels with integrity verification before accepting any external model files.