NVIDIA patches CVSS 8.2 CVE-2026-47483 in DCGM Exporter blocking remote unauthenticated AI workload crashes

Lava Security researchers have disclosed a high-severity vulnerability in NVIDIA’s DCGM Exporter that allows attackers to remotely disrupt GPU monitoring services. The flaw, tracked as CVE-2026-47483, enables unauthenticated users to exhaust server resources and crash the monitoring agent, potentially halting AI workloads. NVIDIA has acknowledged the issue and released a fix.

In plain English

DCGM Exporter is a monitoring tool used to collect performance data from GPU servers, such as temperature, power usage, and memory utilization. These metrics are critical for managing AI and machine learning infrastructure. The vulnerability allows anyone with internet access to the monitoring endpoint to interfere with this service. An attacker does not need a password or login credentials to exploit the flaw. By sending specific requests, they can force the monitoring service to crash or consume excessive server resources. This can lead to the loss of telemetry data and potentially disrupt the AI computations running on those GPUs.

The background

According to Help Net Security, Lava Security reported the vulnerability to NVIDIA through a responsible disclosure process. NVIDIA rated the issue at 8.2 on the CVSS scale, classifying it as high severity, and published a security bulletin on July 28, 2026. Prior to the fix, Lava Security’s research indicated widespread exposure. The researchers identified more than 2,100 servers with exposed GPU monitoring endpoints on the public internet.

These endpoints revealed telemetry from over 12,000 GPUs, including high-end models like H100s, H200s, and Blackwell units. The research noted that approximately 60% of the exposed GPUs showed zero utilization during scans, suggesting they may have been idle or misconfigured. Additionally, 12,096 servers exposed further infrastructure details, such as operating system versions, firmware information, and networking configurations.

What changes now

NVIDIA has patched the vulnerability in DCGM Exporter. Organizations using this monitoring tool must apply the update to prevent unauthenticated attacks. The discovery highlights the risks of exposing diagnostic and monitoring ports to the public internet. Even if the GPUs themselves are not directly targeted for computation theft, the ability to crash monitoring services can cause operational downtime and obscure the status of expensive hardware. The exposure of infrastructure details on thousands of servers also provides attackers with reconnaissance data that could be used for further targeted attacks.

What to do and how to stay safe: NVIDIA

• Verify if your organization uses NVIDIA DCGM Exporter and check for the latest vendor update.
• Ensure monitoring endpoints are not accessible from the public internet; restrict access to internal networks.
• Review firewall rules to block unauthenticated traffic to GPU management ports.
• Audit exposed server configurations to prevent leakage of OS and firmware details.

Step-by-step guide: Automatic Updates Mistakes That Leave Systems Exposed

General security guidance from the RedTeam Monitor newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is CVE-2026-47483?

It is a high-severity vulnerability in NVIDIA DCGM Exporter that allows unauthenticated attackers to crash GPU monitoring services and disrupt workloads.

Is there a fix available?

Yes, NVIDIA published a security bulletin and patch on July 28, 2026, following responsible disclosure by Lava Security.

How many servers were exposed?

Researchers found more than 2,100 exposed GPU servers and 12,096 additional servers exposing infrastructure details on the public internet.