Block Public Access to NVIDIA DCGM Exporter Profiling Endpoints Now

NVIDIA has released a patch for CVE-2026-47483, a high-severity vulnerability in its DCGM Exporter tool. According to Cyber Security News, the flaw allows unauthenticated remote attackers to exhaust server resources, crash GPU monitoring services, and disrupt active AI workloads. Organizations using this common monitoring utility for GPU servers and Kubernetes clusters must verify their exposure immediately.

The risk

DCGM Exporter collects hardware data such as GPU model, temperature, memory usage, and power draw, then serves it via HTTP for Prometheus. The vulnerability resides in Go profiling routes under /debug/pprof, which were intended for developer debugging. In vulnerable builds, these routes lacked authentication and request limits, allowing attackers to send concurrent requests that cause uncontrolled CPU and memory consumption.

NVIDIA rated the issue 8.2 out of 10, classifying it as CWE-770, which covers resource allocation without proper limits. A successful attack can slow or crash the exporter, removing the live view of GPU health and errors. This loss of visibility can also degrade nearby AI tasks if the exporter lacks strict container or host resource limits, according to the report.

Who is exposed

Lava Security researcher Michael Katchinskiy discovered the flaw while analyzing public DCGM Exporter systems. Between March and May 2026, four internet scans identified approximately 2,100 hosts serving metrics without authentication. These systems exposed more than 12,000 unique GPU identifiers across roughly 300 organizations, with about one quarter of the hosts also exposing the risky profiling routes.

The exposed hardware included high-end models such as H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090 GPUs. The United States accounted for 5,274 of these exposed GPUs, representing 44 percent of the observed total. Researchers also found endpoints on networks linked to cloud providers including Voltage Park, Lambda, Northern Data, and DigitalOcean.

Cost and compliance

The researchers tested the resource exhaustion only in a controlled environment and did not target the public systems they identified. GPU cloud providers assisted in alerting customers after receiving the findings. This case highlights that AI monitoring ports require the same security controls as other management services to prevent unauthorized access and potential operational disruption.

Our take on the response

That suggests organizations often treat monitoring tools as internal-only utilities without verifying their external exposure. Managers should audit all Prometheus-compatible exporters to ensure they are not publicly accessible. This incident underscores the need to treat debug endpoints as critical security boundaries.

What to do and how to stay safe: NVIDIA

• Audit firewall rules to ensure DCGM Exporter ports are not accessible from the public internet.
• Verify that all Prometheus exporters require authentication before serving metrics data.
• Disable or restrict access to /debug/pprof routes in production environments immediately.
• Monitor server resource usage for unusual spikes that may indicate exploitation attempts.

Step-by-step guide: Multi-Factor Authentication Fatigue: Why Users Click Yes Too Fast

General security guidance from the BlueTeam Herald 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 flaw in NVIDIA DCGM Exporter that allows unauthenticated attackers to exhaust server resources and crash monitoring services.

Which GPUs were found exposed?

Scans identified exposed H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090 GPUs across multiple organizations.

Is there a fix available?

Yes, NVIDIA has patched the vulnerability. Users should apply the latest update to their DCGM Exporter installations.