NVIDIA Patches DCGM Exporter Flaw CVE-2026-47483 Exposing GPU Servers to Resource Exhaustion
NVIDIA released a patch for a high-severity vulnerability in DCGM Exporter that allows unauthenticated attackers to crash GPU monitoring and disrupt AI workloads.
Key points
- CVE-2026-47483 allows unauthenticated remote attackers to exhaust server resources via exposed Go profiling routes.
- NVIDIA rated the flaw 8.2 out of 10, classifying it as CWE-770 resource allocation without proper limits.
- Scans identified approximately 2,100 hosts serving metrics without authentication across about 300 organizations.
NVIDIA has patched a high-severity vulnerability in its DCGM Exporter tool that enables unauthenticated remote attackers to exhaust server resources and crash GPU monitoring services. The flaw, tracked as CVE-2026-47483, was discovered by Lava Security researcher Michael Katchinskiy and affects systems relying on this common monitoring utility for GPU servers and Kubernetes clusters.
The vulnerability allows attackers to send multiple profiling requests simultaneously, causing uncontrolled memory and CPU consumption. This resource exhaustion can slow down the exporter until it stops responding or crashes entirely. According to Cyber Security News, such a crash removes the live view of GPU health, including temperature, errors, and usage metrics, while potentially disrupting nearby AI tasks if strict container or host limits are not enforced.
Technical Context and Exposure
NVIDIA rated the issue 8.2 out of 10 on the CVSS scale, classifying it under CWE-770, which covers resource allocation without proper limits. The weakness resides in Go profiling routes under the /debug/pprof endpoint. These routes were originally designed to help developers inspect CPU and memory usage during development. However, vulnerable builds exposed these endpoints without login checks or safe request limits, making them accessible to anyone with network access.
DCGM Exporter is widely used to read details such as GPU model, temperature, memory use, power draw, and workload activity. It serves this data in a Prometheus-friendly format over HTTP. Each GPU is listed with a unique UUID. If this endpoint is public, outsiders can map valuable hardware, observe busy periods, and learn details about the software and workloads running on the system.
Affected Systems and Research Findings
Lava Security researchers conducted four internet scans between March and May 2026, identifying about 2,100 hosts serving metrics without authentication. These exposed systems belonged to approximately 300 organizations and revealed more than 12,000 unique GPU identifiers. Around one quarter of the DCGM hosts also exposed the risky profiling routes that enable the resource exhaustion attack.
The visible hardware included high-end GPUs such as the H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090. The United States accounted for 5,274 of the exposed GPUs, representing 44 percent of the observed total. Researchers also found customer-run endpoints on networks linked to GPU cloud providers, including Voltage Park, Lambda, Northern Data, and DigitalOcean. These providers helped alert their customers after receiving the findings.
What should you do?
The researchers tested the resource exhaustion only in a controlled environment and did not target the public systems they found. The case highlights the need for AI monitoring ports to receive the same security care as other management services. Organizations should verify that their DCGM Exporter instances are updated to the patched version provided by NVIDIA to mitigate this specific risk.
What to do and how to stay safe: NVIDIA
- Verify that your DCGM Exporter instances are running the latest patched version provided by NVIDIA to address CVE-2026-47483.
- Ensure that monitoring endpoints are not publicly accessible and require authentication to prevent unauthorized access to hardware details.
- Implement strict request limits and container resource constraints to mitigate the impact of potential resource exhaustion attacks.
- Review firewall rules to restrict access to Go profiling routes and other debug endpoints to internal development networks only.
Step-by-step guide: Social Engineering Warning Signs and How to Respond
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is CVE-2026-47483?
CVE-2026-47483 is a high-severity vulnerability in NVIDIA DCGM Exporter that allows unauthenticated attackers to exhaust server resources via exposed profiling routes.
Which GPUs were found exposed in the scans?
The scans identified exposed H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090 GPUs across multiple organizations.
Did the researchers exploit the vulnerability in the wild?
No, the researchers tested the resource exhaustion only in a controlled environment and did not target the public systems they discovered.
