A high-severity bug in Nvidia's DCGM Exporter, a widely used tool for collecting GPU health and usage data, could let anyone who can reach it crash the service and cut off operators' view of their hardware, The Register reported Thursday. The researcher who found it says about 2,100 GPU servers worldwide were exposing the affected metrics to the internet without authentication.

What Happened

The flaw, tracked as CVE-2026-47483 and rated 8.2 out of 10 on the CVSS scale, is a resource-exhaustion problem. Michael Katchinskiy of datacenter security startup Lava found that a large number of concurrent, unauthenticated requests could make the exporter run out of memory and crash, according to The Register.

Nvidia released a fix in DCGM Exporter version 4.8.2 in September, and operators should upgrade to that version or later. No exploitation in the wild has been reported, and The Register's story did not include comment from Nvidia.

RelatedSuspected Chinese Spies Exploited BeyondTrust Remote Support Flaw Since August, Google Says

How Exposed Systems Were Found

Lava ran internet scans in four periods between March and May 2026 and found about 2,100 GPU servers publishing DCGM metrics with no authentication. Those endpoints revealed roughly 12,000 unique GPU identifiers belonging to about 300 organizations. About 5,274 of the GPUs, or 44%, were in the United States.

The exposed machines included Nvidia's Blackwell Ultra B300 data center GPUs, H200 and H100 accelerators, and consumer RTX 5090 and RTX 4090 cards. Lava estimated the hardware was worth about $100 million. The Register reported that the exposed services touched customer infrastructure at GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data and DigitalOcean.

Why It Matters

DCGM Exporter is part of Nvidia's Data Center GPU Manager tooling. It reads telemetry such as utilization, memory use, power draw and error events and publishes it in a format that monitoring systems like Prometheus collect. In AI clusters, that feed is how operators spot failing cards, overheating and idle capacity.

Crashing the exporter does not stop the GPUs themselves, but it removes the signal operators rely on to notice problems. For a company renting out or training on thousands of accelerators, losing that visibility could hide hardware faults or unusual activity.

The exposure is also a problem without any exploit. Because the metrics are published in plain text, anyone who finds an open endpoint can learn which GPU models an organization runs, how many it has and how heavily they are being used, and can track individual cards by their unique identifiers. That inventory detail could help an attacker decide which clusters are worth targeting.

What to Watch

The fix requires operators to upgrade, so exposed systems running older releases will remain vulnerable until they are patched. Beyond the update, Lava's findings point to a configuration change: DCGM Exporter, Node Exporter and Prometheus should be reachable only from authorized monitoring systems, not the open internet.

Neocloud providers named in the report host infrastructure for many customers, and it is not yet clear whether any of them will push upgrades or network restrictions on customers' behalf.