Lava has disclosed a high-severity vulnerability in NVIDIA's DCGM Exporter and reported widespread public exposure of AI GPU monitoring servers affecting internet-reachable systems across AI infrastructure.
The research identified about 2,100 publicly accessible NVIDIA DCGM Exporter hosts exposing more than 12,000 unique GPUs, alongside 12,096 servers exposing data through node_exporter. These systems were linked to nearly 300 organisations across neoclouds, traditional hosting providers, universities, telecommunications groups and public cloud environments.
NVIDIA assigned the flaw CVE-2026-47483, rated it 8.2 in severity and released a fix in version 4.8.2. According to Lava, unauthenticated attackers able to reach an exposed DCGM Exporter instance could exploit the vulnerability to trigger uncontrolled resource consumption, denial of service and information disclosure.
The monitoring tool is widely used to track the health and performance of GPUs in AI servers and clusters. When exposed directly to the public internet, Lava said, an internal monitoring component became an attack surface.
Researchers said every identified host exposed GPU telemetry without authentication, allowing anyone on the internet to view details of deployed hardware, usage levels and elements of AI environment configuration.
Based on approximate market values for the GPU models observed, Lava estimated the exposed hardware was worth more than USD $100 million. The systems included NVIDIA Blackwell Ultra B300 GPUs, H200s and H100s, as well as consumer RTX 5090 and 4090 machines.
Exposure pattern
The scans were carried out in four sweeps between March and May 2026. Lava said consumer GPU and mining farms accounted for 35% of the observed GPUs, neoclouds 25%, and traditional hosting and colocation providers 19%.
The findings also showed that about 60% of the exposed GPUs reported zero utilisation when observed. Public telemetry also made it possible to identify individual GPUs across multiple observations and track changes in workload activity even when endpoints or hostnames changed.
During the investigation, Lava found that about a quarter of the public DCGM Exporter instances also exposed internal Go profiling endpoints. Researchers initially believed this stemmed from operator misconfiguration, but later reproduced the behaviour using NVIDIA's official DCGM Exporter container.
Lava said it validated a resource exhaustion attack in a controlled environment rather than against publicly exposed systems. An unauthenticated attacker could remotely exhaust system resources and crash the monitoring service, while the resulting CPU and memory pressure could also affect AI training or inference workloads running on the same server.
Broader visibility
The exposure was not limited to GPU telemetry. Lava also found more than 12,000 publicly accessible servers exposing node_exporter data, including server models, operating systems, firmware versions, hostnames, storage paths and networking hardware used in GPU clusters.
In some cases, researchers said they could link GPU activity to the underlying server hardware, software stacks and high-speed networking infrastructure. This created a broader picture of how AI environments were built without requiring authentication.
Lava identified publicly exposed monitoring services on customer infrastructure across major neocloud and GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data and DigitalOcean. The providers indicated that most of the exposed services had been deployed by customers rather than by the providers themselves.
Lava reported the findings to the affected providers, which then worked with customers to address the exposures. It recommended preventing NVIDIA DCGM Exporter, Node Exporter and Prometheus services from being directly reachable from the public internet and restricting them to authorised monitoring infrastructure.
Yakir Kadkoda, Chief Technology Officer and Co-Founder at Lava, said the pace of AI infrastructure growth was outstripping security controls. "Neoclouds are racing to add GPU capacity, and customers are racing to use it," Kadkoda said. "That speed is creating security gaps on both sides - providers are moving faster than they can harden the environment, while customers often don't know exactly what they're inheriting or exposing."