High-severity Nvidia bug could crash GPU monitoring on exposed servers

Researchers found thousands of GPU servers exposing Nvidia's DCGM Exporter to the internet, with hundreds potentially vulnerable to a high-severity flaw that could let unauthenticated attackers crash the GPU monitoring service and disrupt AI workloads. DCGM Exporters read telemetry from the GPUs on a host, including its hardware, utilization, memory usage, power consumption, and error events. Each GPU has its own unique ID, or UUID, and all of these metrics are exposed in plaintext over HTTP. This exposure provides would-be attackers with detailed information useful for reconnaissance, including mapping GPU infrastructure, identifying potentially vulnerable systems, and monitoring workload activity. Michael Katchinskiy, a researcher at datacenter security startup Lava, found and reported the bug in the GPU health and performance monitoring service. In September, the GPU giant Nvidia released a fix for the flaw, tracked as CVE-2026-47483, and gave it an 8.2 CVSS high-severity rating. Once we realized how much these endpoints revealed, the next question was: How many of them are exposed to the internet?" Katchinskiy said in a Thursday blog. So the researchers started scanning the internet for exposed DCGM Exporters. And the scale of exposure proved especially significant," he wrote. Over the course of four scans between March and May, the threat hunters found about 2,100 GPU servers exposing DCGM Exporter metrics to the open internet. These included 12,000 GPU UUIDs. None of these required authentication. The hosts belonged to about 300 organizations, according to Katchinskiy, and nearly half - 5,274 of the exposed GPUs, or 44 percent of the total - were located in the US. These GPUs represented about $100 million in hardware, and included Nvidia Blackwell Ultra B300 GPUs, H200s, and H100s - used to run large-scale AI workloads - plus consumer RTX 5090 and 4090 systems. While investigating the exposed systems, the Lava team found that about 25 percent of the exposed DCGM hosts also revealed data from Go's /debug/pprof/ built-in profiling tool. The profiler collects and exposes runtime performance data such as CPU and memory usage for running Go applications. This includes CPU usage, memory allocations, goroutine states, and blocking events. With enough concurrent unauthenticated requests, the exporter could run out of memory and crash, cutting off visibility into GPU health and activity," Katchinskiy wrote. The CPU and memory pressure could also affect AI training or inference workloads. Nvidia fixed the issue in version 4.8.2, and operators should upgrade to that version or later. In addition to GPU telemetry, Lava looked into Prometheus Node Exporter, which monitors server hardware and operating systems, and also exposes metrics over HTTP. The team found 12,096 public Node Exporter hosts exposing data on server models, operating systems, firmware versions, hostnames, storage paths and networking hardware commonly used in GPU clusters. This information reveals how environments are built and configured, which could also be used by attackers for reconnaissance, matching the system to known vulnerabilities. The publicly exposed monitoring services affected customer infrastructure across neocloud and GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data, and DigitalOcean. Lava reported all of this to the affected providers, and Katchinskiy says that these providers worked with customers to address the exposures. For operators: the security shop says Nvidia DCGM Exporter, Node Exporter and Prometheus services should not be directly reachable from the public internet and recommends restricting them to authorized monitoring infrastructure. The findings highlight a growing security gap in AI infrastructure: companies are spending millions on GPUs while leaving critical systems exposed," Katchinskiy wrote. Those exposures can reveal how AI environments are built and, in some cases, allow attackers to disrupt them." (R)