Nvidia Security Flaw Exposes 2,100 GPU Servers Running $100M Worth of AI Hardware

High-Severity Nvidia Bug Threatens Exposed GPU Infrastructure

A critical Nvidia security flaw in the company's GPU monitoring service has left thousands of servers vulnerable to attacks that could crash GPU monitoring and disrupt AI workloads. Security researchers at datacenter security startup Lava discovered the high-severity Nvidia bug, tracked as CVE-2026-47483, which earned a CVSS 8.2 rating and affects the widely-used DCGM Exporter tool1

1

. The vulnerability allows unauthenticated attackers to send multiple requests to debugging endpoints, forcing servers to consume excessive memory and CPU power, ultimately leading to a denial of service attack that crashes the monitoring service2

2

.

Massive Scale of Exposed GPU Servers Revealed

Michael Katchinskiy, a researcher at Lava who discovered and reported the flaw, conducted internet-wide scans between March and May that uncovered approximately 2,100 GPU servers exposing DCGM Exporter metrics without authentication. These exposed GPU servers revealed telemetry data from over 12,000 GPU UUIDs, representing an estimated $100 million in hardware across roughly 300 organizations1

1

. The geographic distribution showed 5,274 GPUs—44 percent of the total—located in the United States, making it the most affected region2

2

.

How DCGM Exporter Vulnerability Works

DCGM Exporter functions as a health monitor for Nvidia GPUs, collecting critical telemetry data including temperature, memory usage, power consumption, performance metrics, and error events. Each GPU carries a unique identifier, and this information is exposed in plaintext over HTTP without requiring authentication1

1

. The vulnerability stems from DCGM Exporter accidentally exposing Go's /debug/pprof/ built-in profiling tool alongside standard metrics. According to Katchinskiy, "About a quarter of the exposed DCGM hosts in our scan were serving Go's /debug/pprof/ profiling endpoints alongside /metrics"2

2

. These debugging endpoints collect runtime performance data including CPU usage, memory allocations, goroutine states, and blocking events. With enough concurrent unauthenticated requests to these pprof endpoints, attackers could force the exporter to run out of memory and crash, cutting off visibility into GPU health and activity while potentially affecting AI training or inference workloads running on the same infrastructure1

1

.

Premium AI Hardware at Risk

The exposed systems included some of Nvidia's most advanced and expensive AI infrastructure. Researchers identified Nvidia Blackwell Ultra B300 GPUs, H200s, and H100 processors used for large-scale AI workloads, alongside consumer-grade RTX 5090 and 4090 systems1

1

. Consumer GPUs and mining farms accounted for approximately 35 percent of the exposed systems, demonstrating the vulnerability extended beyond enterprise AI infrastructure2

2

. The publicly exposed monitoring services affected customer infrastructure across neocloud and GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data, and DigitalOcean1

1

.

Reconnaissance Risks Beyond Service Disruption

Beyond enabling denial of service attacks, the exposed telemetry data provides attackers with detailed intelligence for reconnaissance operations. The plaintext HTTP exposure reveals information useful for mapping GPU infrastructure, identifying potentially vulnerable systems, and monitoring workload activity patterns1

1

. Lava's investigation also examined Prometheus Node Exporter exposures, finding 12,096 public hosts revealing data on server models, operating systems, firmware versions, hostnames, storage paths, and networking hardware commonly deployed in GPU clusters. This information exposes how AI infrastructure environments are built and configured, allowing attackers to match systems to known vulnerabilities1

1

.

Nvidia Releases Fix and Security Recommendations

Nvidia addressed the Nvidia security flaw in September by releasing DCGM Exporter version 4.8.2, and administrators should upgrade immediately to this version or later1

1

. Security experts recommend that DCGM Exporter, Node Exporter, and Prometheus services should never be directly reachable from the public internet. Organizations must restrict these monitoring tools to authorized monitoring infrastructure only and keep them on private networks protected by firewalls2

2

. Lava reported all exposures to affected providers, and according to Katchinskiy, these providers worked with customers to address the security gaps. "The findings highlight a growing security gap in AI infrastructure: companies are spending millions on GPUs while leaving critical systems exposed," Katchinskiy emphasized, noting that "those exposures can reveal how AI environments are built and, in some cases, allow attackers to disrupt them"1

1

. With GPU prices soaring amid the ongoing AI boom, securing these valuable AI security assets has become more critical than ever2

2

.