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

Nvidia's DCGM Exporter bug could expose GPU monitoring to unauthenticated attackers, potentially disrupting AI workloads. Researchers discovered thousands of servers exposing sensitive GPU telemetry to the internet, many of which could be vulnerable to the flaw. The exporters provide detailed information like GPU hardware specs, utilization, and error events, all exposed in plaintext over HTTP.

This exposes potential targets for reconnaissance, including GPU infrastructure mapping and workload monitoring. Nvidia released a fix in version 4.8.2, rating the issue as high severity with a CVSS score of 8.2. The exposed systems include powerful GPUs like Blackwell Ultra B300, H200, and H100, as well as RTX 5090 and 4090 consumer systems, totaling about $100 million in hardware.

The exposed data also includes Go's built-in profiling tool, which could reveal CPU and memory usage for running Go applications. Over 2,100 GPU servers were found exposing DCGM Exporter metrics, with nearly half located in the US. The issue extends to Prometheus Node Exporter, which also exposes server hardware and operating system data over HTTP, potentially aiding attackers in mapping environments and identifying known vulnerabilities. Operators are advised to restrict these services to authorized monitoring infrastructure.