Thousands of Exposed Nvidia GPU Servers Could Be Knocked Offline by High-Severity Flaw

Researchers have found thousands of GPU servers exposing Nvidia's DCGM Exporter to the open internet, and hundreds may be vulnerable to a high-severity flaw. The bug could let an unauthenticated attacker crash the GPU monitoring service and disrupt AI workloads. Nvidia released a fix in September, and the flaw has an 8.2 CVSS high-severity rating. The Register's report gives the identifier as CVE-2026-4748 3, with a stray space in the number, so check Nvidia's advisory for the exact reference.

The DCGM Exporter reads data from the GPUs on a host, including hardware details, utilisation, memory use, power consumption and error events. Each GPU has a unique ID, and all of this is shared in plaintext over HTTP. That detail can help attackers map GPU infrastructure, spot potentially vulnerable systems and watch workload activity.

Michael Katchinskiy of security startup Lava found and reported the bug. Across four scans between March and May, his team found about 2,100 GPU servers exposing metrics with no authentication, covering 12,000 GPU IDs and about 300 organisations. Nearly half of the exposed GPUs (44 percent) were in the US. The hardware was worth about $100 million and included high-end AI GPUs as well as consumer RTX cards. About 25 percent of the exposed hosts also revealed data from Go's built-in profiling tool, which shares runtime performance details such as CPU and memory use.

This summary is based on the opening of the article, so it does not cover the full story.